datadog 2.17.0 → 2.37.0

data/ext/datadog_profiling_native_extension/collectors_cpu_and_wall_time_worker.c

@@ -17,6 +17,10 @@
 #include "setup_signal_handler.h"
 #include "time_helpers.h"
 
+#ifdef __APPLE__
+  #include "macos_sampler_thread.h"
+#endif
+
 // Used to trigger the execution of Collectors::ThreadContext, which implements all of the sampling logic
 // itself; this class only implements the "when to do it" part.
 //
@@ -76,8 +80,6 @@
 //
 // ---
 
-#define ERR_CLOCK_FAIL "failed to get clock time"
-
 // Maximum allowed value for an allocation weight. Attempts to use higher values will result in clamping.
 // See https://docs.google.com/document/d/1lWLB714wlLBBq6T4xZyAc4a5wtWhSmr4-hgiPKeErlA/edit#heading=h.ugp0zxcj5iqh
 // (Datadog-only link) for research backing the choice of this value.
@@ -89,6 +91,7 @@ unsigned int MAX_ALLOC_WEIGHT = 10000;
   static rb_postponed_job_handle_t sample_from_postponed_job_handle;
   static rb_postponed_job_handle_t after_gc_from_postponed_job_handle;
   static rb_postponed_job_handle_t after_gvl_running_from_postponed_job_handle;
+  static rb_postponed_job_handle_t after_allocation_from_postponed_job_handle;
 #endif
 
 // Contains state for a single CpuAndWallTimeWorker instance
@@ -102,6 +105,8 @@ typedef struct {
   bool allocation_counting_enabled;
   bool gvl_profiling_enabled;
   bool skip_idle_samples_for_testing;
+  bool sighandler_sampling_enabled;
+  uint32_t cpu_sampling_interval_ms;
   VALUE self_instance;
   VALUE thread_context_collector_instance;
   VALUE idle_sampling_helper_instance;
@@ -116,13 +121,18 @@ typedef struct {
   // When something goes wrong during sampling, we record the Ruby exception here, so that it can be "re-raised" on
   // the CpuAndWallTimeWorker thread
   VALUE failure_exception;
+  const char *failure_exception_during_operation;
   // Used by `_native_stop` to flag the worker thread to start (see comment on `_native_sampling_loop`)
   VALUE stop_thread;
 
   // Others
 
   // Used to detect/avoid nested sampling, e.g. when on_newobj_event gets triggered by a memory allocation
-  // that happens during another sample.
+  // that happens during another sample, or when the signal handler gets triggered while we're already in the middle of
+  // sampling.
+  //
+  // @ivoanjo: Right now we always sample inside `safely_call`; if that ever changes, this flag may need to become
+  // volatile/atomic/have some barriers to ensure it's visible during e.g. signal handlers.
   bool during_sample;
 
   #ifndef NO_GVL_INSTRUMENTATION
@@ -134,27 +144,21 @@ typedef struct {
     // # Generic stats
     // How many times we tried to trigger a sample
     unsigned int trigger_sample_attempts;
+    // How many times extra sleep was triggered by dynamic sampling rate
+    unsigned int trigger_sample_extra_sleep;
     // How many times we tried to simulate signal delivery
     unsigned int trigger_simulated_signal_delivery_attempts;
     // How many times we actually simulated signal delivery
     unsigned int simulated_signal_delivery;
-    // How many times we actually called rb_postponed_job_register_one from a signal handler
+    // How many times we actually called rb_postponed_job_register_one from the signal handler
     unsigned int signal_handler_enqueued_sample;
+    // How many times we prepared a sample (sampled directly) from the signal handler
+    unsigned int signal_handler_prepared_sample;
     // How many times the signal handler was called from the wrong thread
     unsigned int signal_handler_wrong_thread;
     // How many times we actually tried to interrupt a thread for sampling
     unsigned int interrupt_thread_attempts;
 
-    // # Stats for the results of calling rb_postponed_job_register_one
-    // The same function was already waiting to be executed
-    unsigned int postponed_job_skipped_already_existed;
-    // The function was added to the queue successfully
-    unsigned int postponed_job_success;
-    // The queue was full
-    unsigned int postponed_job_full;
-    // The function returned an unknown result code
-    unsigned int postponed_job_unknown_result;
-
     // # CPU/Walltime sampling stats
     // How many times we actually CPU/wall sampled
     unsigned int cpu_sampled;
@@ -178,6 +182,11 @@ typedef struct {
     unsigned int allocations_during_sample;
 
     // # GVL profiling stats
+    // Note that this tracks two kind of samples from RESUMED:
+    // * samples when Waiting for GVL for more than the threshold
+    // * samples for the skip-samples-while-without-GVL optimization,
+    //   which are needed to attribute the time without GVL to the correct Ruby stack
+
     // How many times we triggered the after_gvl_running sampling
     unsigned int after_gvl_running;
     // How many times we skipped the after_gvl_running sampling
@@ -186,6 +195,11 @@ typedef struct {
     uint64_t gvl_sampling_time_ns_min;
     uint64_t gvl_sampling_time_ns_max;
     uint64_t gvl_sampling_time_ns_total;
+
+    struct vm_metrics {
+      // Total time spent waiting for the GVL
+      uint64_t gvl_waiting_time_ns_total;
+    } vm_metrics;
   } stats;
 } cpu_and_wall_time_worker_state;
 
@@ -194,17 +208,17 @@ static VALUE _native_initialize(int argc, VALUE *argv, DDTRACE_UNUSED VALUE _sel
 static void cpu_and_wall_time_worker_typed_data_mark(void *state_ptr);
 static VALUE _native_sampling_loop(VALUE self, VALUE instance);
 static VALUE _native_stop(DDTRACE_UNUSED VALUE _self, VALUE self_instance, VALUE worker_thread);
-static VALUE stop(VALUE self_instance, VALUE optional_exception);
-static void stop_state(cpu_and_wall_time_worker_state *state, VALUE optional_exception);
+static VALUE stop(VALUE self_instance, VALUE optional_exception, const char *optional_exception_during_operation);
+static void stop_state(cpu_and_wall_time_worker_state *state, VALUE optional_exception, const char *optional_operation_name);
 static void handle_sampling_signal(DDTRACE_UNUSED int _signal, DDTRACE_UNUSED siginfo_t *_info, DDTRACE_UNUSED void *_ucontext);
 static void *run_sampling_trigger_loop(void *state_ptr);
 static void interrupt_sampling_trigger_loop(void *state_ptr);
 static void sample_from_postponed_job(DDTRACE_UNUSED void *_unused);
 static VALUE rescued_sample_from_postponed_job(VALUE self_instance);
-static VALUE handle_sampling_failure(VALUE self_instance, VALUE exception);
 static VALUE _native_current_sigprof_signal_handler(DDTRACE_UNUSED VALUE self);
 static VALUE release_gvl_and_run_sampling_trigger_loop(VALUE instance);
 static VALUE _native_is_running(DDTRACE_UNUSED VALUE self, VALUE instance);
+static VALUE _native_failure_exception_during_operation(DDTRACE_UNUSED VALUE self, VALUE instance);
 static void testing_signal_handler(DDTRACE_UNUSED int _signal, DDTRACE_UNUSED siginfo_t *_info, DDTRACE_UNUSED void *_ucontext);
 static VALUE _native_install_testing_signal_handler(DDTRACE_UNUSED VALUE self);
 static VALUE _native_remove_testing_signal_handler(DDTRACE_UNUSED VALUE self);
@@ -212,7 +226,12 @@ static VALUE _native_trigger_sample(DDTRACE_UNUSED VALUE self);
 static VALUE _native_gc_tracepoint(DDTRACE_UNUSED VALUE self, VALUE instance);
 static void on_gc_event(VALUE tracepoint_data, DDTRACE_UNUSED void *unused);
 static void after_gc_from_postponed_job(DDTRACE_UNUSED void *_unused);
-static VALUE safely_call(VALUE (*function_to_call_safely)(VALUE), VALUE function_to_call_safely_arg, VALUE instance);
+static VALUE safely_call(
+  VALUE (*function_to_call_safely)(VALUE),
+  VALUE function_to_call_safely_arg,
+  VALUE instance,
+  VALUE (*handle_sampling_failure)(VALUE, VALUE)
+);
 static VALUE _native_simulate_handle_sampling_signal(DDTRACE_UNUSED VALUE self);
 static VALUE _native_simulate_sample_from_postponed_job(DDTRACE_UNUSED VALUE self);
 static VALUE _native_reset_after_fork(DDTRACE_UNUSED VALUE self, VALUE instance);
@@ -228,16 +247,26 @@ static void on_newobj_event(DDTRACE_UNUSED VALUE unused1, DDTRACE_UNUSED void *u
 static void disable_tracepoints(cpu_and_wall_time_worker_state *state);
 static VALUE _native_with_blocked_sigprof(DDTRACE_UNUSED VALUE self);
 static VALUE rescued_sample_allocation(VALUE tracepoint_data);
+static VALUE rescued_after_allocation(VALUE self_instance);
 static void delayed_error(cpu_and_wall_time_worker_state *state, const char *error);
+static void delayed_error_clock_failure(cpu_and_wall_time_worker_state *state);
 static VALUE _native_delayed_error(DDTRACE_UNUSED VALUE self, VALUE instance, VALUE error_msg);
 static VALUE _native_hold_signals(DDTRACE_UNUSED VALUE self);
 static VALUE _native_resume_signals(DDTRACE_UNUSED VALUE self);
 #ifndef NO_GVL_INSTRUMENTATION
-static void on_gvl_event(rb_event_flag_t event_id, const rb_internal_thread_event_data_t *event_data, DDTRACE_UNUSED void *_unused);
-static void after_gvl_running_from_postponed_job(DDTRACE_UNUSED void *_unused);
+  static void on_gvl_event(rb_event_flag_t event_id, const rb_internal_thread_event_data_t *event_data, DDTRACE_UNUSED void *_unused);
+  static void after_gvl_running_from_postponed_job(DDTRACE_UNUSED void *_unused);
+  static VALUE rescued_after_gvl_running_from_postponed_job(VALUE self_instance);
+  static VALUE handle_sampling_failure_rescued_after_gvl_running_from_postponed_job(VALUE self_instance, VALUE exception);
 #endif
-static VALUE rescued_after_gvl_running_from_postponed_job(VALUE self_instance);
 static VALUE _native_gvl_profiling_hook_active(DDTRACE_UNUSED VALUE self, VALUE instance);
+static VALUE handle_sampling_failure_rescued_sample_from_postponed_job(VALUE self_instance, VALUE exception);
+static VALUE handle_sampling_failure_thread_context_collector_sample_after_gc(VALUE self_instance, VALUE exception);
+static VALUE handle_sampling_failure_rescued_sample_allocation(VALUE self_instance, VALUE exception);
+static VALUE handle_sampling_failure_rescued_after_allocation(VALUE self_instance, VALUE exception);
+static inline void during_sample_enter(cpu_and_wall_time_worker_state* state);
+static inline void during_sample_exit(cpu_and_wall_time_worker_state* state);
+static void after_allocation_from_postponed_job(DDTRACE_UNUSED void *_unused);
 
 // We're using `on_newobj_event` function with `rb_add_event_hook2`, which requires in its public signature a function
 // with signature `rb_event_hook_func_t` which doesn't match `on_newobj_event`.
@@ -263,6 +292,7 @@ static VALUE _native_gvl_profiling_hook_active(DDTRACE_UNUSED VALUE self, VALUE
 // (e.g. signal handler) where it's impossible or just awkward to pass it as an argument.
 static VALUE active_sampler_instance = Qnil;
 static cpu_and_wall_time_worker_state *active_sampler_instance_state = NULL;
+static VALUE clock_failure_exception_class = Qnil;
 
 // See handle_sampling_signal for details on what this does
 #ifdef NO_POSTPONED_TRIGGER
@@ -284,13 +314,15 @@ void collectors_cpu_and_wall_time_worker_init(VALUE profiling_module) {
     sample_from_postponed_job_handle = rb_postponed_job_preregister(unused_flags, sample_from_postponed_job, NULL);
     after_gc_from_postponed_job_handle = rb_postponed_job_preregister(unused_flags, after_gc_from_postponed_job, NULL);
     after_gvl_running_from_postponed_job_handle = rb_postponed_job_preregister(unused_flags, after_gvl_running_from_postponed_job, NULL);
+    after_allocation_from_postponed_job_handle = rb_postponed_job_preregister(unused_flags, after_allocation_from_postponed_job, NULL);
 
     if (
       sample_from_postponed_job_handle == POSTPONED_JOB_HANDLE_INVALID ||
       after_gc_from_postponed_job_handle == POSTPONED_JOB_HANDLE_INVALID ||
-      after_gvl_running_from_postponed_job_handle == POSTPONED_JOB_HANDLE_INVALID
+      after_gvl_running_from_postponed_job_handle == POSTPONED_JOB_HANDLE_INVALID ||
+      after_allocation_from_postponed_job_handle == POSTPONED_JOB_HANDLE_INVALID
     ) {
-      rb_raise(rb_eRuntimeError, "Failed to register profiler postponed jobs (got POSTPONED_JOB_HANDLE_INVALID)");
+      raise_error(rb_eRuntimeError, "Failed to register profiler postponed jobs (got POSTPONED_JOB_HANDLE_INVALID)");
     }
   #else
     gc_finalize_deferred_workaround = objspace_ptr_for_gc_finalize_deferred_workaround();
@@ -300,6 +332,8 @@ void collectors_cpu_and_wall_time_worker_init(VALUE profiling_module) {
   VALUE collectors_cpu_and_wall_time_worker_class = rb_define_class_under(collectors_module, "CpuAndWallTimeWorker", rb_cObject);
   // Hosts methods used for testing the native code using RSpec
   VALUE testing_module = rb_define_module_under(collectors_cpu_and_wall_time_worker_class, "Testing");
+  clock_failure_exception_class = rb_define_class_under(collectors_cpu_and_wall_time_worker_class, "ClockFailure", rb_eRuntimeError);
+  rb_gc_register_mark_object(clock_failure_exception_class);
 
   // Instances of the CpuAndWallTimeWorker class are "TypedData" objects.
   // "TypedData" objects are special objects in the Ruby VM that can wrap C structs.
@@ -319,6 +353,7 @@ void collectors_cpu_and_wall_time_worker_init(VALUE profiling_module) {
   rb_define_singleton_method(collectors_cpu_and_wall_time_worker_class, "_native_stats_reset_not_thread_safe", _native_stats_reset_not_thread_safe, 1);
   rb_define_singleton_method(collectors_cpu_and_wall_time_worker_class, "_native_allocation_count", _native_allocation_count, 0);
   rb_define_singleton_method(collectors_cpu_and_wall_time_worker_class, "_native_is_running?", _native_is_running, 1);
+  rb_define_singleton_method(collectors_cpu_and_wall_time_worker_class, "_native_failure_exception_during_operation", _native_failure_exception_during_operation, 1);
   rb_define_singleton_method(testing_module, "_native_current_sigprof_signal_handler", _native_current_sigprof_signal_handler, 0);
   rb_define_singleton_method(collectors_cpu_and_wall_time_worker_class, "_native_hold_signals", _native_hold_signals, 0);
   rb_define_singleton_method(collectors_cpu_and_wall_time_worker_class, "_native_resume_signals", _native_resume_signals, 0);
@@ -362,6 +397,8 @@ static VALUE _native_new(VALUE klass) {
   state->allocation_counting_enabled = false;
   state->gvl_profiling_enabled = false;
   state->skip_idle_samples_for_testing = false;
+  state->sighandler_sampling_enabled = false;
+  state->cpu_sampling_interval_ms = 10;
   state->thread_context_collector_instance = Qnil;
   state->idle_sampling_helper_instance = Qnil;
   state->owner_thread = Qnil;
@@ -370,9 +407,10 @@ static VALUE _native_new(VALUE klass) {
 
   atomic_init(&state->should_run, false);
   state->failure_exception = Qnil;
+  state->failure_exception_during_operation = NULL;
   state->stop_thread = Qnil;
 
-  state->during_sample = false;
+  during_sample_exit(state);
 
   #ifndef NO_GVL_INSTRUMENTATION
     state->gvl_profiling_hook = NULL;
@@ -404,6 +442,8 @@ static VALUE _native_initialize(int argc, VALUE *argv, DDTRACE_UNUSED VALUE _sel
   VALUE allocation_counting_enabled = rb_hash_fetch(options, ID2SYM(rb_intern("allocation_counting_enabled")));
   VALUE gvl_profiling_enabled = rb_hash_fetch(options, ID2SYM(rb_intern("gvl_profiling_enabled")));
   VALUE skip_idle_samples_for_testing = rb_hash_fetch(options, ID2SYM(rb_intern("skip_idle_samples_for_testing")));
+  VALUE sighandler_sampling_enabled = rb_hash_fetch(options, ID2SYM(rb_intern("sighandler_sampling_enabled")));
+  VALUE cpu_sampling_interval_ms = rb_hash_fetch(options, ID2SYM(rb_intern("cpu_sampling_interval_ms")));
 
   ENFORCE_BOOLEAN(gc_profiling_enabled);
   ENFORCE_BOOLEAN(no_signals_workaround_enabled);
@@ -413,6 +453,8 @@ static VALUE _native_initialize(int argc, VALUE *argv, DDTRACE_UNUSED VALUE _sel
   ENFORCE_BOOLEAN(allocation_counting_enabled);
   ENFORCE_BOOLEAN(gvl_profiling_enabled);
   ENFORCE_BOOLEAN(skip_idle_samples_for_testing)
+  ENFORCE_BOOLEAN(sighandler_sampling_enabled)
+  ENFORCE_TYPE(cpu_sampling_interval_ms, T_FIXNUM);
 
   cpu_and_wall_time_worker_state *state;
   TypedData_Get_Struct(self_instance, cpu_and_wall_time_worker_state, &cpu_and_wall_time_worker_typed_data, state);
@@ -424,6 +466,8 @@ static VALUE _native_initialize(int argc, VALUE *argv, DDTRACE_UNUSED VALUE _sel
   state->allocation_counting_enabled = (allocation_counting_enabled == Qtrue);
   state->gvl_profiling_enabled = (gvl_profiling_enabled == Qtrue);
   state->skip_idle_samples_for_testing = (skip_idle_samples_for_testing == Qtrue);
+  state->sighandler_sampling_enabled = (sighandler_sampling_enabled == Qtrue);
+  state->cpu_sampling_interval_ms = NUM2INT(cpu_sampling_interval_ms);
 
   double total_overhead_target_percentage = NUM2DBL(dynamic_sampling_rate_overhead_target_percentage);
   if (!state->allocation_profiling_enabled) {
@@ -469,10 +513,7 @@ static VALUE _native_sampling_loop(DDTRACE_UNUSED VALUE _self, VALUE instance) {
   cpu_and_wall_time_worker_state *old_state = active_sampler_instance_state;
   if (old_state != NULL) {
     if (is_thread_alive(old_state->owner_thread)) {
-      rb_raise(
-        rb_eRuntimeError,
-        "Could not start CpuAndWallTimeWorker: There's already another instance of CpuAndWallTimeWorker active in a different thread"
-      );
+      raise_error(rb_eRuntimeError, "Could not start CpuAndWallTimeWorker: There's already another instance of CpuAndWallTimeWorker active in a different thread");
     } else {
       // The previously active thread seems to have died without cleaning up after itself.
       // In this case, we can still go ahead and start the profiler BUT we make sure to disable any existing tracepoint
@@ -489,6 +530,8 @@ static VALUE _native_sampling_loop(DDTRACE_UNUSED VALUE _self, VALUE instance) {
   // situation we stop immediately and never even start the sampling trigger loop.
   if (state->stop_thread == rb_thread_current()) return Qnil;
 
+  thread_context_collector_profiler_internal_thread_started();
+
   // Reset the dynamic sampling rate state, if any (reminder: the monotonic clock reference may change after a fork)
   dynamic_sampling_rate_reset(&state->cpu_dynamic_sampling_rate);
   long now = monotonic_wall_time_now_ns(RAISE_ON_FAILURE);
@@ -503,6 +546,14 @@ static VALUE _native_sampling_loop(DDTRACE_UNUSED VALUE _self, VALUE instance) {
 
   block_sigprof_signal_handler_from_running_in_current_thread(); // We want to interrupt the thread with the global VM lock, never this one
 
+  #ifdef __APPLE__
+    // Promote the native thread that will run the sampling loop to a realtime scheduling policy so
+    // its periodic wake-ups are not subject to the default ~1-2ms timeshare wake latency. The matching
+    // demote happens below in this same function, alongside the SIGPROF unblock dance, so both per-thread
+    // policy changes are reverted before Ruby reuses this native thread for an unrelated Ruby thread.
+    promote_sampler_thread_to_realtime(MILLIS_AS_NS(state->cpu_sampling_interval_ms));
+  #endif
+
   // Release GVL, get to the actual work!
   int exception_state;
   rb_protect(release_gvl_and_run_sampling_trigger_loop, instance, &exception_state);
@@ -524,6 +575,14 @@ static VALUE _native_sampling_loop(DDTRACE_UNUSED VALUE _self, VALUE instance) {
   // had SIGPROF delivery blocked. :hide_the_pain_harold:
   unblock_sigprof_signal_handler_from_running_in_current_thread();
 
+  #ifdef __APPLE__
+    // Pairs with promote_sampler_thread_to_realtime above. Same reasoning as the SIGPROF unblock:
+    // Ruby may reuse this native thread for an unrelated Ruby thread that would otherwise inherit
+    // our scheduler policy. Doing it here also covers the exception path, since grab_gvl_and_sample
+    // can raise and unwind out of the sampling loop.
+    demote_sampler_thread_from_realtime();
+  #endif
+
   // Why replace and not use remove the signal handler? We do this because when a process receives a SIGPROF without
   // having an explicit signal handler set up, the process will instantly terminate with a confusing
   // "Profiling timer expired" message left behind. (This message doesn't come from us -- it's the default message for
@@ -534,8 +593,9 @@ static VALUE _native_sampling_loop(DDTRACE_UNUSED VALUE _self, VALUE instance) {
   // @ivoanjo: I suspect this will never happen, but the cost of getting it wrong is really high (VM terminates) so this
   // is a just-in-case situation.
   //
-  // Note 2: This can raise exceptions as well, so make sure that all cleanups are done by the time we get here.
-  replace_sigprof_signal_handler_with_empty_handler(handle_sampling_signal);
+  // Note 2: If exception_state is set, we have a pending exception that we'll re-raise below.
+  // In that case, we don't want replace_sigprof_signal_handler_with_empty_handler to raise another exception.
+  replace_sigprof_signal_handler_with_empty_handler(handle_sampling_signal, !exception_state);
 
   // Ensure that instance is not garbage collected while the native sampling loop is running; this is probably not needed, but just in case
   RB_GC_GUARD(instance);
@@ -551,22 +611,24 @@ static VALUE _native_stop(DDTRACE_UNUSED VALUE _self, VALUE self_instance, VALUE
 
   state->stop_thread = worker_thread;
 
-  return stop(self_instance, /* optional_exception: */ Qnil);
+  return stop(self_instance, Qnil, NULL);
 }
 
-static void stop_state(cpu_and_wall_time_worker_state *state, VALUE optional_exception) {
+// When providing an `optional_exception`, `optional_exception_during_operation` should be provided as well
+static void stop_state(cpu_and_wall_time_worker_state *state, VALUE optional_exception, const char *optional_exception_during_operation) {
   atomic_store(&state->should_run, false);
   state->failure_exception = optional_exception;
+  state->failure_exception_during_operation = optional_exception_during_operation;
 
   // Disable the tracepoints as soon as possible, so the VM doesn't keep on calling them
   disable_tracepoints(state);
 }
 
-static VALUE stop(VALUE self_instance, VALUE optional_exception) {
+static VALUE stop(VALUE self_instance, VALUE optional_exception, const char *optional_exception_during_operation) {
   cpu_and_wall_time_worker_state *state;
   TypedData_Get_Struct(self_instance, cpu_and_wall_time_worker_state, &cpu_and_wall_time_worker_typed_data, state);
 
-  stop_state(state, optional_exception);
+  stop_state(state, optional_exception, optional_exception_during_operation);
 
   return Qtrue;
 }
@@ -589,25 +651,30 @@ static void handle_sampling_signal(DDTRACE_UNUSED int _signal, DDTRACE_UNUSED si
     return;
   }
 
-  // We implicitly assume there can be no concurrent nor nested calls to handle_sampling_signal because
-  // a) we get triggered using SIGPROF, and the docs state a second SIGPROF will not interrupt an existing one
+  // We assume there can be no concurrent nor nested calls to handle_sampling_signal because
+  // a) we get triggered using SIGPROF, and the docs state a second SIGPROF will not interrupt an existing one (see sigaction docs on sa_mask)
   // b) we validate we are in the thread that has the global VM lock; if a different thread gets a signal, it will return early
   //    because it will not have the global VM lock
 
-  // Note: rb_postponed_job_register_one ensures that if there's a previous sample_from_postponed_job queued for execution
-  // then we will not queue a second one. It does this by doing a linear scan on the existing jobs; in the future we
-  // may want to implement that check ourselves.
-
   state->stats.signal_handler_enqueued_sample++;
 
-  // Note: If we ever want to get rid of rb_postponed_job_register_one, remember not to clobber Ruby exceptions, as
-  // this function does this helpful job for us now -- https://github.com/ruby/ruby/commit/a98e343d39c4d7bf1e2190b076720f32d9f298b3.
+  bool sample_from_signal_handler =
+    state->sighandler_sampling_enabled &&
+    // Don't sample if we're already in the middle of processing a sample
+    !state->during_sample;
+
+  if (sample_from_signal_handler) {
+    // Buffer current stack trace. Note that this will not actually record the sample, for that we still need to wait
+    // until the postponed job below gets run.
+    bool prepared = thread_context_collector_prepare_sample_inside_signal_handler();
+
+    if (prepared) state->stats.signal_handler_prepared_sample++;
+  }
+
   #ifndef NO_POSTPONED_TRIGGER // Ruby 3.3+
     rb_postponed_job_trigger(sample_from_postponed_job_handle);
-    state->stats.postponed_job_success++; // Always succeeds
   #else
-
-    // This is a workaround for https://bugs.ruby-lang.org/issues/19991 (for Ruby < 3.3)
+    // Passing in `gc_finalize_deferred_workaround` is a workaround for https://bugs.ruby-lang.org/issues/19991 (for Ruby < 3.3)
     //
     // TL;DR the `rb_postponed_job_register_one` API is not atomic (which is why it got replaced by `rb_postponed_job_trigger`)
     // and in rare cases can cause VM crashes.
@@ -631,20 +698,7 @@ static void handle_sampling_signal(DDTRACE_UNUSED int _signal, DDTRACE_UNUSED si
     //
     // Thus, our workaround is simple: we pass in objspace as our argument, just in case the clobbering happens.
     // In the happy path, we never use this argument so it makes no difference. In the buggy path, we avoid crashing the VM.
-    int result = rb_postponed_job_register(0, sample_from_postponed_job, gc_finalize_deferred_workaround /* instead of NULL */);
-
-    // Officially, the result of rb_postponed_job_register_one is documented as being opaque, but in practice it does not
-    // seem to have changed between Ruby 2.3 and 3.2, and so we track it as a debugging mechanism
-    switch (result) {
-      case 0:
-        state->stats.postponed_job_full++; break;
-      case 1:
-        state->stats.postponed_job_success++; break;
-      case 2:
-        state->stats.postponed_job_skipped_already_existed++; break;
-      default:
-        state->stats.postponed_job_unknown_result++;
-    }
+    rb_postponed_job_register(0, sample_from_postponed_job, gc_finalize_deferred_workaround /* instead of NULL */);
   #endif
 }
 
@@ -652,7 +706,7 @@ static void handle_sampling_signal(DDTRACE_UNUSED int _signal, DDTRACE_UNUSED si
 static void *run_sampling_trigger_loop(void *state_ptr) {
   cpu_and_wall_time_worker_state *state = (cpu_and_wall_time_worker_state *) state_ptr;
 
-  uint64_t minimum_time_between_signals = MILLIS_AS_NS(10);
+  uint64_t minimum_time_between_signals = MILLIS_AS_NS(state->cpu_sampling_interval_ms);
 
   while (atomic_load(&state->should_run)) {
     state->stats.trigger_sample_attempts++;
@@ -674,6 +728,10 @@ static void *run_sampling_trigger_loop(void *state_ptr) {
         // we're doing this, so we may still not signal the correct thread from time to time, but our signal handler
         // includes a check to see if it got called in the right thread
         state->stats.interrupt_thread_attempts++;
+
+        // Pick up any last-minute attempts to stop before we send the signal
+        if (!atomic_load(&state->should_run)) return NULL;
+
         pthread_kill(owner.owner, SIGPROF);
       } else {
         if (state->skip_idle_samples_for_testing) {
@@ -701,7 +759,10 @@ static void *run_sampling_trigger_loop(void *state_ptr) {
     // `dynamic_sampling_rate_get_sleep` may have changed while the above sleep was ongoing.
     uint64_t extra_sleep =
       dynamic_sampling_rate_get_sleep(&state->cpu_dynamic_sampling_rate, monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE));
-    if (state->dynamic_sampling_rate_enabled && extra_sleep > 0) sleep_for(extra_sleep);
+    if (state->dynamic_sampling_rate_enabled && extra_sleep > 0) {
+      state->stats.trigger_sample_extra_sleep++;
+      sleep_for(extra_sleep);
+    }
   }
 
   return NULL; // Unused
@@ -714,6 +775,8 @@ static void interrupt_sampling_trigger_loop(void *state_ptr) {
   atomic_store(&state->should_run, false);
 }
 
+  // Note: If we ever want to get rid of the postponed job execution, remember not to clobber Ruby exceptions, as
+  // this function does this helpful job for us now -- https://github.com/ruby/ruby/commit/a98e343d39c4d7bf1e2190b076720f32d9f298b3.
 static void sample_from_postponed_job(DDTRACE_UNUSED void *_unused) {
   cpu_and_wall_time_worker_state *state = active_sampler_instance_state; // Read from global variable, see "sampler global state safety" note above
 
@@ -726,12 +789,17 @@ static void sample_from_postponed_job(DDTRACE_UNUSED void *_unused) {
     return; // We're not on the main Ractor; we currently don't support profiling non-main Ractors
   }
 
-  state->during_sample = true;
+  during_sample_enter(state);
 
   // Rescue against any exceptions that happen during sampling
-  safely_call(rescued_sample_from_postponed_job, state->self_instance, state->self_instance);
+  safely_call(
+    rescued_sample_from_postponed_job,
+    state->self_instance,
+    state->self_instance,
+    handle_sampling_failure_rescued_sample_from_postponed_job
+  );
 
-  state->during_sample = false;
+  during_sample_exit(state);
 }
 
 static VALUE rescued_sample_from_postponed_job(VALUE self_instance) {
@@ -747,8 +815,7 @@ static VALUE rescued_sample_from_postponed_job(VALUE self_instance) {
 
   state->stats.cpu_sampled++;
 
-  VALUE profiler_overhead_stack_thread = state->owner_thread; // Used to attribute profiler overhead to a different stack
-  thread_context_collector_sample(state->thread_context_collector_instance, wall_time_ns_before_sample, profiler_overhead_stack_thread);
+  thread_context_collector_sample(state->thread_context_collector_instance, wall_time_ns_before_sample);
 
   long wall_time_ns_after_sample = monotonic_wall_time_now_ns(RAISE_ON_FAILURE);
   long delta_ns = wall_time_ns_after_sample - wall_time_ns_before_sample;
@@ -766,11 +833,6 @@ static VALUE rescued_sample_from_postponed_job(VALUE self_instance) {
   return Qnil;
 }
 
-static VALUE handle_sampling_failure(VALUE self_instance, VALUE exception) {
-  stop(self_instance, exception);
-  return Qnil;
-}
-
 // This method exists only to enable testing Datadog::Profiling::Collectors::CpuAndWallTimeWorker behavior using RSpec.
 // It SHOULD NOT be used for other purposes.
 static VALUE _native_current_sigprof_signal_handler(DDTRACE_UNUSED VALUE self) {
@@ -809,22 +871,19 @@ static VALUE release_gvl_and_run_sampling_trigger_loop(VALUE instance) {
 
   if (state->gvl_profiling_enabled) {
     #ifndef NO_GVL_INSTRUMENTATION
-      #ifdef USE_GVL_PROFILING_3_2_WORKAROUNDS
-        gvl_profiling_state_thread_tracking_workaround();
-      #endif
-
       state->gvl_profiling_hook = rb_internal_thread_add_event_hook(
         on_gvl_event,
         (
           // For now we're only asking for these events, even though there's more
           // (e.g. check docs or gvl-tracing gem)
+          RUBY_INTERNAL_THREAD_EVENT_SUSPENDED | /* released gvl */
           RUBY_INTERNAL_THREAD_EVENT_READY /* waiting for gvl */ |
           RUBY_INTERNAL_THREAD_EVENT_RESUMED /* running/runnable */
         ),
         NULL
       );
     #else
-      rb_raise(rb_eArgError, "GVL profiling is not supported in this Ruby version");
+      raise_error(rb_eArgError, "GVL profiling is not supported in this Ruby version");
     #endif
   }
 
@@ -836,6 +895,8 @@ static VALUE release_gvl_and_run_sampling_trigger_loop(VALUE instance) {
   // If we stopped sampling due to an exception, re-raise it (now in the worker thread)
   if (state->failure_exception != Qnil) rb_exc_raise(state->failure_exception);
 
+  thread_context_collector_profiler_internal_thread_done(state->thread_context_collector_instance);
+
   return Qnil;
 }
 
@@ -847,6 +908,15 @@ static VALUE _native_is_running(DDTRACE_UNUSED VALUE self, VALUE instance) {
   return (state != NULL && is_thread_alive(state->owner_thread) && state->self_instance == instance) ? Qtrue : Qfalse;
 }
 
+static VALUE _native_failure_exception_during_operation(DDTRACE_UNUSED VALUE self, VALUE instance) {
+  cpu_and_wall_time_worker_state *state;
+  TypedData_Get_Struct(instance, cpu_and_wall_time_worker_state, &cpu_and_wall_time_worker_typed_data, state);
+
+  if (state->failure_exception_during_operation == NULL) return Qnil;
+
+  return rb_str_new_cstr(state->failure_exception_during_operation);
+}
+
 static void testing_signal_handler(DDTRACE_UNUSED int _signal, DDTRACE_UNUSED siginfo_t *_info, DDTRACE_UNUSED void *_ucontext) {
   /* Does nothing on purpose */
 }
@@ -937,16 +1007,26 @@ static void after_gc_from_postponed_job(DDTRACE_UNUSED void *_unused) {
     return; // We're not on the main Ractor; we currently don't support profiling non-main Ractors
   }
 
-  state->during_sample = true;
+  during_sample_enter(state);
 
-  safely_call(thread_context_collector_sample_after_gc, state->thread_context_collector_instance, state->self_instance);
+  safely_call(
+    thread_context_collector_sample_after_gc,
+    state->thread_context_collector_instance,
+    state->self_instance,
+    handle_sampling_failure_thread_context_collector_sample_after_gc
+  );
 
-  state->during_sample = false;
+  during_sample_exit(state);
 }
 
 // Equivalent to Ruby begin/rescue call, where we call a C function and jump to the exception handler if an
 // exception gets raised within
-static VALUE safely_call(VALUE (*function_to_call_safely)(VALUE), VALUE function_to_call_safely_arg, VALUE instance) {
+static VALUE safely_call(
+  VALUE (*function_to_call_safely)(VALUE),
+  VALUE function_to_call_safely_arg,
+  VALUE instance,
+  VALUE (*handle_sampling_failure)(VALUE, VALUE)
+) {
   VALUE exception_handler_function_arg = instance;
   return rb_rescue2(
     function_to_call_safely,
@@ -980,6 +1060,9 @@ static VALUE _native_simulate_sample_from_postponed_job(DDTRACE_UNUSED VALUE sel
 //
 // In the future, if we add more other components with tracepoints, we will need to coordinate stopping all such
 // tracepoints before doing the other cleaning steps.
+//
+// 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 instance) {
   cpu_and_wall_time_worker_state *state;
   TypedData_Get_Struct(instance, cpu_and_wall_time_worker_state, &cpu_and_wall_time_worker_typed_data, state);
@@ -1016,14 +1099,12 @@ static VALUE _native_stats(DDTRACE_UNUSED VALUE self, VALUE instance) {
   VALUE stats_as_hash = rb_hash_new();
   VALUE arguments[] = {
     ID2SYM(rb_intern("trigger_sample_attempts")),                    /* => */ UINT2NUM(state->stats.trigger_sample_attempts),
+    ID2SYM(rb_intern("trigger_sample_extra_sleep")),                 /* => */ UINT2NUM(state->stats.trigger_sample_extra_sleep),
     ID2SYM(rb_intern("trigger_simulated_signal_delivery_attempts")), /* => */ UINT2NUM(state->stats.trigger_simulated_signal_delivery_attempts),
     ID2SYM(rb_intern("simulated_signal_delivery")),                  /* => */ UINT2NUM(state->stats.simulated_signal_delivery),
     ID2SYM(rb_intern("signal_handler_enqueued_sample")),             /* => */ UINT2NUM(state->stats.signal_handler_enqueued_sample),
+    ID2SYM(rb_intern("signal_handler_prepared_sample")),             /* => */ UINT2NUM(state->stats.signal_handler_prepared_sample),
     ID2SYM(rb_intern("signal_handler_wrong_thread")),                /* => */ UINT2NUM(state->stats.signal_handler_wrong_thread),
-    ID2SYM(rb_intern("postponed_job_skipped_already_existed")),      /* => */ UINT2NUM(state->stats.postponed_job_skipped_already_existed),
-    ID2SYM(rb_intern("postponed_job_success")),                      /* => */ UINT2NUM(state->stats.postponed_job_success),
-    ID2SYM(rb_intern("postponed_job_full")),                         /* => */ UINT2NUM(state->stats.postponed_job_full),
-    ID2SYM(rb_intern("postponed_job_unknown_result")),               /* => */ UINT2NUM(state->stats.postponed_job_unknown_result),
     ID2SYM(rb_intern("interrupt_thread_attempts")),                  /* => */ UINT2NUM(state->stats.interrupt_thread_attempts),
 
     // CPU Stats
@@ -1053,8 +1134,12 @@ static VALUE _native_stats(DDTRACE_UNUSED VALUE self, VALUE instance) {
     ID2SYM(rb_intern("gvl_sampling_time_ns_max")),   /* => */ RUBY_NUM_OR_NIL(state->stats.gvl_sampling_time_ns_max, > 0, ULL2NUM),
     ID2SYM(rb_intern("gvl_sampling_time_ns_total")), /* => */ RUBY_NUM_OR_NIL(state->stats.gvl_sampling_time_ns_total, > 0, ULL2NUM),
     ID2SYM(rb_intern("gvl_sampling_time_ns_avg")),   /* => */ RUBY_AVG_OR_NIL(state->stats.gvl_sampling_time_ns_total, state->stats.after_gvl_running),
+    ID2SYM(rb_intern("gvl_waiting_time_ns_total")),  /* => */ state->gvl_profiling_enabled ? ULL2NUM(state->stats.vm_metrics.gvl_waiting_time_ns_total) : Qnil,
   };
   for (long unsigned int i = 0; i < VALUE_COUNT(arguments); i += 2) rb_hash_aset(stats_as_hash, arguments[i], arguments[i+1]);
+
+  thread_context_collector_stats(state->thread_context_collector_instance, stats_as_hash);
+
   return stats_as_hash;
 }
 
@@ -1062,6 +1147,7 @@ static VALUE _native_stats_reset_not_thread_safe(DDTRACE_UNUSED VALUE self, VALU
   cpu_and_wall_time_worker_state *state;
   TypedData_Get_Struct(instance, cpu_and_wall_time_worker_state, &cpu_and_wall_time_worker_typed_data, state);
   reset_stats_not_thread_safe(state);
+  thread_context_collector_stats_reset_not_thread_safe(state->thread_context_collector_instance);
   return Qnil;
 }
 
@@ -1073,8 +1159,7 @@ void *simulate_sampling_signal_delivery(DDTRACE_UNUSED void *_unused) {
 
   state->stats.simulated_signal_delivery++;
 
-  // @ivoanjo: We could instead directly call sample_from_postponed_job, but I chose to go through the signal handler
-  // so that the simulated case is as close to the original one as well (including any metrics increases, etc).
+  // `handle_sampling_signal` does a few things extra on top of `sample_from_postponed_job` so that's why we don't shortcut here
   handle_sampling_signal(0, NULL, NULL);
 
   return NULL; // Unused
@@ -1094,6 +1179,7 @@ static void reset_stats_not_thread_safe(cpu_and_wall_time_worker_state *state) {
     .cpu_sampling_time_ns_min        = UINT64_MAX,
     .allocation_sampling_time_ns_min = UINT64_MAX,
     .gvl_sampling_time_ns_min        = UINT64_MAX,
+    // Other fields are reset to 0
   };
 }
 
@@ -1126,7 +1212,7 @@ static VALUE _native_allocation_count(DDTRACE_UNUSED VALUE self) {
 #define HANDLE_CLOCK_FAILURE(call) ({ \
     long _result = (call); \
     if (_result == 0) { \
-        delayed_error(state, ERR_CLOCK_FAIL); \
+        delayed_error_clock_failure(state); \
         return; \
     } \
     _result; \
@@ -1171,13 +1257,22 @@ static void on_newobj_event(DDTRACE_UNUSED VALUE unused1, DDTRACE_UNUSED void *u
     return;
   }
 
+  VALUE current_thread = rb_thread_current();
+
   // If Ruby is in the middle of raising an exception, we don't want to try to sample. This is because if we accidentally
   // trigger an exception inside the profiler code, bad things will happen (specifically, Ruby will try to kill off the
   // thread even though we may try to catch the exception).
   //
   // Note that "in the middle of raising an exception" means the exception itself has already been allocated.
   // What's getting allocated now is probably the backtrace objects (@ivoanjo or at least that's what I've observed)
-  if (is_raised_flag_set(rb_thread_current())) {
+  if (is_raised_flag_set(current_thread)) {
+    return;
+  }
+
+  per_thread_context *thread_context = get_per_thread_context(current_thread);
+  if (!thread_context) {
+    // Context is created eagerly via on_thread_begin_event, so this should not normally be NULL.
+    // We keep the guard since we can't allocate here (inside on_newobj_event).
     return;
   }
 
@@ -1207,19 +1302,20 @@ static void on_newobj_event(DDTRACE_UNUSED VALUE unused1, DDTRACE_UNUSED void *u
     &state->allocation_sampler, HANDLE_CLOCK_FAILURE(monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE))
   );
 
-  // @ivoanjo: Strictly speaking, this is not needed because Ruby should not call the same tracepoint while a previous
-  // invocation is still pending, (e.g. it wouldn't call `on_newobj_event` while it's already running), but I decided
-  // to keep this here for consistency -- every call to the thread context (other than the special gc calls which are
-  // defined as not being able to allocate) sets this.
-  state->during_sample = true;
+  during_sample_enter(state);
 
   // Rescue against any exceptions that happen during sampling
-  safely_call(rescued_sample_allocation, Qnil, state->self_instance);
+  safely_call(
+    rescued_sample_allocation,
+    (VALUE) thread_context,
+    state->self_instance,
+    handle_sampling_failure_rescued_sample_allocation
+  );
 
   if (state->dynamic_sampling_rate_enabled) {
     long now = monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE);
     if (now == 0) {
-      delayed_error(state, ERR_CLOCK_FAIL);
+      delayed_error_clock_failure(state);
       // NOTE: Not short-circuiting here to make sure cleanup happens
     }
     uint64_t sampling_time_ns = discrete_dynamic_sampler_after_sample(&state->allocation_sampler, now);
@@ -1232,7 +1328,7 @@ static void on_newobj_event(DDTRACE_UNUSED VALUE unused1, DDTRACE_UNUSED void *u
 
   state->stats.allocation_sampled++;
 
-  state->during_sample = false;
+  during_sample_exit(state);
 }
 
 static void disable_tracepoints(cpu_and_wall_time_worker_state *state) {
@@ -1263,7 +1359,8 @@ static VALUE _native_with_blocked_sigprof(DDTRACE_UNUSED VALUE self) {
   }
 }
 
-static VALUE rescued_sample_allocation(DDTRACE_UNUSED VALUE unused) {
+static VALUE rescued_sample_allocation(VALUE arg) {
+  per_thread_context *thread_context = (per_thread_context*) arg;
   cpu_and_wall_time_worker_state *state = active_sampler_instance_state; // Read from global variable, see "sampler global state safety" note above
 
   // This should not happen in a normal situation because on_newobj_event already checked for this, but just in case...
@@ -1282,20 +1379,33 @@ static VALUE rescued_sample_allocation(DDTRACE_UNUSED VALUE unused) {
   // To control bias from sampling, we clamp the maximum weight attributed to a single allocation sample. This avoids
   // assigning a very large number to a sample, if for instance the dynamic sampling mechanism chose a really big interval.
   unsigned int weight = allocations_since_last_sample > MAX_ALLOC_WEIGHT ? MAX_ALLOC_WEIGHT : (unsigned int) allocations_since_last_sample;
-  thread_context_collector_sample_allocation(state->thread_context_collector_instance, weight, new_object);
+  bool needs_after_allocation = thread_context_collector_sample_allocation(state->thread_context_collector_instance, thread_context, weight, new_object);
   // ...but we still represent the skipped samples in the profile, thus the data will account for all allocations.
   if (weight < allocations_since_last_sample) {
     uint32_t skipped_samples = (uint32_t) uint64_min_of(allocations_since_last_sample - weight, UINT32_MAX);
     thread_context_collector_sample_skipped_allocation_samples(state->thread_context_collector_instance, skipped_samples);
   }
 
+  if (needs_after_allocation) {
+    #ifndef NO_POSTPONED_TRIGGER
+      rb_postponed_job_trigger(after_allocation_from_postponed_job_handle);
+    #else
+      // Not needed on legacy rubies
+    #endif
+  }
+
   // Return a dummy VALUE because we're called from rb_rescue2 which requires it
   return Qnil;
 }
 
 static void delayed_error(cpu_and_wall_time_worker_state *state, const char *error) {
   // If we can't raise an immediate exception at the calling site, use the asynchronous flow through the main worker loop.
-  stop_state(state, rb_exc_new_cstr(rb_eRuntimeError, error));
+  stop_state(state, rb_exc_new_cstr(rb_eRuntimeError, error), "delayed_error");
+}
+
+static void delayed_error_clock_failure(cpu_and_wall_time_worker_state *state) {
+  // If we can't raise an immediate exception at the calling site, use the asynchronous flow through the main worker loop.
+  stop_state(state, rb_exc_new_cstr(clock_failure_exception_class, "failed to get clock time"), "delayed_error_clock_failure");
 }
 
 static VALUE _native_delayed_error(DDTRACE_UNUSED VALUE self, VALUE instance, VALUE error_msg) {
@@ -1326,44 +1436,65 @@ static VALUE _native_resume_signals(DDTRACE_UNUSED VALUE self) {
 #ifndef NO_GVL_INSTRUMENTATION
   static void on_gvl_event(rb_event_flag_t event_id, const rb_internal_thread_event_data_t *event_data, DDTRACE_UNUSED void *_unused) {
     // Be very careful about touching the `state` here or doing anything at all:
-    // This function gets called without the GVL, and potentially from background Ractors!
-    //
-    // In fact, the `target_thread` that this event is about may not even be the current thread. (So be careful with thread locals that
-    // are not directly tied to the `target_thread` object and the like)
-    gvl_profiling_thread target_thread = thread_from_event(event_data);
+    // This function gets called without the GVL, and potentially from non-main Ractors!
+
+    // The thread that this event is about may not be the current thread
+    // (as documented on rb_internal_thread_add_event_hook(), and this is notably the case for READY on Ruby 4.0),
+    // so be careful with native thread locals that are not directly tied to the thread object and the like.
+
+    // On Ruby 3.2 the event does not carry the thread, but all events always fire on the event thread on Ruby 3.2.
+    // However, during early thread startup rb_thread_current() can crash because the execution context (Fiber) isn't
+    // stored in TLS yet; ruby_native_thread_p() guards against this.
+    #ifdef HAVE_RUBY_THREAD_STORAGE_API
+      VALUE target_thread = event_data->thread;
+    #else
+      if (!ruby_native_thread_p()) return;
+      VALUE target_thread = rb_thread_current();
+    #endif
 
-    if (event_id == RUBY_INTERNAL_THREAD_EVENT_READY) { /* waiting for gvl */
-      thread_context_collector_on_gvl_waiting(target_thread);
+    per_thread_context* thread_context = get_per_thread_context(target_thread);
+    if (!thread_context) return;
+    // If non-NULL the thread is profiled and from the main Ractor
+
+    if (event_id == RUBY_INTERNAL_THREAD_EVENT_SUSPENDED) { /* released gvl */
+      thread_context_collector_on_gvl_released(thread_context);
+    } else if (event_id == RUBY_INTERNAL_THREAD_EVENT_READY) { /* waiting for gvl */
+      thread_context_collector_on_gvl_waiting(thread_context);
     } else if (event_id == RUBY_INTERNAL_THREAD_EVENT_RESUMED) { /* running/runnable */
-      // Interesting note: A RUBY_INTERNAL_THREAD_EVENT_RESUMED is guaranteed to be called with the GVL being acquired.
-      // (And... I think target_thread will be == rb_thread_current()?)
+      // Interesting note: A RUBY_INTERNAL_THREAD_EVENT_RESUMED is guaranteed to be called with the GVL being acquired
+      // and on the event thread.
+      // However, on_gvl_event() is called while holding the scheduler lock, so we do as little work as possible here,
+      // and perform the sample in a postponed_job.
+      cpu_and_wall_time_worker_state *state = active_sampler_instance_state; // Read from global variable, see "sampler global state safety" note above
+      if (state == NULL) return; // This should not happen, but just in case...
+
+      // on_gvl_running prepares a stack sample so we need to gate it with `during_sample_enter` to avoid a
+      // SIGPROF signal coming in during that function and potentially causing a corrupted final state.
       //
-      // But we're not sure if we're on the main Ractor yet. The thread context collector actually can actually help here:
-      // it tags threads it's tracking, so if a thread is tagged then by definition we know that thread belongs to the main
-      // Ractor. Thus, if we get a ON_GVL_RUNNING_UNKNOWN result we shouldn't touch any state, but otherwise we're good to go.
+      // TODO: Currently, the sample prepared in on_gvl_running can still be clobbered if the signal handler runs
+      // after `during_sample_exit` but before `after_gvl_running_from_postponed_job` gets to run. We'll need to fix
+      // that next.
+      during_sample_enter(state);
 
-      #ifdef USE_GVL_PROFILING_3_2_WORKAROUNDS
-        target_thread = gvl_profiling_state_maybe_initialize();
-      #endif
+      on_gvl_running_result result = thread_context_collector_on_gvl_running(state->thread_context_collector_instance, target_thread, thread_context);
 
-      on_gvl_running_result result = thread_context_collector_on_gvl_running(target_thread);
+      during_sample_exit(state);
+
+      if (result.waiting_for_gvl_duration_ns > 0) {
+        state->stats.vm_metrics.gvl_waiting_time_ns_total += (uint64_t) result.waiting_for_gvl_duration_ns;
+      }
 
-      if (result == ON_GVL_RUNNING_SAMPLE) {
+      if (result.action == ON_GVL_RUNNING_SAMPLE) {
         #ifndef NO_POSTPONED_TRIGGER
           rb_postponed_job_trigger(after_gvl_running_from_postponed_job_handle);
         #else
           rb_postponed_job_register_one(0, after_gvl_running_from_postponed_job, NULL);
         #endif
-      } else if (result == ON_GVL_RUNNING_DONT_SAMPLE) {
-        cpu_and_wall_time_worker_state *state = active_sampler_instance_state; // Read from global variable, see "sampler global state safety" note above
-
-        if (state == NULL) return; // This should not happen, but just in case...
-
+      } else if (result.action == ON_GVL_RUNNING_DONT_SAMPLE) {
         state->stats.gvl_dont_sample++;
       }
     } else {
-      // This is a very delicate time and it's hard for us to raise an exception so let's at least complain to stderr
-      fprintf(stderr, "[ddtrace] Unexpected value in on_gvl_event (%d)\n", event_id);
+      rb_bug("[ddtrace] Unexpected value in on_gvl_event (%d)\n", event_id);
     }
   }
 
@@ -1373,12 +1504,17 @@ static VALUE _native_resume_signals(DDTRACE_UNUSED VALUE self) {
     // This can potentially happen if the CpuAndWallTimeWorker was stopped while the postponed job was waiting to be executed; nothing to do
     if (state == NULL) return;
 
-    state->during_sample = true;
+    during_sample_enter(state);
 
     // Rescue against any exceptions that happen during sampling
-    safely_call(rescued_after_gvl_running_from_postponed_job, state->self_instance, state->self_instance);
+    safely_call(
+      rescued_after_gvl_running_from_postponed_job,
+      state->self_instance,
+      state->self_instance,
+      handle_sampling_failure_rescued_after_gvl_running_from_postponed_job
+    );
 
-    state->during_sample = false;
+    during_sample_exit(state);
   }
 
   static VALUE rescued_after_gvl_running_from_postponed_job(VALUE self_instance) {
@@ -1409,8 +1545,90 @@ static VALUE _native_resume_signals(DDTRACE_UNUSED VALUE self) {
 
     return state->gvl_profiling_hook != NULL ? Qtrue : Qfalse;
   }
+
+  static VALUE handle_sampling_failure_rescued_after_gvl_running_from_postponed_job(VALUE self_instance, VALUE exception) {
+    stop(self_instance, exception, "rescued_after_gvl_running_from_postponed_job");
+    return Qnil;
+  }
 #else
   static VALUE _native_gvl_profiling_hook_active(DDTRACE_UNUSED VALUE self, DDTRACE_UNUSED VALUE instance) {
     return Qfalse;
   }
 #endif
+
+static VALUE handle_sampling_failure_rescued_sample_from_postponed_job(VALUE self_instance, VALUE exception) {
+  stop(self_instance, exception, "rescued_sample_from_postponed_job");
+  return Qnil;
+}
+
+static VALUE handle_sampling_failure_thread_context_collector_sample_after_gc(VALUE self_instance, VALUE exception) {
+  stop(self_instance, exception, "thread_context_collector_sample_after_gc");
+  return Qnil;
+}
+
+static VALUE handle_sampling_failure_rescued_sample_allocation(VALUE self_instance, VALUE exception) {
+  stop(self_instance, exception, "rescued_sample_allocation");
+  return Qnil;
+}
+
+static VALUE handle_sampling_failure_rescued_after_allocation(VALUE self_instance, VALUE exception) {
+  stop(self_instance, exception, "rescued_after_allocation");
+  return Qnil;
+}
+
+static VALUE rescued_after_allocation(VALUE self_instance) {
+  cpu_and_wall_time_worker_state *state;
+  TypedData_Get_Struct(self_instance, cpu_and_wall_time_worker_state, &cpu_and_wall_time_worker_typed_data, state);
+
+  thread_context_collector_after_allocation(state->thread_context_collector_instance);
+
+  // Return a dummy VALUE because we're called from rb_rescue2 which requires it
+  return Qnil;
+}
+
+// This postponed job callback is used to finalize heap allocation recordings on Ruby 4+.
+// During on_newobj_event, calling rb_obj_id() is unsafe because it mutates the object.
+// So we defer getting the object_id until after the event completes.
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wunused-function" // This is only used for some Rubies, but we want to build on all to make it easier to dev
+static void after_allocation_from_postponed_job(DDTRACE_UNUSED void *_unused) {
+  cpu_and_wall_time_worker_state *state = active_sampler_instance_state;
+
+  if (state == NULL || !ddtrace_rb_ractor_main_p()) return;
+
+  // Protect against nested operations
+  if (state->during_sample) return;
+
+  during_sample_enter(state);
+
+  // NOTE: We're not updating the allocation_sampler here.
+  // This means work done in this function isn't accounted for as profiler overhead.
+  // This is acceptable as the amount of work done here is expected to be small.
+  safely_call(
+    rescued_after_allocation,
+    state->self_instance,
+    state->self_instance,
+    handle_sampling_failure_rescued_after_allocation
+  );
+
+  during_sample_exit(state);
+}
+#pragma GCC diagnostic pop
+
+static inline void during_sample_enter(cpu_and_wall_time_worker_state* state) {
+  // Tell the compiler it's not allowed to reorder the `during_sample` flag with anything that happens after.
+  //
+  // In a few cases, we may be checking this flag from a signal handler, so we need to make sure the compiler didn't
+  // get clever and reordered things in such a way that makes us miss the flag update.
+  //
+  // See https://github.com/ruby/ruby/pull/11036 for a similar change made to the Ruby VM with more context.
+  state->during_sample = true;
+  atomic_signal_fence(memory_order_seq_cst);
+}
+
+static inline void during_sample_exit(cpu_and_wall_time_worker_state* state) {
+  // See `during_sample_enter` for more context; in this case we set the fence before to make sure anything that
+  // happens before the fence is not reordered with the flag update.
+  atomic_signal_fence(memory_order_seq_cst);
+  state->during_sample = false;
+}