datadog 2.4.0 → 2.44.0

data/ext/datadog_profiling_native_extension/collectors_thread_context.c

@@ -1,5 +1,7 @@
 #include <ruby.h>
+#include <ruby/debug.h>
 
+#include "datadog_ruby_common.h"
 #include "collectors_thread_context.h"
 #include "clock_id.h"
 #include "collectors_stack.h"
@@ -7,8 +9,11 @@
 #include "helpers.h"
 #include "libdatadog_helpers.h"
 #include "private_vm_api_access.h"
+#include "ruby_helpers.h"
 #include "stack_recorder.h"
 #include "time_helpers.h"
+#include "unsafe_api_calls_check.h"
+#include "extconf.h"
 
 // Used to trigger sampling of threads, based on external "events", such as:
 // * periodic timer for cpu-time and wall-time
@@ -38,11 +43,8 @@
 // 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.)
+// While `cpu_time_at_gc_start_ns` is set, we don't expect the thread to be sampled: the VM is doing GC
+// on the thread holding the GVL so no other samples can/will be triggered until GC finishes.
 //
 // 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`
@@ -71,15 +73,19 @@
 
 #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
+
+// Used as markers in the code
+#define OTEL_CURRENT_SPAN_KEY_NOT_EXTRACTED INT2FIX(1)
+#define OTEL_CURRENT_SPAN_KEY_EXTRACTION_NEEDED INT2FIX(2)
+
 #define TIME_BETWEEN_GC_EVENTS_NS MILLIS_AS_NS(10)
+#define GVL_SUSPENDED ((uint64_t)1)
+#define GVL_RUNNING ((uint64_t)0)
+
+#define MAX(a, b) ((a) < (b) ? (b) : (a))
 
-// 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 dd_per_thread_context_id; // Hidden ivar (no @ prefix, inaccessible from Ruby)
 
 static ID at_active_span_id;  // id of :@active_span in Ruby
 static ID at_active_trace_id; // id of :@active_trace in Ruby
@@ -101,60 +107,78 @@ 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
+static ID otel_fiber_context_storage_id; // id of :@opentelemetry_context in Ruby
+
+// This is mutable and gets set last-writer-wins style by
+// `thread_context_collector_reset_all_per_thread_contexts`, which is called whenever
+// profiling is starting or restarting.
+//
+// Note: We must be careful to not change this value while the profiler is still running,
+// otherwise new threads can get the new value and cause profiling to stop with an
+// exception due to the mismatched size.
+//
+// The initial value should be kept in sync with the default for DD_PROFILING_MAX_FRAMES
+// in settings.rb. See `initialize_context` for details on why this is needed/used.
+static uint16_t latest_max_frames = 400;
 
-// 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);
+// Global tracepoint for RUBY_EVENT_THREAD_BEGIN. Created and enabled once when the first ThreadContext collector is initialized.
+static VALUE thread_begin_tracepoint = Qnil;
 
-enum otel_context_enabled {otel_context_enabled_false, otel_context_enabled_only, otel_context_enabled_both};
+typedef enum { OTEL_CONTEXT_ENABLED_FALSE, OTEL_CONTEXT_ENABLED_ONLY, OTEL_CONTEXT_ENABLED_BOTH } otel_context_enabled;
+typedef enum { OTEL_CONTEXT_SOURCE_UNKNOWN, OTEL_CONTEXT_SOURCE_FIBER_IVAR, OTEL_CONTEXT_SOURCE_FIBER_LOCAL } otel_context_source;
 
 // Contains state for a single ThreadContext instance
-struct thread_context_collector_state {
+typedef struct {
   // 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
-  ddog_prof_Location *locations;
-  uint16_t max_frames;
-  // Hashmap <Thread Object, struct per_thread_context>
-  st_table *hash_map_per_thread_context;
+  // Output buffer for stack traces, passed to sample_thread()
+  sample_locations locations;
   // 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 control context collection
-  enum otel_context_enabled otel_context_enabled;
-  // Used to omit class information from collected allocation data
-  bool allocation_type_enabled;
+  otel_context_enabled otel_context_enabled;
+  // Used to remember where otel context is being stored after we observe it the first time
+  otel_context_source otel_context_source;
   // 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;
   // 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.
+  // * `OTEL_CURRENT_SPAN_KEY_NOT_EXTRACTED`: we've not needed it yet
+  // * `OTEL_CURRENT_SPAN_KEY_EXTRACTION_NEEDED`: a sample needed it, and it's waiting to be extracted
+  // * anything else: the extracted value, or Qnil if we tried to extract it and failed
   VALUE otel_current_span_key;
+  // Used to enable native filenames in stack traces
+  bool native_filenames_enabled;
+  // Used to include class/module names in method names (e.g. "Foo#bar" instead of "bar")
+  bool show_classes;
+  // Used to cache native filename lookup results (Map[void *function_pointer, char *filename])
+  st_table *native_filenames_cache;
+  // Used to attribute overhead during sampling to this component
+  VALUE overhead_filename;
 
   struct stats {
+    // Track how many regular samples we've taken. Does not include garbage collection samples.
+    unsigned int sample_count;
     // 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;
+    // See thread_context_collector_sample_after_gc for details
+    unsigned int gc_samples_skipped_nothing_to_flush;
+    // How many per-thread samples were skipped because the thread has been continuously suspended
+    // (no GVL) since its previous sample, so its Ruby stack cannot have changed.
+    unsigned int inactive_thread_samples_skipped;
+    unsigned int profiler_thread_samples_skipped;
   } stats;
 
   struct {
@@ -164,18 +188,93 @@ struct thread_context_collector_state {
     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
+} thread_context_collector_state;
+
+// Tracks per-thread state.
+// This state is global and lives forever on the Ruby Thread (until the Thread is GC'd).
+// The state is created early on for all threads on the main Ractor
+// (enabling a TracePoint only enables it for the current Ractor).
+// The state is either created when the Thread starts running (via the RUBY_EVENT_THREAD_BEGIN TracePoint),
+// or via `thread_context_collector_reset_all_per_thread_contexts` when the profiler starts.
+//
+// Unfortunately that RUBY_EVENT_THREAD_BEGIN TracePoint still fires after some other events:
+// * RUBY_INTERNAL_THREAD_EVENT_RESUMED for the Thread acquiring the GVL for the first time
+// * an early SIGPROF calling handle_sampling_signal()
+// * Ruby might check for interrupts and run postponed jobs (e.g. thread_context_collector_sample)
+// * another RUBY_EVENT_THREAD_BEGIN TracePoint which might run before ours
+// * etc
+// For those cases we have to ignore those events and we cannot assume the state is always set,
+// however this only matters for a very short period when a thread starts.
 struct per_thread_context {
-  sampling_buffer *sampling_buffer;
+  sampling_buffer sampling_buffer;
+  bool prepare_sample_on_gc_finish;
   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
+  long cpu_time_at_previous_sample_ns;
+  long wall_time_at_previous_sample_ns;
+
+  // There are 3 possible states for the GVL (per thread), and 3 transitions for which we receive GVL events:
+  // Thread holds the GVL
+  //   on_gvl_released() the thread releases the GVL (RUBY_INTERNAL_THREAD_EVENT_SUSPENDED)
+  // Thread runs without the GVL
+  //   on_gvl_waiting() the thread wants the GVL (RUBY_INTERNAL_THREAD_EVENT_READY)
+  // Thread is "Waiting for GVL"
+  //   on_gvl_running() the thread now got the GVL (RUBY_INTERNAL_THREAD_EVENT_RESUMED)
+  // ... and the cycle restarts
+
+
+  // --- GVL waiting tracking state machine ---
+  //
+  // gvl_waiting_at tracks the GVL wait state for each profiled thread:
+  //
+  //        ┌───────────────────────────────────┐
+  //        │           on_gvl_waiting          │
+  //        │                                   ▼
+  //   Not Waiting (0) ◀────────────────── Waiting (> 0)
+  //        ▲            on_gvl_running         │
+  //        │          (below threshold)        │ on_gvl_running (above threshold)
+  //        │                                   ▼
+  //        └─────────────────────────── Sample Pending (< 0)
+  //      sample / sample_after_gvl_running
+  //
+  // Not Waiting (0):      thread is running or not waiting for the GVL
+  // Waiting (> 0):        monotonic wall time (ns) when the thread started waiting
+  // Sample Pending (< 0): negated timestamp; the wait ended and a sample is pending
+  //
+  // The field is accessed under the GVL for most functions EXCEPT on_gvl_waiting() which writes to it without the GVL.
+  // So we need to pack the above state in a single long to ensure atomicity.
+  long gvl_waiting_at;
+
+  // Per-thread "state + version" word, updated on every GVL state transition. The encoding is:
+  //   - low bit:  current state (1 = currently suspended, 0 = currently running)
+  //   - bits 1+:  monotonic event counter (incremented on every RESUMED)
+  // The hooks set the state bit explicitly rather than relying on parity, so the encoding stays
+  // correct even when events are not paired properly (as in tests).
+  //
+  // Note that SUSPENDED can happen multiple times in a row on Ruby 3.2,
+  // see https://github.com/DataDog/dd-trace-rb/pull/5777#discussion_r3388560254,
+  // the encoding is designed to naturally not change the field in such a case.
+  uint64_t gvl_state_change_count;
+  // Snapshot of the thread's gvl_state_change_count at the moment we last sampled it.
+  // Equality with this snapshot means no GVL transition since the last sample.
+  uint64_t gvl_state_change_count_at_previous_sample;
+  // True when the previous per-tick sample was skipped by the SUSPENDED-skip optimization, so the
+  // flush-before-serialize pass knows it needs to report this thread.
+  // As a result, we will accumulate all wall & CPU time as a single batch per reporting period,
+  // but this is deemed worth it for this optimization. In any case we don't know exactly
+  // at what time a thread was doing CPU work (unless it's on CPU 100% of the time).
+  bool was_skipped_at_last_sample;
+  // Set as true for CpuAndWallTimeWorker and IdleSamplingHelper threads.
+  // When true, per-tick samples are skipped entirely; the thread is sampled only once per
+  // reporting period during the on_serialize flush.
+  //
+  // These threads are always in native code so their stacks aren't interesting;
+  // the Profiling::Scheduler thread on the other hand does a lot of different
+  // things using a mix of Ruby and native code, so that one isn't considered internal.
+  bool is_profiler_internal_thread;
 
   struct {
     // Both of these fields are set by on_gc_start and kept until on_gc_finish is called.
@@ -186,55 +285,40 @@ struct per_thread_context {
 };
 
 // Used to correlate profiles with traces
-struct trace_identifiers {
+typedef struct {
   bool valid;
   uint64_t local_root_span_id;
   uint64_t span_id;
   VALUE trace_endpoint;
-};
+} trace_identifiers;
 
-struct otel_span {
+typedef struct {
   VALUE span;
   VALUE span_id;
   VALUE trace_id;
-};
+} otel_span;
 
 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 void per_thread_context_typed_data_mark(void *ctx_ptr);
+static void per_thread_context_typed_data_free(void *ctx_ptr);
 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 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_initialize(int argc, VALUE *argv, DDTRACE_UNUSED VALUE _self);
+static VALUE _native_sample(VALUE self, VALUE collector_instance, VALUE allow_exception);
 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, VALUE reset_monotonic_to_system_state);
+static VALUE _native_sample_after_gc(DDTRACE_UNUSED VALUE self, VALUE collector_instance);
 static void update_metrics_and_sample(
-  struct thread_context_collector_state *state,
+  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,
-  long current_monotonic_wall_time_ns
+  per_thread_context *thread_context,
+  long current_monotonic_wall_time_ns,
+  bool force_sample
 );
 static void trigger_sample_for_thread(
-  struct thread_context_collector_state *state,
-  VALUE thread,
-  VALUE stack_from_thread,
-  struct per_thread_context *thread_context,
-  sampling_buffer* sampling_buffer,
+  thread_context_collector_state *state,
+  VALUE thread_being_sampled,
+  per_thread_context *thread_context,
   sample_values values,
   long current_monotonic_wall_time_ns,
   ddog_CharSlice *ruby_vm_type,
@@ -242,32 +326,33 @@ static void trigger_sample_for_thread(
   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);
-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 void free_context(struct per_thread_context* thread_context);
+static void check_frozen_thread(VALUE thread);
+static per_thread_context *get_or_create_context_for(VALUE thread);
+static void initialize_context(VALUE thread, per_thread_context *thread_context);
 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 per_thread_context_to_ruby_hash(per_thread_context *thread_context);
+static VALUE stats_to_ruby_hash(thread_context_collector_state *state, VALUE hash);
+static VALUE gc_tracking_as_ruby_hash(thread_context_collector_state *state);
 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 update_cpu_time_since_previous_sample(per_thread_context *thread_context, long current_cpu_time_ns);
+static long update_wall_time_since_previous_sample(per_thread_context *thread_context, long current_wall_time_ns);
+static long cpu_time_now_ns(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 void trace_identifiers_for(
+  thread_context_collector_state *state,
+  VALUE thread,
+  trace_identifiers *trace_identifiers_result
+);
 static bool should_collect_resource(VALUE root_span);
 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 thread_list(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);
 static void ddtrace_otel_trace_identifiers_for(
-  struct thread_context_collector_state *state,
+  thread_context_collector_state *state,
   VALUE *active_trace,
   VALUE *root_span,
   VALUE *numeric_span_id,
@@ -276,29 +361,41 @@ static void ddtrace_otel_trace_identifiers_for(
 );
 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,
+  thread_context_collector_state *state,
   VALUE thread_being_sampled,
-  VALUE stack_from_thread,
-  struct per_thread_context *thread_context,
-  sampling_buffer* sampling_buffer,
+  per_thread_context *thread_context,
   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 VALUE _native_on_gvl_released(DDTRACE_UNUSED VALUE self, VALUE thread);
+#ifndef NO_GVL_INSTRUMENTATION
+  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, VALUE waiting_for_gvl_threshold_ns);
+  static VALUE _native_sample_after_gvl_running(DDTRACE_UNUSED VALUE self, VALUE collector_instance, VALUE thread);
+#endif
+static VALUE _native_apply_delta_to_cpu_time_at_previous_sample_ns(DDTRACE_UNUSED VALUE self, VALUE thread, VALUE delta_ns);
 static void otel_without_ddtrace_trace_identifiers_for(
-  struct thread_context_collector_state *state,
+  thread_context_collector_state *state,
   VALUE thread,
-  struct trace_identifiers *trace_identifiers_result
+  trace_identifiers *trace_identifiers_result
 );
-static struct otel_span otel_span_from(VALUE otel_context, VALUE otel_current_span_key);
+static otel_span otel_span_from(VALUE otel_context, VALUE otel_current_span_key);
+static bool is_otel_current_span_key_needed(thread_context_collector_state *state);
 static uint64_t otel_span_id_to_uint(VALUE otel_span_id);
+static VALUE safely_lookup_hash_without_going_into_ruby_code(VALUE hash, VALUE key);
+static VALUE _native_system_epoch_time_now_ns(DDTRACE_UNUSED VALUE self, VALUE collector_instance);
+static VALUE _native_prepare_sample_inside_signal_handler(DDTRACE_UNUSED VALUE self);
+static VALUE _native_request_prepare_on_gc_finish(DDTRACE_UNUSED VALUE self);
+static VALUE _native_mark_thread_as_profiler_internal(DDTRACE_UNUSED VALUE self, VALUE thread);
+static VALUE _native_remove_per_thread_context_for(DDTRACE_UNUSED VALUE self, VALUE thread);
+static VALUE _native_global_reset_per_thread_context(DDTRACE_UNUSED VALUE self, VALUE collector_instance);
+static bool skip_sample(thread_context_collector_state *state, per_thread_context *thread_context, bool is_gvl_waiting_state, bool force_sample);
+static void on_thread_begin_event(VALUE tracepoint_data, DDTRACE_UNUSED void *unused);
 
 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");
 
@@ -312,27 +409,34 @@ 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, 9);
+  rb_define_singleton_method(collectors_thread_context_class, "_native_initialize", _native_initialize, -1);
   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, 2);
+  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);
   rb_define_singleton_method(testing_module, "_native_sample_skipped_allocation_samples", _native_sample_skipped_allocation_samples, 2);
+  rb_define_singleton_method(testing_module, "_native_system_epoch_time_now_ns", _native_system_epoch_time_now_ns, 1);
+  rb_define_singleton_method(testing_module, "_native_prepare_sample_inside_signal_handler", _native_prepare_sample_inside_signal_handler, 0);
+  rb_define_singleton_method(testing_module, "_native_request_prepare_on_gc_finish", _native_request_prepare_on_gc_finish, 0);
+  rb_define_singleton_method(testing_module, "_native_remove_per_thread_context_for", _native_remove_per_thread_context_for, 1);
+  rb_define_singleton_method(testing_module, "_native_global_reset_per_thread_context", _native_global_reset_per_thread_context, 1);
+  rb_define_singleton_method(testing_module, "_native_mark_thread_as_profiler_internal", _native_mark_thread_as_profiler_internal, 1);
+  rb_define_singleton_method(testing_module, "_native_on_gvl_waiting", _native_on_gvl_waiting, 1);
+  rb_define_singleton_method(testing_module, "_native_on_gvl_released", _native_on_gvl_released, 1);
   #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_on_gvl_running", _native_on_gvl_running, 2);
     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
+  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, 2);
 
   at_active_span_id = rb_intern_const("@active_span");
   at_active_trace_id = rb_intern_const("@active_trace");
@@ -351,16 +455,19 @@ void collectors_thread_context_init(VALUE profiling_module) {
   at_name_id = rb_intern_const("@name");
   server_id = rb_intern_const("server");
   otel_context_storage_id = rb_intern_const("__opentelemetry_context_storage__");
+  otel_fiber_context_storage_id = rb_intern_const("@opentelemetry_context");
 
-  #ifndef NO_GVL_INSTRUMENTATION
-    // This will raise if Ruby already ran out of thread-local keys
-    gvl_profiling_init();
-  #endif
+  dd_per_thread_context_id = rb_intern_const("dd_per_thread_context");
+
+  // This will raise if Ruby already ran out of thread-local keys
+  per_thread_context_tls_init();
+
+  rb_global_variable(&thread_begin_tracepoint);
 
   gc_profiling_init();
 }
 
-// This structure is used to define a Ruby object that stores a pointer to a struct thread_context_collector_state
+// This structure is used to define a Ruby object that stores a pointer to a 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",
@@ -376,71 +483,94 @@ static const rb_data_type_t thread_context_collector_typed_data = {
 // 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;
+  thread_context_collector_state *state = (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);
   rb_gc_mark(state->otel_current_span_key);
+  rb_gc_mark(state->overhead_filename);
 }
 
 static void thread_context_collector_typed_data_free(void *state_ptr) {
-  struct thread_context_collector_state *state = (struct thread_context_collector_state *) state_ptr;
+  thread_context_collector_state *state = (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->locations != NULL) ruby_xfree(state->locations);
+  sample_locations_free(&state->locations);
 
-  // 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);
+  st_free_table(state->native_filenames_cache);
 
   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;
+// per_thread_context is wrapped in a TypedData Ruby object stored as an ivar on each Ruby Thread.
+// This gives us automatic GC marking (for sampling_buffer iseq VALUEs) and lifecycle management.
+static const rb_data_type_t per_thread_context_typed_data = {
+  .wrap_struct_name = "Datadog::Profiling::PerThreadContext",
+  .function = {
+    .dmark = per_thread_context_typed_data_mark,
+    .dfree = per_thread_context_typed_data_free,
+    .dsize = NULL,
+  },
+  .flags = RUBY_TYPED_FREE_IMMEDIATELY,
+};
+
+static void per_thread_context_typed_data_mark(void *ctx_ptr) {
+  per_thread_context *ctx = (per_thread_context *) ctx_ptr;
+  if (sampling_buffer_needs_marking(&ctx->sampling_buffer)) {
+    sampling_buffer_mark(&ctx->sampling_buffer);
+  }
 }
 
-// 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 *thread_context = (struct per_thread_context*) value_per_thread_context;
-  free_context(thread_context);
-  return ST_CONTINUE;
+static void per_thread_context_typed_data_free(void *ctx_ptr) {
+  per_thread_context *ctx = (per_thread_context *) ctx_ptr;
+  sampling_buffer_free(&ctx->sampling_buffer);
+  free(ctx);
+}
+
+// Only for testing: removes the per-thread context without recreating it, so the thread has no context.
+// This simulates situations where we observe a thread running before our RUBY_EVENT_THREAD_BEGIN tracepoint fires,
+// so that we can test our `get_per_thread_context(...) => null` code paths.
+static VALUE _native_remove_per_thread_context_for(DDTRACE_UNUSED VALUE self, VALUE thread) {
+  check_frozen_thread(thread);
+  per_thread_context *ctx = get_per_thread_context(thread);
+  if (ctx != NULL) {
+    set_per_thread_context(thread, NULL);
+    rb_ivar_set(thread, dd_per_thread_context_id, Qnil);
+  }
+  return Qnil;
+}
+
+static VALUE _native_global_reset_per_thread_context(DDTRACE_UNUSED VALUE self, VALUE collector_instance) {
+  thread_context_collector_reset_all_per_thread_contexts(collector_instance);
+  return Qnil;
 }
 
 static VALUE _native_new(VALUE klass) {
-  struct thread_context_collector_state *state = ruby_xcalloc(1, sizeof(struct thread_context_collector_state));
+  thread_context_collector_state *state = ruby_xcalloc(1, sizeof(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->locations = NULL;
-  state->max_frames = 0;
-  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;
   VALUE thread_list_buffer = rb_ary_new();
   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->native_filenames_enabled = false;
+  state->show_classes = false;
+  state->native_filenames_cache = st_init_numtable();
+  state->otel_context_enabled = OTEL_CONTEXT_ENABLED_FALSE;
+  state->otel_context_source = OTEL_CONTEXT_SOURCE_UNKNOWN;
   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 = Qtrue;
+  state->otel_current_span_key = OTEL_CURRENT_SPAN_KEY_NOT_EXTRACTED;
   state->gc_tracking.wall_time_at_previous_gc_ns = INVALID_TIME;
   state->gc_tracking.wall_time_at_last_flushed_gc_event_ns = 0;
 
@@ -456,46 +586,49 @@ 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,
-  VALUE recorder_instance,
-  VALUE max_frames,
-  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_initialize(int argc, VALUE *argv, DDTRACE_UNUSED VALUE _self) {
+  VALUE options;
+  rb_scan_args(argc, argv, "0:", &options);
+  if (options == Qnil) options = rb_hash_new();
+
+  VALUE self_instance = rb_hash_fetch(options, ID2SYM(rb_intern("self_instance")));
+  VALUE recorder_instance = rb_hash_fetch(options, ID2SYM(rb_intern("recorder")));
+  VALUE max_frames = rb_hash_fetch(options, ID2SYM(rb_intern("max_frames")));
+  VALUE tracer_context_key = rb_hash_fetch(options, ID2SYM(rb_intern("tracer_context_key")));
+  VALUE endpoint_collection_enabled = rb_hash_fetch(options, ID2SYM(rb_intern("endpoint_collection_enabled")));
+  VALUE otel_context_enabled = rb_hash_fetch(options, ID2SYM(rb_intern("otel_context_enabled")));
+  VALUE native_filenames_enabled = rb_hash_fetch(options, ID2SYM(rb_intern("native_filenames_enabled")));
+  VALUE show_classes = rb_hash_fetch(options, ID2SYM(rb_intern("show_classes")));
+  VALUE overhead_filename = rb_hash_fetch(options, ID2SYM(rb_intern("overhead_filename")));
+
+  ENFORCE_TYPE(max_frames, T_FIXNUM);
   ENFORCE_BOOLEAN(endpoint_collection_enabled);
-  ENFORCE_BOOLEAN(timeline_enabled);
-  ENFORCE_TYPE(waiting_for_gvl_threshold_ns, T_FIXNUM);
-  ENFORCE_BOOLEAN(allocation_type_enabled);
+  ENFORCE_BOOLEAN(native_filenames_enabled);
+  ENFORCE_BOOLEAN(show_classes);
+  ENFORCE_TYPE(overhead_filename, T_STRING);
 
-  struct thread_context_collector_state *state;
-  TypedData_Get_Struct(collector_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+  uint16_t max_frame_int = sampling_buffer_check_max_frames(NUM2INT(max_frames));
+
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(self_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
 
   // Update this when modifying state struct
-  state->max_frames = sampling_buffer_check_max_frames(NUM2INT(max_frames));
-  state->locations = ruby_xcalloc(state->max_frames, sizeof(ddog_prof_Location));
-  // hash_map_per_thread_context is already initialized, nothing to do here
+  sample_locations_initialize(&state->locations, max_frame_int, show_classes == Qtrue);
   state->recorder_instance = enforce_recorder_instance(recorder_instance);
+  recorder_install_on_serialize(recorder_instance, self_instance);
   state->endpoint_collection_enabled = (endpoint_collection_enabled == Qtrue);
-  state->timeline_enabled = (timeline_enabled == Qtrue);
+  state->native_filenames_enabled = (native_filenames_enabled == Qtrue);
+  state->show_classes = (show_classes == Qtrue);
+  state->overhead_filename = overhead_filename;
   if (otel_context_enabled == Qfalse || otel_context_enabled == Qnil) {
-    state->otel_context_enabled = otel_context_enabled_false;
+    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;
+    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;
+    state->otel_context_enabled = OTEL_CONTEXT_ENABLED_BOTH;
   } else {
-    rb_raise(rb_eArgError, "Unexpected value for otel_context_enabled: %+" PRIsVALUE, otel_context_enabled);
+    raise_error(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);
@@ -505,48 +638,104 @@ static VALUE _native_initialize(
     state->tracer_context_key = rb_to_id(tracer_context_key);
   }
 
+  if (thread_begin_tracepoint == Qnil) {
+    thread_begin_tracepoint = rb_tracepoint_new(Qnil, RUBY_EVENT_THREAD_BEGIN, on_thread_begin_event, NULL);
+    rb_tracepoint_enable(thread_begin_tracepoint);
+  }
+
   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");
+static VALUE _native_sample(DDTRACE_UNUSED VALUE _self, VALUE collector_instance, VALUE allow_exception) {
+  ENFORCE_BOOLEAN(allow_exception);
+
+  if (allow_exception == Qfalse) debug_enter_unsafe_context();
+
+  bool needs_otel_span_key =
+    thread_context_collector_sample(collector_instance, monotonic_wall_time_now_ns(RAISE_ON_FAILURE));
+
+  if (allow_exception == Qfalse) debug_leave_unsafe_context();
+
+  // Same as the CpuAndWallTimeWorker does: this can't happen during a sample (and thus not inside the unsafe context)
+  if (needs_otel_span_key) thread_context_collector_resolve_otel_span_key_may_lose_gvl(collector_instance);
 
-  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) {
+  debug_enter_unsafe_context();
+
   thread_context_collector_on_gc_start(collector_instance);
+
+  debug_leave_unsafe_context();
+
   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) {
-  (void) !thread_context_collector_on_gc_finish(collector_instance);
+  debug_enter_unsafe_context();
+
+  (void) !thread_context_collector_on_gc_finish(collector_instance, true);
+
+  debug_leave_unsafe_context();
+
   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, VALUE reset_monotonic_to_system_state) {
-  ENFORCE_BOOLEAN(reset_monotonic_to_system_state);
+static VALUE _native_sample_after_gc(DDTRACE_UNUSED VALUE self, VALUE collector_instance) {
+  debug_enter_unsafe_context();
 
-  struct thread_context_collector_state *state;
-  TypedData_Get_Struct(collector_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+  thread_context_collector_sample_after_gc(collector_instance);
 
-  if (reset_monotonic_to_system_state == Qtrue) {
-    state->time_converter_state = (monotonic_to_system_epoch_state) MONOTONIC_TO_SYSTEM_EPOCH_INITIALIZER;
-  }
+  debug_leave_unsafe_context();
 
-  thread_context_collector_sample_after_gc(collector_instance);
   return Qtrue;
 }
 
+// Record profiler sampling overhead as a placeholder stack
+static void record_sampling_overhead(thread_context_collector_state *state, per_thread_context *current_thread_context) {
+  long wall_time_after_sampling = monotonic_wall_time_now_ns(RAISE_ON_FAILURE);
+  long cpu_time_after_sampling = cpu_time_now_ns(current_thread_context);
+
+  long overhead_cpu_time_ns = update_cpu_time_since_previous_sample(current_thread_context, cpu_time_after_sampling);
+  long overhead_wall_time_ns = update_wall_time_since_previous_sample(current_thread_context, wall_time_after_sampling);
+
+  ddog_prof_Label overhead_labels[] = {
+    {.key = DDOG_CHARSLICE_C("thread id"), .str = DDOG_CHARSLICE_C("0"), .num = 0},
+    {.key = DDOG_CHARSLICE_C("thread name"), .str = DDOG_CHARSLICE_C("Datadog::Profiling::Sampling"), .num = 0},
+    {.key = DDOG_CHARSLICE_C("state"), .str = DDOG_CHARSLICE_C("had cpu"), .num = 0},
+    {.key = DDOG_CHARSLICE_C("profiler overhead"), .num = 1},
+  };
+
+  int64_t end_timestamp_ns = monotonic_to_system_epoch_ns(&state->time_converter_state, wall_time_after_sampling);
+
+  ddog_prof_Location overhead_location = {
+    .mapping = {.filename = DDOG_CHARSLICE_C(""), .build_id = DDOG_CHARSLICE_C(""), .build_id_id = {}},
+    .function = {
+      .name = DDOG_CHARSLICE_C("sampling"),
+      .filename = char_slice_from_ruby_string(state->overhead_filename),
+    },
+    .line = 0,
+  };
+
+  record_sample(
+    state->recorder_instance,
+    (ddog_prof_Slice_Location) {.ptr = &overhead_location, .len = 1},
+    (sample_values) {.cpu_time_ns = overhead_cpu_time_ns, .cpu_or_wall_samples = 1, .wall_time_ns = overhead_wall_time_ns},
+    (sample_labels) {
+      .labels = (ddog_prof_Slice_Label) {.ptr = overhead_labels, .len = sizeof(overhead_labels) / sizeof(overhead_labels[0])},
+      .state_label = NULL,
+      .end_timestamp_ns = end_timestamp_ns,
+    }
+  );
+}
+
 // 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.
@@ -554,92 +743,79 @@ static VALUE _native_sample_after_gc(DDTRACE_UNUSED VALUE self, VALUE collector_
 // 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).
+// Assumption 6: When called while the profiler is active, `during_sample` MUST be set.
 //
-// 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);
+bool thread_context_collector_sample(VALUE self_instance, long current_monotonic_wall_time_ns) {
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(self_instance, 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);
+  per_thread_context *current_thread_context = get_or_create_context_for(current_thread);
+
+  // Sample the current thread (which will use the current CPU-time) first.
+  // This ensures any CPU-time the profiler spends sampling (e.g. the rest of the function) is not wrongly blamed on this thread,
+  // and instead is separately accounted by `record_sampling_overhead()` below.
+  update_metrics_and_sample(
+    state,
+    current_thread,
+    current_thread_context,
+    current_monotonic_wall_time_ns,
+    false);
 
   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);
+    VALUE thread = rb_ary_entry(threads, i);
+    ENFORCE_THREAD(thread);
 
-    // 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;
+    if (thread == current_thread) continue; // Already sampled above
 
     update_metrics_and_sample(
       state,
-      /* thread_being_sampled: */ thread,
-      /* stack_from_thread: */ thread,
-      thread_context,
-      thread_context->sampling_buffer,
-      current_cpu_time_ns,
-      current_monotonic_wall_time_ns
-    );
+      thread,
+      get_or_create_context_for(thread),
+      current_monotonic_wall_time_ns,
+      false);
   }
 
-  state->sample_count++;
+  state->stats.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);
+  // If the current thread is a profiler-internal thread, we don't use record_sampling_overhead()
+  // and accept the sampling overhead is attributed to the profiler-internal thread
+  // (both are under the same datadog group in the flamegraph).
+  // The reason we need to do this is profiler-internal threads are skipped during per-tick sampling
+  // so their timestamps are not updated. record_sampling_overhead() would see the stale previous-sample timestamps and
+  // attribute the entire sleep interval incorrectly as overhead instead of just the sampling work.
+  if (!current_thread_context->is_profiler_internal_thread) {
+    record_sampling_overhead(state, current_thread_context);
+  }
 
-  update_metrics_and_sample(
-    state,
-    /* thread_being_sampled: */ current_thread,
-    /* stack_from_thread: */ profiler_overhead_stack_thread,
-    current_thread_context,
-    // Here we use the overhead thread's sampling buffer so as to not invalidate the cache in the buffer of the thread being sampled
-    get_or_create_context_for(profiler_overhead_stack_thread, state)->sampling_buffer,
-    cpu_time_now_ns(current_thread_context),
-    monotonic_wall_time_now_ns(RAISE_ON_FAILURE)
-  );
+  return is_otel_current_span_key_needed(state);
 }
 
 static void update_metrics_and_sample(
-  struct thread_context_collector_state *state,
+  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,
-  sampling_buffer* sampling_buffer,
-  long current_cpu_time_ns,
-  long current_monotonic_wall_time_ns
+  per_thread_context *thread_context,
+  long current_monotonic_wall_time_ns,
+  bool force_sample
 ) {
+  long current_cpu_time_ns = cpu_time_now_ns(thread_context);
   bool is_gvl_waiting_state =
-    handle_gvl_waiting(state, thread_being_sampled, stack_from_thread, thread_context, sampling_buffer, current_cpu_time_ns);
+    handle_gvl_waiting(state, thread_being_sampled, thread_context, current_cpu_time_ns);
+
+  if (skip_sample(state, thread_context, is_gvl_waiting_state, force_sample)) return;
 
   // 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 cpu_time_elapsed_ns = is_gvl_waiting_state ? 0 : update_cpu_time_since_previous_sample(thread_context, current_cpu_time_ns);
 
-  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
-  );
+  long wall_time_elapsed_ns = update_wall_time_since_previous_sample(thread_context, current_monotonic_wall_time_ns);
 
   // 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)
+  // 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
@@ -654,9 +830,7 @@ static void update_metrics_and_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 = wall_time_elapsed_ns},
     current_monotonic_wall_time_ns,
     NULL,
@@ -665,6 +839,37 @@ static void update_metrics_and_sample(
   );
 }
 
+static bool skip_sample(thread_context_collector_state *state, per_thread_context *thread_context, bool is_gvl_waiting_state, bool force_sample) {
+  if (!force_sample && thread_context->is_profiler_internal_thread) {
+    state->stats.profiler_thread_samples_skipped++;
+    return true;
+  }
+
+  // Racy read but harmless, can only cause an extra sample
+  uint64_t gvl_state_change_count = thread_context->gvl_state_change_count;
+
+  // Skip this per-tick sample entirely when the thread does not have the GVL and did not acquire
+  // it since the previous sample: its Ruby-level stack has not changed. The skipped wall-time will
+  // be picked up by either by an extra sample when the thread acquires the GVL, or by
+  // the on-serialize flush in the stack recorder (using was_skipped_at_last_sample).
+  // The check is gated by `!is_gvl_waiting_state` so the existing "Waiting for GVL" machinery
+  // in handle_gvl_waiting (situation 1 extra sample, situation 2 regular sample) keeps running.
+  // TODO: we could probably also skip while "Waiting for GVL"
+  if (!is_gvl_waiting_state &&
+      !force_sample &&
+      (gvl_state_change_count & GVL_SUSPENDED) &&
+      gvl_state_change_count == thread_context->gvl_state_change_count_at_previous_sample) {
+    state->stats.inactive_thread_samples_skipped++;
+    thread_context->was_skipped_at_last_sample = true;
+    return true; // Do NOT update wall_time_at_previous_sample_ns or cpu_time_at_previous_sample_ns
+  } else {
+    // We are going to sample, update the state accordingly:
+    thread_context->gvl_state_change_count_at_previous_sample = gvl_state_change_count;
+    thread_context->was_skipped_at_last_sample = false;
+    return false;
+  }
+}
+
 // 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.
@@ -676,24 +881,27 @@ static void update_metrics_and_sample(
 // 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;
+  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);
+  // This should never fail when the above check passes
+  TypedData_Get_Struct(self_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
 
-  struct per_thread_context *thread_context = get_context_for(rb_thread_current(), state);
+  per_thread_context *thread_context = get_per_thread_context(rb_thread_current());
 
-  // 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)!
+  // Context is created eagerly via on_thread_begin_event, so this should not normally be NULL.
+  // We can't get_or_create_context_for() here since we can't allocate (GC context).
   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);
+  long wall_time_at_start_ns = monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE);
+
+  // If our start timestamp is not OK, we skip tracking this GC as well
+  if (wall_time_at_start_ns == 0) return;
+
+  thread_context->gc_tracking.wall_time_at_start_ns = wall_time_at_start_ns;
   thread_context->gc_tracking.cpu_time_at_start_ns = cpu_time_now_ns(thread_context);
 }
 
@@ -707,23 +915,29 @@ 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).
+// Assumption 3: If allow_prepare_sample is true while profiling is active, during_sample MUST be set by the caller.
 __attribute__((warn_unused_result))
-bool thread_context_collector_on_gc_finish(VALUE self_instance) {
-  struct thread_context_collector_state *state;
+bool thread_context_collector_on_gc_finish(VALUE self_instance, bool allow_prepare_sample) {
+  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);
+  // This should never fail when the above check passes
+  TypedData_Get_Struct(self_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
 
-  struct per_thread_context *thread_context = get_context_for(rb_thread_current(), state);
+  per_thread_context *thread_context = get_per_thread_context(rb_thread_current());
 
-  // 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.
+  // Context is created eagerly, so this should not normally be NULL (see on_gc_start).
   if (thread_context == NULL) return false;
 
+  if (thread_context->prepare_sample_on_gc_finish) {
+    thread_context->prepare_sample_on_gc_finish = false;
+    if (allow_prepare_sample) prepare_sample_thread(rb_thread_current(), &thread_context->sampling_buffer);
+  }
+
   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) {
+  // Both fields are always set and cleared together, so checking one is enough
+  if (cpu_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.
@@ -758,9 +972,9 @@ bool thread_context_collector_on_gc_finish(VALUE self_instance) {
   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`).
+  // Update cpu-time accounting so it doesn't include the cpu-time spent in GC during the next sample.
+  // We don't do the same for wall-time, because GC is just like any other reason a thread didn't make
+  // progress -- time always goes forward regardless of the thread making progress on what it wanted.
   if (thread_context->cpu_time_at_previous_sample_ns != INVALID_TIME) {
     thread_context->cpu_time_at_previous_sample_ns += gc_cpu_time_elapsed_ns;
   }
@@ -786,12 +1000,20 @@ bool thread_context_collector_on_gc_finish(VALUE self_instance) {
 // 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).
+// Assumption 5: When called while the profiler is active, `during_sample` MUST be set.
 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);
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(self_instance, 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");
+    // Rarely, we might be called with nothing to do, as an earlier call already flushed the needed info.
+    // This is because Ruby clears the "pending" flag for an entire batch of postponed jobs BEFORE running any of them
+    // (see `rb_postponed_job_flush`). Thus, if another GC happens while an earlier job in the postponed batch is
+    // running (and other parts of the profiler such as the heap profiler can trigger this)
+    // then `on_gc_finish` records that data and asks for a second flush, even though Ruby has not yet run the
+    // first one. Because of this, Ruby can call us once more with nothing left to do.
+    state->stats.gc_samples_skipped_nothing_to_flush++;
+    return Qnil;
   }
 
   int max_labels_needed_for_gc = 7; // Magic number gets validated inside gc_profiling_set_metadata
@@ -801,11 +1023,7 @@ VALUE thread_context_collector_sample_after_gc(VALUE self_instance) {
   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);
-  }
+  int64_t 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->recorder_instance,
@@ -833,12 +1051,23 @@ VALUE thread_context_collector_sample_after_gc(VALUE self_instance) {
   return Qnil;
 }
 
+// Let the recorder do any cleanup/updates it requires after a GC step.
+//
+// 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.
+VALUE thread_context_collector_heap_update_may_lose_gvl(VALUE self_instance) {
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(self_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  recorder_heap_update_may_lose_gvl(state->recorder_instance);
+
+  return Qnil; // for rb_rescue2
+}
+
 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,
-  sampling_buffer* sampling_buffer,
+  thread_context_collector_state *state,
+  VALUE thread_being_sampled,
+  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?
@@ -849,7 +1078,6 @@ static void trigger_sample_for_thread(
   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
@@ -861,13 +1089,13 @@ static void trigger_sample_for_thread(
     .str = thread_context->thread_id_char_slice
   };
 
-  VALUE thread_name = thread_name_for(thread);
+  VALUE thread_name = thread_name_for(thread_being_sampled);
   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
+  } else if (thread_being_sampled == 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"),
@@ -882,12 +1110,12 @@ 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);
+  trace_identifiers trace_identifiers_result = {.valid = false, .trace_endpoint = Qnil};
+  trace_identifiers_for(state, thread_being_sampled, &trace_identifiers_result);
 
-  if (!trace_identifiers_result.valid && state->otel_context_enabled != otel_context_enabled_false) {
+  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);
+    otel_without_ddtrace_trace_identifiers_for(state, thread_being_sampled, &trace_identifiers_result);
   }
 
   if (trace_identifiers_result.valid) {
@@ -912,13 +1140,6 @@ static void trigger_sample_for_thread(
     }
   }
 
-  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"),
@@ -952,20 +1173,21 @@ static void trigger_sample_for_thread(
   // @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);
+    raise_error(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) {
+  if (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,
-    sampling_buffer,
+    thread_being_sampled,
+    &thread_context->sampling_buffer,
+    state->locations,
     state->recorder_instance,
     values,
     (sample_labels) {
@@ -973,7 +1195,10 @@ static void trigger_sample_for_thread(
       .state_label = state_label,
       .end_timestamp_ns = end_timestamp_ns,
       .is_gvl_waiting_state = is_gvl_waiting_state,
-    }
+    },
+    state->native_filenames_enabled,
+    state->native_filenames_cache,
+    state->show_classes
   );
 }
 
@@ -981,36 +1206,48 @@ static void trigger_sample_for_thread(
 // It SHOULD NOT be used for other purposes.
 static VALUE _native_thread_list(DDTRACE_UNUSED VALUE _self) {
   VALUE result = rb_ary_new();
+
+  debug_enter_unsafe_context();
+
   ddtrace_thread_list(result);
+
+  debug_leave_unsafe_context();
+
   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);
+static void check_frozen_thread(VALUE thread) {
+  if (RB_OBJ_FROZEN(thread)) {
+    raise_error(rb_eFrozenError, "Cannot setup profiler state for Thread %"PRIsVALUE" because it is frozen. Please avoid freezing Thread instances and/or report the issue to dd-trace-rb", thread);
   }
-
-  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;
+// See the docs on struct per_thread_context.
+// This allocates a Ruby object and therefore needs the GVL and is not safe to call from RUBY_INTERNAL_EVENT_* hooks.
+static per_thread_context *get_or_create_context_for(VALUE thread) {
+  per_thread_context *thread_context = get_per_thread_context(thread);
+  if (thread_context != NULL) return thread_context;
 
-  if (st_lookup(state->hash_map_per_thread_context, (st_data_t) thread, &value_context)) {
-    thread_context = (struct per_thread_context*) value_context;
-  }
+  check_frozen_thread(thread);
+
+  thread_context = calloc(1, sizeof(per_thread_context)); // See "note on calloc vs ruby_xcalloc use" in heap_recorder.c
+  initialize_context(thread, thread_context);
+
+  VALUE wrapper = TypedData_Wrap_Struct(rb_cObject, &per_thread_context_typed_data, thread_context);
+  rb_ivar_set(thread, dd_per_thread_context_id, wrapper);
 
+  set_per_thread_context(thread, thread_context);
   return thread_context;
 }
 
+static void on_thread_begin_event(VALUE tracepoint_data, DDTRACE_UNUSED void *unused) {
+  if (!ddtrace_rb_ractor_main_p()) return;
+
+  VALUE thread = rb_tracearg_self(rb_tracearg_from_tracepoint(tracepoint_data));
+  ENFORCE_THREAD(thread);
+  get_or_create_context_for(thread);
+}
+
 #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
@@ -1032,8 +1269,11 @@ static bool is_logging_gem_monkey_patch(VALUE invoke_file_location) {
   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) {
-  thread_context->sampling_buffer = sampling_buffer_new(state->max_frames, state->locations);
+static void initialize_context(VALUE thread, per_thread_context *thread_context) {
+  // We always create per_thread_context's with latest_max_frames; that value is kept in sync with the
+  // active profiler's max_frames by the global reset that runs when profiling starts
+  // so we expect to always see here the latest correct value to be used.
+  sampling_buffer_initialize(&thread_context->sampling_buffer, latest_max_frames);
 
   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)};
@@ -1052,7 +1292,7 @@ static void initialize_context(VALUE thread, struct per_thread_context *thread_c
     } else {
       snprintf(thread_context->thread_invoke_location, THREAD_INVOKE_LOCATION_LIMIT_CHARS, "%s", "(Unnamed thread)");
     }
-  } else if (thread != state->main_thread) {
+  } else if (thread != rb_thread_main()) {
     // 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.
@@ -1066,52 +1306,76 @@ static void initialize_context(VALUE thread, struct per_thread_context *thread_c
 
   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;
+  thread_context->wall_time_at_previous_sample_ns = monotonic_wall_time_now_ns(RAISE_ON_FAILURE);
+  thread_context->cpu_time_at_previous_sample_ns = cpu_time_now_ns(thread_context);
 
   // 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
+  thread_context->gvl_waiting_at = 0;
+  thread_context->gvl_state_change_count = 0;
 }
 
-static void free_context(struct per_thread_context* thread_context) {
-  sampling_buffer_free(thread_context->sampling_buffer);
-  ruby_xfree(thread_context);
+// This MUST be called before profiling starts, so that a new profiler session starts from a fresh state and never
+// observes or includes any leftover stale state from a previous session.
+// Such a call MUST happen while the CpuAndWallTimeWorker is stopped (e.g. no tracepoints active, no signals
+// triggering samples, no gvl hooks, etc).
+//
+// It updates the global `latest_max_frames` from the given (latest) ThreadContext and (re)creates every per-thread
+// context's sampling buffer sized accordingly, so the buffers always match the collector that's about to start
+// sampling -- even if a previous session used a different max_frames.
+void thread_context_collector_reset_all_per_thread_contexts(VALUE self_instance) {
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(self_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  // Update global max frames to be used when allocating sampling buffers
+  if (state->locations.len == 0) {
+    raise_error(rb_eRuntimeError, "BUG: Unexpected locations.len == 0. Is this ThreadContext not initialized?");
+  }
+
+  latest_max_frames = state->locations.len;
+
+  VALUE threads = thread_list(state);
+  const long thread_count = RARRAY_LEN(threads);
+  for (long i = 0; i < thread_count; i++) {
+    VALUE thread = rb_ary_entry(threads, i);
+    ENFORCE_THREAD(thread);
+
+    per_thread_context *thread_context = get_per_thread_context(thread);
+    if (thread_context != NULL) {
+      bool is_profiler_internal_thread = thread_context->is_profiler_internal_thread;
+
+      sampling_buffer_free(&thread_context->sampling_buffer);
+      memset(thread_context, 0, sizeof(per_thread_context));
+      initialize_context(thread, thread_context);
+
+      thread_context->is_profiler_internal_thread = is_profiler_internal_thread;
+    } else {
+      // If thread didn't have a context, let's trigger its creation
+      get_or_create_context_for(thread);
+    }
+  }
 }
 
 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);
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(collector_instance, 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(" max_frames=%d", state->max_frames));
-  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(" max_frames=%d", state->locations.len));
   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(" stats=%"PRIsVALUE, stats_to_ruby_hash(state, rb_hash_new())));
   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(" native_filenames_enabled=%"PRIsVALUE, state->native_filenames_enabled ? Qtrue : Qfalse));
+  rb_str_concat(result, rb_sprintf(" show_classes=%"PRIsVALUE, state->show_classes ? Qtrue : Qfalse));
+  // Note: `st_table_size()` is available from Ruby 3.2+ but not before
+  rb_str_concat(result, rb_sprintf(" native_filenames_cache_size=%zu", state->native_filenames_cache->num_entries));
   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}",
     state->time_converter_state.system_epoch_ns_reference,
@@ -1120,56 +1384,55 @@ 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;
 }
 
-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;
+static VALUE per_thread_context_to_ruby_hash(per_thread_context *thread_context) {
   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),
+    #ifdef __APPLE__
+      ID2SYM(rb_intern("thread_cpu_time_id")),              /* => */ ULL2NUM(thread_context->thread_cpu_time_id.clock_id),
+    #else
+      ID2SYM(rb_intern("thread_cpu_time_id")),              /* => */ CLOCKID2NUM(thread_context->thread_cpu_time_id.clock_id),
+    #endif
     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),
 
-    #ifndef NO_GVL_INSTRUMENTATION
-      ID2SYM(rb_intern("gvl_waiting_at")), /* => */ LONG2NUM(gvl_profiling_state_thread_object_get(thread)),
-    #endif
+    ID2SYM(rb_intern("gvl_waiting_at")), /* => */ LONG2NUM(thread_context->gvl_waiting_at),
+    ID2SYM(rb_intern("gvl_state_change_count")), /* => */ ULL2NUM(thread_context->gvl_state_change_count),
+    ID2SYM(rb_intern("gvl_state_change_count_at_previous_sample")), /* => */ ULL2NUM(thread_context->gvl_state_change_count_at_previous_sample),
+    ID2SYM(rb_intern("was_skipped_at_last_sample")), /* => */ thread_context->was_skipped_at_last_sample ? Qtrue : Qfalse,
+    ID2SYM(rb_intern("prepare_sample_on_gc_finish")), /* => */ thread_context->prepare_sample_on_gc_finish ? Qtrue : Qfalse,
+    ID2SYM(rb_intern("is_profiler_internal_thread")), /* => */ thread_context->is_profiler_internal_thread ? Qtrue : Qfalse,
   };
   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;
+  return context_as_hash;
 }
 
-static VALUE stats_as_ruby_hash(struct thread_context_collector_state *state) {
+static VALUE stats_to_ruby_hash(thread_context_collector_state *state, VALUE hash) {
   // Update this when modifying state struct (stats inner struct)
-  VALUE stats_as_hash = rb_hash_new();
   VALUE arguments[] = {
+    ID2SYM(rb_intern("sample_count")),                             /* => */ UINT2NUM(state->stats.sample_count),
     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),
+    ID2SYM(rb_intern("gc_samples_skipped_nothing_to_flush")),      /* => */ UINT2NUM(state->stats.gc_samples_skipped_nothing_to_flush),
+    ID2SYM(rb_intern("inactive_thread_samples_skipped")),          /* => */ UINT2NUM(state->stats.inactive_thread_samples_skipped),
+    ID2SYM(rb_intern("profiler_thread_samples_skipped")),          /* => */ UINT2NUM(state->stats.profiler_thread_samples_skipped),
   };
-  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;
+  for (long unsigned int i = 0; i < VALUE_COUNT(arguments); i += 2) rb_hash_aset(hash, arguments[i], arguments[i+1]);
+  return hash;
 }
 
-static VALUE gc_tracking_as_ruby_hash(struct thread_context_collector_state *state) {
+static VALUE gc_tracking_as_ruby_hash(thread_context_collector_state *state) {
   // Update this when modifying state struct (gc_tracking inner struct)
   VALUE result = rb_hash_new();
   VALUE arguments[] = {
@@ -1182,72 +1445,88 @@ static VALUE gc_tracking_as_ruby_hash(struct thread_context_collector_state *sta
   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;
-
-  free_context(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.
+// Returns the whole contents of the per_thread_context structs being tracked, by iterating all live threads.
 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);
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(collector_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
 
-  return per_thread_context_st_table_as_ruby_hash(state);
-}
+  VALUE result = rb_hash_new();
+  VALUE threads = thread_list(state);
+  const long thread_count = RARRAY_LEN(threads);
+  for (long i = 0; i < thread_count; i++) {
+    VALUE thread = rb_ary_entry(threads, i);
+    ENFORCE_THREAD(thread);
 
-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;
+    per_thread_context *thread_context = get_per_thread_context(thread);
+    if (thread_context != NULL) {
+      rb_hash_aset(result, thread, per_thread_context_to_ruby_hash(thread_context));
+    }
+  }
+  return result;
+}
 
-  bool is_thread_doing_gc = gc_start_time_ns != INVALID_TIME;
-  long elapsed_time_ns = -1;
+static long update_time_since_previous_sample(long *time_at_previous_sample_ns, long current_time_ns, per_thread_context *thread_context) {
+  // We don't expect to be sampling a thread (and thus updating these counters) while Ruby is doing GC (between
+  // `thread_context_collector_on_gc_start` and `thread_context_collector_on_gc_finish`)
+  if (thread_context->gc_tracking.cpu_time_at_start_ns != INVALID_TIME) {
+    raise_error(
+      rb_eRuntimeError,
+      "BUG: Unexpected sample during GC (thread_id=%s, gc_tracking.cpu_time_at_start_ns=%ld, "
+      "gc_tracking.wall_time_at_start_ns=%ld, monotonic_wall_time_now_ns=%ld)",
+      thread_context->thread_id,
+      thread_context->gc_tracking.cpu_time_at_start_ns,
+      thread_context->gc_tracking.wall_time_at_start_ns,
+      monotonic_wall_time_now_ns(RAISE_ON_FAILURE)
+    );
+  }
 
-  if (is_thread_doing_gc) {
-    bool previous_sample_was_during_gc = gc_start_time_ns <= *time_at_previous_sample_ns;
+  long 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 (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
-    }
+  return elapsed_time_ns;
+}
 
-    // 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;
+static long update_cpu_time_since_previous_sample(per_thread_context *thread_context, long current_cpu_time_ns) {
+  // A previous `cpu_time_now_ns` may have failed to read the clock and thus invalidated our baseline; in that case we
+  // use the current time, so this sample gets no cpu-time and the next one gets an accurate delta.
+  // Note wall-time never needs this: it is always seeded with a real value in `initialize_context`.
+  if (thread_context->cpu_time_at_previous_sample_ns == INVALID_TIME) {
+    thread_context->cpu_time_at_previous_sample_ns = current_cpu_time_ns;
   }
 
+  long elapsed_time_ns = update_time_since_previous_sample(
+    &thread_context->cpu_time_at_previous_sample_ns,
+    current_cpu_time_ns,
+    thread_context
+  );
+
+  // We don't expect cpu-time to go backwards, so let's flag this as a bug
   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");
-    }
+    raise_error(rb_eRuntimeError, "BUG: Unexpected CPU time going backwards between samples");
   }
 
   return elapsed_time_ns;
 }
 
+static long update_wall_time_since_previous_sample(per_thread_context *thread_context, long current_wall_time_ns) {
+  long elapsed_time_ns = update_time_since_previous_sample(
+    &thread_context->wall_time_at_previous_sample_ns,
+    current_wall_time_ns,
+    thread_context
+  );
+
+  // 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.
+  return long_max_of(elapsed_time_ns, 0);
+}
+
 // Safety: This function is assumed never to raise exceptions by callers
-static long cpu_time_now_ns(struct per_thread_context *thread_context) {
+static long cpu_time_now_ns(per_thread_context *thread_context) {
   thread_cpu_time cpu_time = thread_cpu_time_for(thread_context->thread_cpu_time_id);
 
   if (!cpu_time.valid) {
@@ -1278,31 +1557,45 @@ static long thread_id_for(VALUE thread) {
 }
 
 VALUE enforce_thread_context_collector_instance(VALUE object) {
-  Check_TypedStruct(object, &thread_context_collector_typed_data);
+  ENFORCE_TYPED_DATA(object, &thread_context_collector_typed_data);
   return object;
 }
 
+// Commit any pending heap allocation recordings.
+// Recording an allocation only puts it on a pending list, because taking the weak reference the heap profiler needs
+// can't be done from inside the NEWOBJ tracepoint -- see `heap_recorder_commit_recordings_may_lose_gvl`.
+void thread_context_collector_commit_heap_recordings_may_lose_gvl(VALUE self_instance) {
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(self_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  recorder_commit_heap_recordings_may_lose_gvl(state->recorder_instance);
+}
+
 // 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);
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(collector_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
 
-  return stats_as_ruby_hash(state);
+  return stats_to_ruby_hash(state, rb_hash_new());
 }
 
 // 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);
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(collector_instance, 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->otel_context_enabled == otel_context_enabled_only) return;
+static void trace_identifiers_for(
+  thread_context_collector_state *state,
+  VALUE thread,
+  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);
@@ -1320,7 +1613,9 @@ static void trace_identifiers_for(struct thread_context_collector_state *state,
 
   VALUE numeric_span_id = Qnil;
 
-  if (otel_values != Qnil) ddtrace_otel_trace_identifiers_for(state, &active_trace, &root_span, &numeric_span_id, active_span, otel_values);
+  if (otel_values != Qnil) {
+    ddtrace_otel_trace_identifiers_for(state, &active_trace, &root_span, &numeric_span_id, active_span, otel_values);
+  }
 
   if (root_span == Qnil || (active_span == Qnil && numeric_span_id == Qnil)) return;
 
@@ -1376,34 +1671,62 @@ static bool should_collect_resource(VALUE root_span) {
 //
 // Assumption: This method gets called BEFORE restarting profiling -- e.g. there are no components attempting to
 // trigger samples at the same time.
+//
+// Note that tests call this method directly in the same process without forking,
+// and in such a case non-current Threads keep running.
 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);
-
-  // 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);
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(collector_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
 
   state->stats = (struct stats) {}; // Resets all stats back to zero
 
+  // No need to clean-up the per-thread context because the CpuAndWallTimeWorker always cleans
+  // it up unconditionally on every start/restart and that includes after a fork.
+
   rb_funcall(state->recorder_instance, rb_intern("reset_after_fork"), 0);
 
   return Qtrue;
 }
 
-static VALUE thread_list(struct thread_context_collector_state *state) {
+static VALUE thread_list(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);
-
+// Inside a signal handler, we don't want to do the whole work of recording a sample, but we only record the stack of
+// the current thread.
+//
+// Assumptions for this function are same as for `thread_context_collector_sample` except that this function is
+// expected to be called from a signal handler and to be async-signal-safe, and `during_sample` MUST be unset.
+//
+// Also, no allocation (Ruby or malloc) can happen.
+bool thread_context_collector_prepare_sample_inside_signal_handler(void) {
   VALUE current_thread = rb_thread_current();
+  per_thread_context *thread_context = get_per_thread_context(current_thread);
+  if (thread_context == NULL) return false;
+
+  return prepare_sample_thread(current_thread, &thread_context->sampling_buffer);
+}
+
+// Called from the sampling signal handler; must not allocate or inspect the Ruby stack during GC.
+void thread_context_collector_request_prepare_on_gc_finish(void) {
+  per_thread_context *thread_context = get_per_thread_context(rb_thread_current());
+  if (thread_context != NULL) thread_context->prepare_sample_on_gc_finish = true;
+}
+
+// This method gets called from inside the RUBY_INTERNAL_EVENT_NEWOBJ tracepoint so it should neither allocate in the
+// Ruby heap nor release the GVL (https://github.com/DataDog/dd-trace-rb/pull/4240).
+//
+// Returns true if `thread_context_collector_commit_heap_recordings_may_lose_gvl` needs to be called (to do work
+// that can't be done from inside the tracepoint, such as allocate new objects), and false if it doesn't
+//
+// Assumption 1: The callers must ensure thread_context is non-NULL.
+// Assumption 2: When called while the profiler is active, `during_sample` MUST be set.
+bool thread_context_collector_sample_allocation(VALUE self_instance, per_thread_context *thread_context, unsigned int sample_weight, VALUE new_object) {
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(self_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
 
   enum ruby_value_type type = rb_type(new_object);
 
@@ -1412,93 +1735,101 @@ void thread_context_collector_sample_allocation(VALUE self_instance, unsigned in
 
   // 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_class2name(klass);
-        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);
-        }
+  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) {
+      VALUE name = ddtrace_alloc_free_rb_mod_name(klass);
+      if (name != Qnil) {
+        class_name = char_slice_from_ruby_string(name);
       } 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
+      // Fallback for objects with no class. Objects with no class are a way for the Ruby VM to mark them
+      // as internal objects; see rb_objspace_internal_object_p for details.
+      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);
+  VALUE current_thread = rb_thread_current();
 
-  struct per_thread_context *thread_context = get_or_create_context_for(current_thread, state);
+  heap_sample_values heap_sample = {.new_object = new_object, .alloc_class = class_name};
 
   trigger_sample_for_thread(
     state,
-    /* thread: */  current_thread,
-    /* stack_from_thread: */ current_thread,
+    current_thread,
     thread_context,
-    thread_context->sampling_buffer,
-    (sample_values) {.alloc_samples = sample_weight, .alloc_samples_unscaled = 1, .heap_sample = true},
+    (sample_values) {.alloc_samples = sample_weight, .alloc_samples_unscaled = 1, .heap_sample = &heap_sample},
     INVALID_TIME, // For now we're not collecting timestamps for allocation events, as per profiling team internal discussions
     &ruby_vm_type,
-    optional_class_name,
-    false
+    &class_name,
+    /* is_gvl_waiting_state: */ false
   );
+
+  return heap_sample.out_needs_commit;
 }
 
 // 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;
+  per_thread_context *thread_context = get_per_thread_context(rb_thread_current());
+  if (thread_context == NULL) {
+    rb_raise(rb_eRuntimeError, "Missing per_thread_context for current thread in _native_sample_allocation");
+  }
+
+  debug_enter_unsafe_context();
+
+  bool needs_commit = thread_context_collector_sample_allocation(collector_instance, thread_context, NUM2UINT(sample_weight), new_object);
+
+  debug_leave_unsafe_context();
+
+  // We could instead choose to automatically trigger the after allocation here; yet, it seems kinda nice to keep it manual for
+  // the tests so we can pull on each lever separately and observe "the sausage being made" in steps
+  return needs_commit ? Qtrue : Qfalse;
 }
 
-static VALUE new_empty_thread_inner(DDTRACE_UNUSED void *arg) { return Qnil; }
+static VALUE new_empty_thread_inner(DDTRACE_UNUSED void *arg) {
+  rb_thread_sleep(INT_MAX);
+  return Qnil;
+}
 
 // This method exists only to enable testing Datadog::Profiling::Collectors::ThreadContext behavior using RSpec.
 // It SHOULD NOT be used for other purposes.
@@ -1542,23 +1873,61 @@ static VALUE read_otel_current_span_key_const(DDTRACE_UNUSED VALUE _unused) {
   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 == 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);
+static VALUE otel_current_span_key_not_found(DDTRACE_UNUSED VALUE _unused, DDTRACE_UNUSED VALUE _exception) {
+  return Qnil;
+}
 
-    // 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;
+static VALUE get_otel_current_span_key(thread_context_collector_state *state) {
+  if (state->otel_current_span_key == OTEL_CURRENT_SPAN_KEY_NOT_EXTRACTED) {
+    // Extracting this needs calling into Ruby code, which we don't want to do in the middle of a sample.
+    // So instead we signal that we want this, and let it be resolved separately.
+    state->otel_current_span_key = OTEL_CURRENT_SPAN_KEY_EXTRACTION_NEEDED;
+    return Qnil;
   }
 
+  // We asked for it, but it hasn't been extracted yet
+  if (is_otel_current_span_key_needed(state)) return Qnil;
+
   return state->otel_current_span_key;
 }
 
+static bool is_otel_current_span_key_needed(thread_context_collector_state *state) {
+  return state->otel_current_span_key == OTEL_CURRENT_SPAN_KEY_EXTRACTION_NEEDED;
+}
+
+// Extracts `otel_current_span_key`, if a sample asked for it (see `get_otel_current_span_key`).
+// This is deliberately not done during sampling because we can lose the GVL.
+//
+// Assumption 1: This function is called in a thread that is holding the Global VM Lock. Caller is responsible for enforcing this.
+void thread_context_collector_resolve_otel_span_key_may_lose_gvl(VALUE self_instance) {
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(self_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  if (!is_otel_current_span_key_needed(state)) return;
+
+  // 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).
+  //
+  // Note that we use `rb_rescue2` and not `rb_protect` so that we only swallow exceptions: anything else that can
+  // unwind through here (e.g. a `Thread#kill`) is not ours to swallow. `rb_rescue2` also takes care of restoring
+  // any previously-pending exception for us.
+  VALUE span_key = rb_rescue2(
+    read_otel_current_span_key_const,
+    Qnil,
+    otel_current_span_key_not_found,
+    Qnil,
+    rb_eException, // rb_eException is the base class of all Ruby exceptions
+    0 // Required by API to be the last argument
+  );
+
+  // 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;
+}
+
 // 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,
+  thread_context_collector_state *state,
   VALUE *active_trace,
   VALUE *root_span,
   VALUE *numeric_span_id,
@@ -1582,7 +1951,7 @@ static void ddtrace_otel_trace_identifiers_for(
   // trace and span representing it. Each ddtrace trace is then connected to the previous otel span, forming a linked
   // list. The local root span is going to be the trace/span we find at the end of this linked list.
   while (otel_values != Qnil) {
-    VALUE otel_span = rb_hash_lookup(otel_values, otel_current_span_key);
+    VALUE otel_span = safely_lookup_hash_without_going_into_ruby_code(otel_values, otel_current_span_key);
     if (otel_span == Qnil) break;
     VALUE next_trace = rb_ivar_get(otel_span, at_datadog_trace_id);
     if (next_trace == Qnil) break;
@@ -1600,9 +1969,10 @@ static void ddtrace_otel_trace_identifiers_for(
   *numeric_span_id = resolved_numeric_span_id;
 }
 
+// Assumption 1: When called while the profiler is active, `during_sample` MUST be set.
 void thread_context_collector_sample_skipped_allocation_samples(VALUE self_instance, unsigned int skipped_samples) {
-  struct thread_context_collector_state *state;
-  TypedData_Get_Struct(self_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(self_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
 
   ddog_prof_Label labels[] = {
     // Providing .num = 0 should not be needed but the tracer-2.7 docker image ships a buggy gcc that complains about this
@@ -1625,10 +1995,51 @@ void thread_context_collector_sample_skipped_allocation_samples(VALUE self_insta
 }
 
 static VALUE _native_sample_skipped_allocation_samples(DDTRACE_UNUSED VALUE self, VALUE collector_instance, VALUE skipped_samples) {
+  debug_enter_unsafe_context();
+
   thread_context_collector_sample_skipped_allocation_samples(collector_instance, NUM2UINT(skipped_samples));
+
+  debug_leave_unsafe_context();
+
   return Qtrue;
 }
 
+#ifndef NO_CURRENT_FIBER_FOR // Ruby 3.1+
+  static VALUE otel_context_storage_for(thread_context_collector_state *state, VALUE thread) {
+    if (state->otel_context_source == OTEL_CONTEXT_SOURCE_FIBER_IVAR) { // otel-api 1.5+
+      VALUE current_fiber = current_fiber_for(thread);
+      return current_fiber == Qnil ? Qnil : rb_ivar_get(current_fiber, otel_fiber_context_storage_id /* @opentelemetry_context */);
+    }
+
+    if (state->otel_context_source == OTEL_CONTEXT_SOURCE_FIBER_LOCAL) { // otel-api < 1.5
+      return rb_thread_local_aref(thread, otel_context_storage_id /* __opentelemetry_context_storage__ */);
+    }
+
+    // If we got here, it means we never observed a context being set. Let's probe which one to use.
+    VALUE current_fiber = current_fiber_for(thread);
+    if (current_fiber != Qnil) {
+      VALUE context_storage = rb_ivar_get(current_fiber, otel_fiber_context_storage_id /* @opentelemetry_context */);
+      if (context_storage != Qnil) {
+        state->otel_context_source = OTEL_CONTEXT_SOURCE_FIBER_IVAR; // Remember for next time
+        return context_storage;
+      }
+    } else {
+      VALUE context_storage = rb_thread_local_aref(thread, otel_context_storage_id /* __opentelemetry_context_storage__ */);
+      if (context_storage != Qnil) {
+        state->otel_context_source = OTEL_CONTEXT_SOURCE_FIBER_LOCAL; // Remember for next time
+        return context_storage;
+      }
+    }
+
+    // There's no context storage attached to the current thread
+    return Qnil;
+  }
+#else
+  static inline VALUE otel_context_storage_for(DDTRACE_UNUSED thread_context_collector_state *state, VALUE thread) {
+    return rb_thread_local_aref(thread, otel_context_storage_id /* __opentelemetry_context_storage__ */);
+  }
+#endif
+
 // 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.
 //
@@ -1651,11 +2062,11 @@ static VALUE _native_sample_skipped_allocation_samples(DDTRACE_UNUSED VALUE self
 //    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,
+  thread_context_collector_state *state,
   VALUE thread,
-  struct trace_identifiers *trace_identifiers_result
+  trace_identifiers *trace_identifiers_result
 ) {
-  VALUE context_storage = rb_thread_local_aref(thread, otel_context_storage_id /* __opentelemetry_context_storage__ */);
+  VALUE context_storage = otel_context_storage_for(state, thread);
 
   // 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;
@@ -1663,17 +2074,17 @@ static void otel_without_ddtrace_trace_identifiers_for(
   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;
+  long 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);
+  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;
+  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);
+  for (long i = active_context_index - 1; i >= 0; i--) {
+    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;
@@ -1693,7 +2104,7 @@ static void otel_without_ddtrace_trace_identifiers_for(
 
   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;
+  if (root_span_type == Qnil || !RB_TYPE_P(root_span_type, T_SYMBOL) || !RB_STATIC_SYM_P(root_span_type) || 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;
@@ -1701,8 +2112,8 @@ static void otel_without_ddtrace_trace_identifiers_for(
   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};
+static otel_span otel_span_from(VALUE otel_context, VALUE otel_current_span_key) {
+  otel_span failed = {.span = Qnil, .span_id = Qnil, .trace_id = Qnil};
 
   if (otel_context == Qnil) return failed;
 
@@ -1710,7 +2121,7 @@ static struct otel_span otel_span_from(VALUE otel_context, VALUE otel_current_sp
   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);
+  VALUE span = safely_lookup_hash_without_going_into_ruby_code(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)
@@ -1721,7 +2132,7 @@ static struct otel_span otel_span_from(VALUE otel_context, VALUE otel_current_sp
   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};
+  return (otel_span) {.span = span, .span_id = span_id, .trace_id = trace_id};
 }
 
 // Otel span ids are represented as a big-endian 8-byte string
@@ -1741,36 +2152,149 @@ static uint64_t otel_span_id_to_uint(VALUE otel_span_id) {
     ((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;
+void thread_context_collector_stats(VALUE self_instance, VALUE stats_hash) {
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(self_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
+  stats_to_ruby_hash(state, stats_hash);
+}
+
+void thread_context_collector_stats_reset_not_thread_safe(VALUE self_instance) {
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(self_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
+  state->stats = (struct stats) {};
+}
+
+static void mark_thread_as_profiler_internal(per_thread_context *ctx) {
+  ctx->is_profiler_internal_thread = true;
+}
 
-    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;
+void thread_context_collector_profiler_internal_thread_started(void) {
+  per_thread_context *ctx = get_or_create_context_for(rb_thread_current());
+  mark_thread_as_profiler_internal(ctx);
+}
+
+static VALUE _native_mark_thread_as_profiler_internal(DDTRACE_UNUSED VALUE self, VALUE thread) {
+  per_thread_context *ctx = get_or_create_context_for(thread);
+  mark_thread_as_profiler_internal(ctx);
+  return Qnil;
+}
+
+// Called via rb_ensure when a profiler-internal thread (worker or idle helper) is about to exit.
+// Records a final sample so the thread's accumulated cpu/wall time since the last on_serialize
+// flush is not lost. on_serialize (below) also flushes profiler-internal threads during periodic
+// serialization, but it can't help at shutdown: by the time the final serialize runs, these
+// threads are already dead and absent from thread_list.
+//
+// Assumption 1: When called while the profiler is active, `during_sample` MUST be set.
+VALUE thread_context_collector_profiler_internal_thread_done(VALUE self_instance) {
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(self_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
 
-    gvl_profiling_state_set(thread, current_monotonic_wall_time_ns);
+  VALUE current_thread = rb_thread_current();
+  per_thread_context *thread_context = get_or_create_context_for(current_thread);
+  if (!thread_context->is_profiler_internal_thread) {
+    rb_raise(rb_eRuntimeError, "current thread %"PRIsVALUE" is not profiler-internal thread", current_thread);
   }
 
-  // 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);
+  long current_monotonic_wall_time_ns = monotonic_wall_time_now_ns(RAISE_ON_FAILURE);
+
+  update_metrics_and_sample(
+    state,
+    current_thread,
+    thread_context,
+    current_monotonic_wall_time_ns,
+    true);
+
+  return Qnil;
+}
 
-    // Thread was not being profiled / not waiting on gvl
-    if (gvl_waiting_at == 0 || gvl_waiting_at == GVL_WAITING_ENABLED_EMPTY) return false;
+// Flushes threads whose last per-tick sample was skipped (either by the SUSPENDED-skip
+// optimization, or by is_profiler_internal_thread) so their accumulated time is recorded.
+// Called by the stack recorder at the start of _native_serialize (regular periodic flush).
+//
+// Assumption 1: When called while the profiler is active, `during_sample` MUST be set.
+void thread_context_collector_on_serialize(VALUE self_instance) {
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(self_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
 
+  long current_monotonic_wall_time_ns = monotonic_wall_time_now_ns(RAISE_ON_FAILURE);
+  VALUE threads = thread_list(state);
+  const long thread_count = RARRAY_LEN(threads);
+
+  for (long i = 0; i < thread_count; i++) {
+    VALUE thread = rb_ary_entry(threads, i);
+    ENFORCE_THREAD(thread);
+
+    per_thread_context *thread_context = get_per_thread_context(thread);
+
+    if (thread_context != NULL && (thread_context->was_skipped_at_last_sample || thread_context->is_profiler_internal_thread)) {
+      update_metrics_and_sample(
+        state,
+        thread,
+        thread_context,
+        current_monotonic_wall_time_ns,
+        true // We need to force_sample=true otherwise this sample would be skipped too
+      );
+    }
+  }
+}
+
+void thread_context_collector_on_gvl_released(per_thread_context *thread_context) {
+  thread_context->gvl_state_change_count |= GVL_SUSPENDED;
+}
+
+void thread_context_collector_on_gvl_waiting(per_thread_context *thread_context) {
+  long current_monotonic_wall_time_ns = monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE);
+  if (current_monotonic_wall_time_ns <= 0) return;
+
+  thread_context->gvl_waiting_at = current_monotonic_wall_time_ns;
+}
+
+static VALUE _native_on_gvl_waiting(DDTRACE_UNUSED VALUE self, VALUE thread) {
+  ENFORCE_THREAD(thread);
+
+  debug_enter_unsafe_context();
+
+  per_thread_context *thread_context = get_per_thread_context(thread);
+  if (thread_context) thread_context_collector_on_gvl_waiting(thread_context);
+
+  debug_leave_unsafe_context();
+
+  return Qnil;
+}
+
+static VALUE _native_on_gvl_released(DDTRACE_UNUSED VALUE self, VALUE thread) {
+  ENFORCE_THREAD(thread);
+
+  debug_enter_unsafe_context();
+
+  per_thread_context *thread_context = get_per_thread_context(thread);
+  if (thread_context) thread_context_collector_on_gvl_released(thread_context);
+
+  debug_leave_unsafe_context();
+
+  return Qnil;
+}
+
+#ifndef NO_GVL_INSTRUMENTATION
+  // We must only use async-signal-safe functions here, see notes in the caller
+  //
+  // Assumption 1: When called while the profiler is active, `during_sample` MUST be set.
+  __attribute__((warn_unused_result))
+  on_gvl_running_result thread_context_collector_on_gvl_running(VALUE thread, per_thread_context *thread_context, uint32_t waiting_for_gvl_threshold_ns) {
+    // Bump the event counter and clears the state bit to "running"
+    uint64_t counter_portion = thread_context->gvl_state_change_count >> 1;
+    thread_context->gvl_state_change_count = ((counter_portion + 1) << 1) | GVL_RUNNING;
+
+    long gvl_waiting_at = thread_context->gvl_waiting_at;
+    // Thread was not waiting on gvl
+    if (gvl_waiting_at == 0) {
+      return (on_gvl_running_result) {.action = ON_GVL_RUNNING_UNKNOWN, .waiting_for_gvl_duration_ns = 0};
+    }
     // @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;
+    if (gvl_waiting_at < 0) {
+      return (on_gvl_running_result) {.action = ON_GVL_RUNNING_SAMPLE, .waiting_for_gvl_duration_ns = 0};
+    }
 
     long waiting_for_gvl_duration_ns = monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE) - gvl_waiting_at;
 
@@ -1778,20 +2302,32 @@ static uint64_t otel_span_id_to_uint(VALUE otel_span_id) {
 
     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;
+      long gvl_waiting_at_is_now_running = -gvl_waiting_at;
 
-      gvl_profiling_state_set(thread, gvl_waiting_at_is_now_running);
+      thread_context->gvl_waiting_at = 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);
+      thread_context->gvl_waiting_at = 0;
+
+      // Even though the GVL wait itself was below threshold, if the thread had skipped samples
+      // (was suspended for a long time without the GVL), we still need to force a sample now.
+      // Otherwise, the accumulated idle wall-time would be reported against whatever stack the
+      // thread runs next, misrepresenting the time spent idle.
+      if (thread_context->was_skipped_at_last_sample) {
+        should_sample = true;
+      }
     }
 
-    return should_sample;
-  }
+    if (should_sample) {
+      // We prepare the sample here because the postponed job might be called some time later,
+      // possibly after some Ruby calls which change the Ruby stack,
+      // and we want to attribute the time acquiring or without the GVL to the correct Ruby stack.
+      prepare_sample_thread(thread, &thread_context->sampling_buffer);
+    }
 
-  __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);
+    return (on_gvl_running_result) {
+      .action = should_sample ? ON_GVL_RUNNING_SAMPLE : ON_GVL_RUNNING_DONT_SAMPLE,
+      .waiting_for_gvl_duration_ns = waiting_for_gvl_duration_ns,
+    };
   }
 
   // Why does this method need to exist?
@@ -1811,7 +2347,7 @@ static uint64_t otel_span_id_to_uint(VALUE otel_span_id) {
   //
   // 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
+  // via the regular cpu/wall-time samples mechanism, the stack can 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
@@ -1820,17 +2356,20 @@ static uint64_t otel_span_id_to_uint(VALUE otel_span_id) {
   //
   // ---
   //
+  // Assumption 1: Always called with the GVL, either from a postponed_job or from tests.
+  // Assumption 2: When called while the profiler is active, `during_sample` MUST be set.
+  //
   // 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);
+  VALUE thread_context_collector_sample_after_gvl_running(VALUE self_instance, VALUE current_thread, long current_monotonic_wall_time_ns) {
+    thread_context_collector_state *state;
+    TypedData_Get_Struct(self_instance, 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");
+    per_thread_context *thread_context = get_or_create_context_for(current_thread);
 
-    intptr_t gvl_waiting_at = gvl_profiling_state_thread_object_get(current_thread);
+    long gvl_waiting_at = thread_context->gvl_waiting_at;
 
-    if (gvl_waiting_at >= 0) {
+    if (gvl_waiting_at >= 0 && !thread_context->was_skipped_at_last_sample) {
       // @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.
@@ -1838,46 +2377,30 @@ static uint64_t otel_span_id_to_uint(VALUE otel_span_id) {
       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,
+      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
-  }
+      current_monotonic_wall_time_ns,
+      false);
 
-  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());
+    return Qtrue;
   }
 
   // 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,
+    thread_context_collector_state *state,
     VALUE thread_being_sampled,
-    VALUE stack_from_thread,
-    struct per_thread_context *thread_context,
-    sampling_buffer* sampling_buffer,
+    per_thread_context *thread_context,
     long current_cpu_time_ns
   ) {
-    intptr_t gvl_waiting_at = gvl_profiling_state_thread_object_get(thread_being_sampled);
+    long gvl_waiting_at = thread_context->gvl_waiting_at;
 
-    bool is_gvl_waiting_state = gvl_waiting_at != 0 && gvl_waiting_at != GVL_WAITING_ENABLED_EMPTY;
+    bool is_gvl_waiting_state = gvl_waiting_at != 0;
 
     if (!is_gvl_waiting_state) return false;
 
@@ -1890,17 +2413,17 @@ static uint64_t otel_span_id_to_uint(VALUE otel_span_id) {
     //  ...──────────────┬───────────────────...
     //       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:
+    //   to represent the remaining cpu/wall time before the "Waiting for GVL" started (for timeline purposes):
     //
     //                             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
@@ -1910,7 +2433,7 @@ static uint64_t otel_span_id_to_uint(VALUE otel_span_id) {
     //  ...──────────────┬───────────────────────────────────────────────...
     //       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.
@@ -1925,91 +2448,153 @@ static uint64_t otel_span_id_to_uint(VALUE otel_span_id) {
 
     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);
+      thread_context->gvl_waiting_at = 0;
     }
 
     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 cpu_time_elapsed_ns = update_cpu_time_since_previous_sample(thread_context, current_cpu_time_ns);
 
-      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
-      );
+      long duration_until_start_of_gvl_waiting_ns =
+        update_wall_time_since_previous_sample(thread_context, gvl_waiting_started_wall_time_ns);
 
       // 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
+        /* is_gvl_waiting_state: */ 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) {
+  static VALUE _native_gvl_waiting_at_for(DDTRACE_UNUSED VALUE self, VALUE thread) {
     ENFORCE_THREAD(thread);
 
-    thread_context_collector_on_gvl_waiting(thread_from_thread_object(thread));
-    return Qnil;
-  }
+    debug_enter_unsafe_context();
 
-  static VALUE _native_gvl_waiting_at_for(DDTRACE_UNUSED VALUE self, VALUE thread) {
-    ENFORCE_THREAD(thread);
+    per_thread_context *thread_context = get_per_thread_context(thread);
+    VALUE result = thread_context ? LONG2NUM(thread_context->gvl_waiting_at) : Qnil;
+
+    debug_leave_unsafe_context();
 
-    intptr_t gvl_waiting_at = gvl_profiling_state_thread_object_get(thread);
-    return LONG2NUM(gvl_waiting_at);
+    return result;
   }
 
-  static VALUE _native_on_gvl_running(DDTRACE_UNUSED VALUE self, VALUE thread) {
+  static VALUE _native_on_gvl_running(DDTRACE_UNUSED VALUE self, VALUE thread, VALUE waiting_for_gvl_threshold_ns) {
     ENFORCE_THREAD(thread);
+    ENFORCE_TYPE(waiting_for_gvl_threshold_ns, T_FIXNUM);
 
-    return thread_context_collector_on_gvl_running(thread_from_thread_object(thread)) ? Qtrue : Qfalse;
-  }
+    debug_enter_unsafe_context();
 
-  static VALUE _native_sample_after_gvl_running(DDTRACE_UNUSED VALUE self, VALUE collector_instance, VALUE thread) {
-    ENFORCE_THREAD(thread);
+    per_thread_context *thread_context = get_per_thread_context(thread);
+    VALUE result;
+    if (thread_context) {
+      on_gvl_running_result gvl_running_result =
+        thread_context_collector_on_gvl_running(thread, thread_context, NUM2UINT(waiting_for_gvl_threshold_ns));
+      result = gvl_running_result.action == ON_GVL_RUNNING_SAMPLE ? Qtrue : Qfalse;
+    } else {
+      result = Qfalse;
+    }
+
+    debug_leave_unsafe_context();
 
-    return thread_context_collector_sample_after_gvl_running_with_thread(collector_instance, thread);
+    return result;
   }
 
-  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 VALUE _native_sample_after_gvl_running(DDTRACE_UNUSED VALUE self, VALUE collector_instance, VALUE thread) {
     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);
+    debug_enter_unsafe_context();
 
-    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");
+    VALUE result = thread_context_collector_sample_after_gvl_running(
+      collector_instance,
+      thread,
+      monotonic_wall_time_now_ns(RAISE_ON_FAILURE)
+    );
 
-    thread_context->cpu_time_at_previous_sample_ns += NUM2LONG(delta_ns);
+    debug_leave_unsafe_context();
 
-    return Qtrue;
+    return result;
   }
 
 #else
   static bool handle_gvl_waiting(
-    DDTRACE_UNUSED struct thread_context_collector_state *state,
+    DDTRACE_UNUSED 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 per_thread_context *thread_context,
     DDTRACE_UNUSED long current_cpu_time_ns
   ) { return false; }
+
 #endif // NO_GVL_INSTRUMENTATION
+
+static VALUE _native_apply_delta_to_cpu_time_at_previous_sample_ns(DDTRACE_UNUSED VALUE self, VALUE thread, VALUE delta_ns) {
+  ENFORCE_THREAD(thread);
+
+  per_thread_context *thread_context = get_per_thread_context(thread);
+  if (thread_context == NULL) raise_error(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;
+}
+
+#define MAX_SAFE_LOOKUP_SIZE 16
+
+typedef struct { VALUE lookup_key; VALUE result; } safe_lookup_hash_state;
+
+static int safe_lookup_hash_iterate(VALUE key, VALUE value, VALUE state_ptr) {
+  safe_lookup_hash_state *state = (safe_lookup_hash_state *) state_ptr;
+
+  if (key == state->lookup_key) {
+    state->result = value;
+    return ST_STOP;
+  }
+
+  return ST_CONTINUE;
+}
+
+// This method exists because we need to look up a hash during sampling, but we don't want to invoke any
+// Ruby code as a side effect. To be able to look up by hash, `rb_hash_lookup` calls `#hash` on the key,
+// which we want to avoid.
+// Thus, instead, we opt to just iterate through the hash and check if we can find what we're looking for.
+//
+// To avoid having too much overhead here we only iterate in hashes up to MAX_SAFE_LOOKUP_SIZE.
+// Note that we don't even try to iterate if the hash is bigger -- this is to avoid flaky behavior where
+// depending on the internal storage order of the hash we may or not find the key, and instead we always
+// enforce the size.
+static VALUE safely_lookup_hash_without_going_into_ruby_code(VALUE hash, VALUE key) {
+  if (!RB_TYPE_P(hash, T_HASH) || RHASH_SIZE(hash) > MAX_SAFE_LOOKUP_SIZE) return Qnil;
+
+  safe_lookup_hash_state state = {.lookup_key = key, .result = Qnil};
+
+  rb_hash_foreach(hash, safe_lookup_hash_iterate, (VALUE) &state);
+
+  return state.result;
+}
+
+static VALUE _native_system_epoch_time_now_ns(DDTRACE_UNUSED VALUE self, VALUE collector_instance) {
+  thread_context_collector_state *state;
+  TypedData_Get_Struct(collector_instance, thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  long current_monotonic_wall_time_ns = monotonic_wall_time_now_ns(RAISE_ON_FAILURE);
+  long system_epoch_time_ns = monotonic_to_system_epoch_ns(&state->time_converter_state, current_monotonic_wall_time_ns);
+
+  return LONG2NUM(system_epoch_time_ns);
+}
+
+static VALUE _native_prepare_sample_inside_signal_handler(DDTRACE_UNUSED VALUE self) {
+  return thread_context_collector_prepare_sample_inside_signal_handler() ? Qtrue : Qfalse;
+}
+
+static VALUE _native_request_prepare_on_gc_finish(DDTRACE_UNUSED VALUE self) {
+  thread_context_collector_request_prepare_on_gc_finish();
+  return Qnil;
+}