datadog 2.17.0 → 2.19.0

data/ext/datadog_profiling_native_extension/heap_recorder.h

@@ -34,7 +34,7 @@ typedef struct {
 
   // The class of the object that we're tracking.
   // NOTE: This is optional and will be set to NULL if not set.
-  char* class;
+  ddog_prof_ManagedStringId class;
 
   // The GC allocation gen in which we saw this object being allocated.
   //
@@ -59,7 +59,7 @@ typedef struct {
 } heap_recorder_iteration_data;
 
 // Initialize a new heap recorder.
-heap_recorder* heap_recorder_new(void);
+heap_recorder* heap_recorder_new(ddog_prof_ManagedStringStorage string_storage);
 
 // Free a previously initialized heap recorder.
 void heap_recorder_free(heap_recorder *heap_recorder);
@@ -164,10 +164,6 @@ VALUE heap_recorder_state_snapshot(heap_recorder *heap_recorder);
 
 // v--- TEST-ONLY APIs ---v
 
-// Assert internal hashing logic is valid for the provided locations and its
-// corresponding internal representations in heap recorder.
-void heap_recorder_testonly_assert_hash_matches(ddog_prof_Slice_Location locations);
-
 // Returns a Ruby string with a representation of internal data helpful to
 // troubleshoot issues such as unexpected test failures.
 VALUE heap_recorder_testonly_debug(heap_recorder *heap_recorder);
@@ -177,3 +173,5 @@ VALUE heap_recorder_testonly_is_object_recorded(heap_recorder *heap_recorder, VA
 
 // Used to ensure that a GC actually triggers an update of the objects
 void heap_recorder_testonly_reset_last_update(heap_recorder *heap_recorder);
+
+void heap_recorder_testonly_benchmark_intern(heap_recorder *heap_recorder, ddog_CharSlice string, int times, bool use_all);

data/ext/datadog_profiling_native_extension/libdatadog_helpers.c

@@ -60,3 +60,25 @@ size_t read_ddogerr_string_and_drop(ddog_Error *error, char *string, size_t capa
   ddog_Error_drop(error);
   return error_msg_size;
 }
+
+ddog_prof_ManagedStringId intern_or_raise(ddog_prof_ManagedStringStorage string_storage, ddog_CharSlice string) {
+  if (string.len == 0) return (ddog_prof_ManagedStringId) { 0 }; // Id 0 is always an empty string, no need to ask
+
+  ddog_prof_ManagedStringStorageInternResult intern_result = ddog_prof_ManagedStringStorage_intern(string_storage, string);
+  if (intern_result.tag == DDOG_PROF_MANAGED_STRING_STORAGE_INTERN_RESULT_ERR) {
+    rb_raise(rb_eRuntimeError, "Failed to intern string: %"PRIsVALUE, get_error_details_and_drop(&intern_result.err));
+  }
+  return intern_result.ok;
+}
+
+void intern_all_or_raise(
+  ddog_prof_ManagedStringStorage string_storage,
+  ddog_prof_Slice_CharSlice strings,
+  ddog_prof_ManagedStringId *output_ids,
+  uintptr_t output_ids_size
+) {
+  ddog_prof_MaybeError result = ddog_prof_ManagedStringStorage_intern_all(string_storage, strings, output_ids, output_ids_size);
+  if (result.tag == DDOG_PROF_OPTION_ERROR_SOME_ERROR) {
+    rb_raise(rb_eRuntimeError, "Failed to intern_all: %"PRIsVALUE, get_error_details_and_drop(&result.some));
+  }
+}

data/ext/datadog_profiling_native_extension/libdatadog_helpers.h

@@ -18,8 +18,11 @@ size_t read_ddogerr_string_and_drop(ddog_Error *error, char *string, size_t capa
 const char *ruby_value_type_to_string(enum ruby_value_type type);
 ddog_CharSlice ruby_value_type_to_char_slice(enum ruby_value_type type);
 
-// Returns a dynamically allocated string from the provided char slice.
-// WARN: The returned string must be explicitly freed with ruby_xfree.
-static inline char* string_from_char_slice(ddog_CharSlice slice) {
-  return ruby_strndup(slice.ptr, slice.len);
-}
+ddog_prof_ManagedStringId intern_or_raise(ddog_prof_ManagedStringStorage string_storage, ddog_CharSlice string);
+
+void intern_all_or_raise(
+  ddog_prof_ManagedStringStorage string_storage,
+  ddog_prof_Slice_CharSlice strings,
+  ddog_prof_ManagedStringId *output_ids,
+  uintptr_t output_ids_size
+);

data/ext/datadog_profiling_native_extension/private_vm_api_access.c

@@ -212,21 +212,6 @@ uint64_t native_thread_id_for(VALUE thread) {
   #endif
 }
 
-// Returns the stack depth by using the same approach as rb_profile_frames and backtrace_each: get the positions
-// of the end and current frame pointers and subtracting them.
-ptrdiff_t stack_depth_for(VALUE thread) {
-  const rb_execution_context_t *ec = thread_struct_from_object(thread)->ec;
-  const rb_control_frame_t *cfp = ec->cfp, *end_cfp = RUBY_VM_END_CONTROL_FRAME(ec);
-
-  if (end_cfp == NULL) return 0;
-
-  // Skip dummy frame, as seen in `backtrace_each` (`vm_backtrace.c`) and our custom rb_profile_frames
-  // ( https://github.com/ruby/ruby/blob/4bd38e8120f2fdfdd47a34211720e048502377f1/vm_backtrace.c#L890-L914 )
-  end_cfp = RUBY_VM_NEXT_CONTROL_FRAME(end_cfp);
-
-  return end_cfp <= cfp ? 0 : end_cfp - cfp - 1;
-}
-
 // This was renamed in Ruby 3.2
 #if !defined(ccan_list_for_each) && defined(list_for_each)
   #define ccan_list_for_each list_for_each
@@ -360,11 +345,16 @@ calc_pos(const rb_iseq_t *iseq, const VALUE *pc, int *lineno, int *node_id)
         }
 #endif
 
-        // In PROF-11475 we spotted a crash when calling `rb_iseq_line_no` from this method. We couldn't reproduce or
-        // figure out the root cause, but "just in case", we're validating that the iseq looks valid and that the
-        // `n` used for the position is also sane, and if they don't look good, we don't calculate the line, rather
-        // than potentially trigger any issues.
-        if (RB_UNLIKELY(!RB_TYPE_P((VALUE) iseq, T_IMEMO) || n < 0 || n > ISEQ_BODY(iseq)->iseq_size)) return 0;
+        // In PROF-11475 we spotted a crash when calling `rb_iseq_line_no` from this method.
+        // We were only able to reproduce this issue on Ruby 2.6 and 2.7, not 2.5 or the 3.x series (tried 3.0, 3.2 and 3.4).
+        // Note that going out of bounds doesn't crash every time, as usual with C we may just read garbage or get lucky.
+        //
+        // For those problematic Rubies, we observed that when we try to take a sample in the middle of processing the
+        // VM `LEAVE` instruction, the value of `n` can violate the documented assumptions above and be
+        // `n > ISEQ_BODY(iseq)->iseq_size)`.
+        //
+        // To work around this and any other potential issues, we validate here that the bytecode position is sane.
+        if (RB_UNLIKELY(n < 0 || n > ISEQ_BODY(iseq)->iseq_size)) return 0;
 
         if (lineno) *lineno = rb_iseq_line_no(iseq, pos);
 #ifdef USE_ISEQ_NODE_ID
@@ -410,6 +400,9 @@ calc_lineno(const rb_iseq_t *iseq, const VALUE *pc)
 // * Add frame_flags.same_frame and logic to skip redoing work if the buffer already contains the same data we're collecting
 // * Skipped use of rb_callable_method_entry_t (cme) for Ruby frames as it doesn't impact us.
 // * Imported fix from https://github.com/ruby/ruby/pull/8280 to keep us closer to upstream
+// * Added potential fix for https://github.com/ruby/ruby/pull/13643 (this one is a just-in-case, unclear if it happens
+//   for ddtrace)
+// * Reversed order of iteration to better enable caching
 //
 // What is rb_profile_frames?
 // `rb_profile_frames` is a Ruby VM debug API added for use by profilers for sampling the stack trace of a Ruby thread.
@@ -445,6 +438,16 @@ int ddtrace_rb_profile_frames(VALUE thread, int start, int limit, frame_info *st
     // support sampling any thread (including the current) passed as an argument
     rb_thread_t *th = thread_struct_from_object(thread);
     const rb_execution_context_t *ec = th->ec;
+
+    // As of this writing, we don't support profiling with MN enabled, and this only happens in that mode, but as we
+    // probably want to experiment with it in the future, I've decided to import https://github.com/ruby/ruby/pull/9310
+    // here.
+    if (ec == NULL) return 0;
+
+    // I suspect this won't happen for ddtrace, but just-in-case we've imported a potential fix for
+    // https://github.com/ruby/ruby/pull/13643 by assuming that these can be NULL/zero with the cfp being non-NULL yet.
+    if (ec->vm_stack == NULL || ec->vm_stack_size == 0) return 0;
+
     const rb_control_frame_t *cfp = ec->cfp, *end_cfp = RUBY_VM_END_CONTROL_FRAME(ec);
     #ifndef NO_JIT_RETURN
       const rb_control_frame_t *top = cfp;
@@ -462,11 +465,6 @@ int ddtrace_rb_profile_frames(VALUE thread, int start, int limit, frame_info *st
     // it from https://github.com/ruby/ruby/pull/7116 in a "just in case" kind of mindset.
     if (cfp == NULL) return 0;
 
-    // As of this writing, we don't support profiling with MN enabled, and this only happens in that mode, but as we
-    // probably want to experiment with it in the future, I've decided to import https://github.com/ruby/ruby/pull/9310
-    // here.
-    if (ec == NULL) return 0;
-
     // Fix: Skip dummy frame that shows up in main thread.
     //
     // According to a comment in `backtrace_each` (`vm_backtrace.c`), there's two dummy frames that we should ignore
@@ -483,7 +481,20 @@ int ddtrace_rb_profile_frames(VALUE thread, int start, int limit, frame_info *st
     // See comment on `record_placeholder_stack_in_native_code` for a full explanation of what this means (and why we don't just return 0)
     if (end_cfp <= cfp) return PLACEHOLDER_STACK_IN_NATIVE_CODE;
 
-    for (i=0; i<limit && cfp != end_cfp; cfp = RUBY_VM_PREVIOUS_CONTROL_FRAME(cfp)) {
+    // This is the position just after the top of the stack -- e.g. where a new frame pushed on the stack would end up.
+    const rb_control_frame_t *top_sentinel = RUBY_VM_NEXT_CONTROL_FRAME(cfp);
+
+    // We iterate the stack from bottom (beginning of thread) to the top (currently-active frame). This is different
+    // from upstream rb_profile_frames, but actually matches what `backtrace_each` does (yes, different Ruby VM APIs
+    // iterate in different directions).
+    // We do this to better take advantage of the `same_frame` caching mechanism: By starting from the bottom of the
+    // stack towards the top, we can usually keep most of the stack intact when the code is only going up and down
+    // a few methods at the top. Before this change, the cache was really only useful if between samples the app had
+    // not moved from the current stack, as adding or removing one frame would invalidate the existing cache (because
+    // every position would shift).
+    cfp = RUBY_VM_NEXT_CONTROL_FRAME(end_cfp);
+
+    for (i=0; i<limit && cfp != top_sentinel; cfp = RUBY_VM_NEXT_CONTROL_FRAME(cfp)) {
         if (cfp->iseq && !cfp->pc) {
           // Fix: Do nothing -- this frame should not be used
           //
@@ -569,6 +580,7 @@ int ddtrace_rb_profile_frames(VALUE thread, int start, int limit, frame_info *st
 
                 stack_buffer[i].as.native_frame.caching_cme = (VALUE)cme;
                 stack_buffer[i].as.native_frame.method_id = cme->def->original_id;
+                stack_buffer[i].as.native_frame.function = cme->def->body.cfunc.func;
                 stack_buffer[i].is_ruby_frame = false;
                 i++;
             }

data/ext/datadog_profiling_native_extension/private_vm_api_access.h

@@ -18,16 +18,19 @@ typedef struct {
   rb_nativethread_id_t owner;
 } current_gvl_owner;
 
+// If a sample is kept around for later use, some of its fields need marking. Remember to
+// update the marking code in `sampling_buffer_mark` if new fields are added.
 typedef struct {
   union {
     struct {
-      VALUE iseq;
-      void *caching_pc; // For caching only
+      VALUE iseq; // Needs marking if kept around
+      void *caching_pc; // For caching validation/invalidation only (does not need marking)
       int line;
     } ruby_frame;
     struct {
-      VALUE caching_cme; // For caching only
+      VALUE caching_cme; // For caching validation/invalidation only (does not need marking)
       ID method_id;
+      void *function;
     } native_frame;
   } as;
   bool is_ruby_frame : 1;
@@ -38,7 +41,6 @@ rb_nativethread_id_t pthread_id_for(VALUE thread);
 bool is_current_thread_holding_the_gvl(void);
 current_gvl_owner gvl_owner(void);
 uint64_t native_thread_id_for(VALUE thread);
-ptrdiff_t stack_depth_for(VALUE thread);
 void ddtrace_thread_list(VALUE result_array);
 bool is_thread_alive(VALUE thread);
 VALUE thread_name_for(VALUE thread);

data/ext/datadog_profiling_native_extension/ruby_helpers.c

@@ -103,16 +103,6 @@ void raise_syserr(
   }
 }
 
-char* ruby_strndup(const char *str, size_t size) {
-  char *dup;
-
-  dup = xmalloc(size + 1);
-  memcpy(dup, str, size);
-  dup[size] = '\0';
-
-  return dup;
-}
-
 static VALUE _id2ref(VALUE obj_id) {
   // Call ::ObjectSpace._id2ref natively. It will raise if the id is no longer valid
   return rb_funcall(module_object_space, _id2ref_id, 1, obj_id);
@@ -122,9 +112,7 @@ static VALUE _id2ref_failure(DDTRACE_UNUSED VALUE _unused1, DDTRACE_UNUSED VALUE
   return Qfalse;
 }
 
-// Native wrapper to get an object ref from an id. Returns true on success and
-// writes the ref to the value pointer parameter if !NULL. False if id doesn't
-// reference a valid object (in which case value is not changed).
+// See notes on header for important details
 bool ruby_ref_from_id(VALUE obj_id, VALUE *value) {
   // Call ::ObjectSpace._id2ref natively. It will raise if the id is no longer valid
   // so we need to call it via rb_rescue2

data/ext/datadog_profiling_native_extension/ruby_helpers.h

@@ -67,20 +67,12 @@ NORETURN(void raise_syserr(
   const char *function_name
 ));
 
-// Alternative to ruby_strdup that takes a size argument.
-// Similar to C's strndup but slightly less smart as size is expected to
-// be smaller or equal to the real size of str (minus null termination if it
-// exists).
-// A new string will be returned with size+1 bytes and last byte set to '\0'.
-// The returned string must be freed explicitly.
-//
-// WARN: Cannot be used during GC or outside the GVL.
-char* ruby_strndup(const char *str, size_t size);
-
 // Native wrapper to get an object ref from an id. Returns true on success and
 // writes the ref to the value pointer parameter if !NULL. False if id doesn't
 // reference a valid object (in which case value is not changed).
-bool ruby_ref_from_id(size_t id, VALUE *value);
+//
+// Note: GVL can be released and other threads may get to run before this method returns
+bool ruby_ref_from_id(VALUE obj_id, VALUE *value);
 
 // Native wrapper to get the approximate/estimated current size of the passed
 // object.

data/ext/datadog_profiling_native_extension/stack_recorder.c

@@ -196,6 +196,10 @@ typedef struct {
   pthread_mutex_t mutex_slot_two;
   profile_slot profile_slot_two;
 
+  ddog_prof_ManagedStringStorage string_storage;
+  ddog_prof_ManagedStringId label_key_allocation_class;
+  ddog_prof_ManagedStringId label_key_gc_gen_age;
+
   short active_slot; // MUST NEVER BE ACCESSED FROM record_sample; this is NOT for the sampler thread to use.
 
   uint8_t position_for[ALL_VALUE_TYPES_COUNT];
@@ -230,6 +234,7 @@ typedef struct {
   ddog_prof_Profile_SerializeResult result;
   long heap_profile_build_time_ns;
   long serialize_no_gvl_time_ns;
+  ddog_prof_MaybeError advance_gen_result;
 
   // Set by both
   bool serialize_ran;
@@ -256,7 +261,6 @@ static void serializer_set_start_timestamp_for_next_profile(stack_recorder_state
 static VALUE _native_record_endpoint(DDTRACE_UNUSED VALUE _self, VALUE recorder_instance, VALUE local_root_span_id, VALUE endpoint);
 static void reset_profile_slot(profile_slot *slot, ddog_Timespec start_timestamp);
 static VALUE _native_track_object(DDTRACE_UNUSED VALUE _self, VALUE recorder_instance, VALUE new_obj, VALUE weight, VALUE alloc_class);
-static VALUE _native_check_heap_hashes(DDTRACE_UNUSED VALUE _self, VALUE locations);
 static VALUE _native_start_fake_slow_heap_serialization(DDTRACE_UNUSED VALUE _self, VALUE recorder_instance);
 static VALUE _native_end_fake_slow_heap_serialization(DDTRACE_UNUSED VALUE _self, VALUE recorder_instance);
 static VALUE _native_debug_heap_recorder(DDTRACE_UNUSED VALUE _self, VALUE recorder_instance);
@@ -265,6 +269,8 @@ static VALUE build_profile_stats(profile_slot *slot, long serialization_time_ns,
 static VALUE _native_is_object_recorded(DDTRACE_UNUSED VALUE _self, VALUE recorder_instance, VALUE object_id);
 static VALUE _native_heap_recorder_reset_last_update(DDTRACE_UNUSED VALUE _self, VALUE recorder_instance);
 static VALUE _native_recorder_after_gc_step(DDTRACE_UNUSED VALUE _self, VALUE recorder_instance);
+static VALUE _native_benchmark_intern(DDTRACE_UNUSED VALUE _self, VALUE recorder_instance, VALUE string, VALUE times, VALUE use_all);
+static VALUE _native_test_managed_string_storage_produces_valid_profiles(DDTRACE_UNUSED VALUE _self);
 
 void stack_recorder_init(VALUE profiling_module) {
   VALUE stack_recorder_class = rb_define_class_under(profiling_module, "StackRecorder", rb_cObject);
@@ -290,7 +296,6 @@ void stack_recorder_init(VALUE profiling_module) {
   rb_define_singleton_method(testing_module, "_native_slot_two_mutex_locked?", _native_is_slot_two_mutex_locked, 1);
   rb_define_singleton_method(testing_module, "_native_record_endpoint", _native_record_endpoint, 3);
   rb_define_singleton_method(testing_module, "_native_track_object", _native_track_object, 4);
-  rb_define_singleton_method(testing_module, "_native_check_heap_hashes", _native_check_heap_hashes, 1);
   rb_define_singleton_method(testing_module, "_native_start_fake_slow_heap_serialization",
       _native_start_fake_slow_heap_serialization, 1);
   rb_define_singleton_method(testing_module, "_native_end_fake_slow_heap_serialization",
@@ -300,6 +305,8 @@ void stack_recorder_init(VALUE profiling_module) {
   rb_define_singleton_method(testing_module, "_native_is_object_recorded?", _native_is_object_recorded, 2);
   rb_define_singleton_method(testing_module, "_native_heap_recorder_reset_last_update", _native_heap_recorder_reset_last_update, 1);
   rb_define_singleton_method(testing_module, "_native_recorder_after_gc_step", _native_recorder_after_gc_step, 1);
+  rb_define_singleton_method(testing_module, "_native_benchmark_intern", _native_benchmark_intern, 4);
+  rb_define_singleton_method(testing_module, "_native_test_managed_string_storage_produces_valid_profiles", _native_test_managed_string_storage_produces_valid_profiles, 0);
 
   ok_symbol = ID2SYM(rb_intern_const("ok"));
   error_symbol = ID2SYM(rb_intern_const("error"));
@@ -334,17 +341,27 @@ static VALUE _native_new(VALUE klass) {
     .serialization_time_ns_min = INT64_MAX,
   };
 
-  // Note: At this point, slot_one_profile/slot_two_profile contain null pointers. Libdatadog validates pointers
+  // Note: At this point, slot_one_profile/slot_two_profile/string_storage contain null pointers. Libdatadog validates pointers
   // before using them so it's ok for us to go ahead and create the StackRecorder object.
 
   VALUE stack_recorder = TypedData_Wrap_Struct(klass, &stack_recorder_typed_data, state);
 
+  ddog_prof_ManagedStringStorageNewResult string_storage = ddog_prof_ManagedStringStorage_new();
+
+  if (string_storage.tag == DDOG_PROF_MANAGED_STRING_STORAGE_NEW_RESULT_ERR) {
+    rb_raise(rb_eRuntimeError, "Failed to initialize string storage: %"PRIsVALUE, get_error_details_and_drop(&string_storage.err));
+  }
+
+  state->string_storage = string_storage.ok;
+  state->label_key_allocation_class = intern_or_raise(state->string_storage, DDOG_CHARSLICE_C("allocation class"));
+  state->label_key_gc_gen_age = intern_or_raise(state->string_storage, DDOG_CHARSLICE_C("gc gen age"));
+
+  initialize_profiles(state, sample_types);
+
   // NOTE: We initialize this because we want a new recorder to be operational even before #initialize runs and our
   //       default is everything enabled. However, if during recording initialization it turns out we don't want
   //       heap samples, we will free and reset heap_recorder back to NULL.
-  state->heap_recorder = heap_recorder_new();
-
-  initialize_profiles(state, sample_types);
+  state->heap_recorder = heap_recorder_new(state->string_storage);
 
   return stack_recorder;
 }
@@ -363,7 +380,7 @@ static void initialize_profiles(stack_recorder_state *state, ddog_prof_Slice_Val
   ddog_Timespec start_timestamp = system_epoch_now_timespec();
 
   ddog_prof_Profile_NewResult slot_one_profile_result =
-    ddog_prof_Profile_new(sample_types, NULL /* period is optional */);
+    ddog_prof_Profile_with_string_storage(sample_types, NULL /* period is optional */, state->string_storage);
 
   if (slot_one_profile_result.tag == DDOG_PROF_PROFILE_NEW_RESULT_ERR) {
     rb_raise(rb_eRuntimeError, "Failed to initialize slot one profile: %"PRIsVALUE, get_error_details_and_drop(&slot_one_profile_result.err));
@@ -372,7 +389,7 @@ static void initialize_profiles(stack_recorder_state *state, ddog_prof_Slice_Val
   state->profile_slot_one = (profile_slot) { .profile = slot_one_profile_result.ok, .start_timestamp = start_timestamp };
 
   ddog_prof_Profile_NewResult slot_two_profile_result =
-    ddog_prof_Profile_new(sample_types, NULL /* period is optional */);
+    ddog_prof_Profile_with_string_storage(sample_types, NULL /* period is optional */, state->string_storage);
 
   if (slot_two_profile_result.tag == DDOG_PROF_PROFILE_NEW_RESULT_ERR) {
     // Note: No need to take any special care of slot one, it'll get cleaned up by stack_recorder_typed_data_free
@@ -393,6 +410,8 @@ static void stack_recorder_typed_data_free(void *state_ptr) {
 
   heap_recorder_free(state->heap_recorder);
 
+  ddog_prof_ManagedStringStorage_drop(state->string_storage);
+
   ruby_xfree(state);
 }
 
@@ -519,6 +538,8 @@ static VALUE _native_serialize(DDTRACE_UNUSED VALUE _self, VALUE recorder_instan
   long heap_iteration_prep_start_time_ns = monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE);
   // Prepare the iteration on heap recorder we'll be doing outside the GVL. The preparation needs to
   // happen while holding on to the GVL.
+  // NOTE: While rare, it's possible for the GVL to be released inside this function (see comments on `heap_recorder_update`)
+  // and thus don't assume this is an "atomic" step -- other threads may get some running time in the meanwhile.
   heap_recorder_prepare_iteration(state->heap_recorder);
   long heap_iteration_prep_time_ns = monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE) - heap_iteration_prep_start_time_ns;
 
@@ -527,7 +548,7 @@ static VALUE _native_serialize(DDTRACE_UNUSED VALUE _self, VALUE recorder_instan
   call_serialize_without_gvl_arguments args = {
     .state = state,
     .finish_timestamp = finish_timestamp,
-    .serialize_ran = false
+    .serialize_ran = false,
   };
 
   while (!args.serialize_ran) {
@@ -551,13 +572,9 @@ static VALUE _native_serialize(DDTRACE_UNUSED VALUE _self, VALUE recorder_instan
   //       really cover the full serialization process but it gives a more useful number since it bypasses
   //       the noise of acquiring GVLs and dealing with interruptions which is highly specific to runtime
   //       conditions and over which we really have no control about.
-  long serialization_time_ns = args.serialize_no_gvl_time_ns;
-  if (serialization_time_ns >= 0) {
-    // Only update stats if our serialization time is valid.
-    state->stats_lifetime.serialization_time_ns_max = long_max_of(state->stats_lifetime.serialization_time_ns_max, serialization_time_ns);
-    state->stats_lifetime.serialization_time_ns_min = long_min_of(state->stats_lifetime.serialization_time_ns_min, serialization_time_ns);
-    state->stats_lifetime.serialization_time_ns_total += serialization_time_ns;
-  }
+  state->stats_lifetime.serialization_time_ns_max = long_max_of(state->stats_lifetime.serialization_time_ns_max, args.serialize_no_gvl_time_ns);
+  state->stats_lifetime.serialization_time_ns_min = long_min_of(state->stats_lifetime.serialization_time_ns_min, args.serialize_no_gvl_time_ns);
+  state->stats_lifetime.serialization_time_ns_total += args.serialize_no_gvl_time_ns;
 
   ddog_prof_Profile_SerializeResult serialized_profile = args.result;
 
@@ -566,14 +583,20 @@ static VALUE _native_serialize(DDTRACE_UNUSED VALUE _self, VALUE recorder_instan
     return rb_ary_new_from_args(2, error_symbol, get_error_details_and_drop(&serialized_profile.err));
   }
 
+  // Note: If we got here, the profile serialized correctly.
+  // Once we wrap this into a Ruby object, our `EncodedProfile` class will automatically manage memory for it and we
+  // can raise exceptions without worrying about leaking the profile.
   state->stats_lifetime.serialization_successes++;
-
-  // Once we wrap this into a Ruby object, our `EncodedProfile` class will automatically manage memory for it
   VALUE encoded_profile = from_ddog_prof_EncodedProfile(serialized_profile.ok);
 
+  ddog_prof_MaybeError result = args.advance_gen_result;
+  if (result.tag == DDOG_PROF_OPTION_ERROR_SOME_ERROR) {
+    rb_raise(rb_eRuntimeError, "Failed to advance string storage gen: %"PRIsVALUE, get_error_details_and_drop(&result.some));
+  }
+
   VALUE start = ruby_time_from(args.slot->start_timestamp);
   VALUE finish = ruby_time_from(finish_timestamp);
-  VALUE profile_stats = build_profile_stats(args.slot, serialization_time_ns, heap_iteration_prep_time_ns, args.heap_profile_build_time_ns);
+  VALUE profile_stats = build_profile_stats(args.slot, args.serialize_no_gvl_time_ns, heap_iteration_prep_time_ns, args.heap_profile_build_time_ns);
 
   return rb_ary_new_from_args(2, ok_symbol, rb_ary_new_from_args(4, start, finish, encoded_profile, profile_stats));
 }
@@ -696,18 +719,15 @@ static bool add_heap_sample_to_active_profile_without_gvl(heap_recorder_iteratio
   ddog_prof_Label labels[2];
   size_t label_offset = 0;
 
-  if (object_data->class != NULL) {
+  if (object_data->class.value > 0) {
     labels[label_offset++] = (ddog_prof_Label) {
-      .key = DDOG_CHARSLICE_C("allocation class"),
-      .str = (ddog_CharSlice) {
-        .ptr = object_data->class,
-        .len = strlen(object_data->class),
-      },
+      .key_id = context->state->label_key_allocation_class,
+      .str_id = object_data->class,
       .num = 0, // This shouldn't be needed but the tracer-2.7 docker image ships a buggy gcc that complains about this
     };
   }
   labels[label_offset++] = (ddog_prof_Label) {
-    .key = DDOG_CHARSLICE_C("gc gen age"),
+    .key_id = context->state->label_key_gc_gen_age,
     .num = object_data->gen_age,
   };
 
@@ -772,8 +792,9 @@ static void *call_serialize_without_gvl(void *call_args) {
 
   // Note: The profile gets reset by the serialize call
   args->result = ddog_prof_Profile_serialize(&args->slot->profile, &args->slot->start_timestamp, &args->finish_timestamp);
+  args->advance_gen_result = ddog_prof_ManagedStringStorage_advance_gen(args->state->string_storage);
   args->serialize_ran = true;
-  args->serialize_no_gvl_time_ns = monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE) - serialize_no_gvl_start_time_ns;
+  args->serialize_no_gvl_time_ns = long_max_of(0, monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE) - serialize_no_gvl_start_time_ns);
 
   return NULL; // Unused
 }
@@ -917,36 +938,6 @@ static VALUE _native_track_object(DDTRACE_UNUSED VALUE _self, VALUE recorder_ins
   return Qtrue;
 }
 
-static VALUE _native_check_heap_hashes(DDTRACE_UNUSED VALUE _self, VALUE locations) {
-  ENFORCE_TYPE(locations, T_ARRAY);
-  size_t locations_len = rb_array_len(locations);
-  ddog_prof_Location locations_arr[locations_len];
-  for (size_t i = 0; i < locations_len; i++) {
-    VALUE location = rb_ary_entry(locations, i);
-    ENFORCE_TYPE(location, T_ARRAY);
-    VALUE name = rb_ary_entry(location, 0);
-    VALUE filename = rb_ary_entry(location, 1);
-    VALUE line = rb_ary_entry(location, 2);
-    ENFORCE_TYPE(name, T_STRING);
-    ENFORCE_TYPE(filename, T_STRING);
-    ENFORCE_TYPE(line, T_FIXNUM);
-    locations_arr[i] = (ddog_prof_Location) {
-      .line = line,
-        .function = (ddog_prof_Function) {
-          .name = char_slice_from_ruby_string(name),
-          .filename = char_slice_from_ruby_string(filename),
-        }
-    };
-  }
-  ddog_prof_Slice_Location ddog_locations = {
-    .len = locations_len,
-    .ptr = locations_arr,
-  };
-  heap_recorder_testonly_assert_hash_matches(ddog_locations);
-
-  return Qnil;
-}
-
 static void reset_profile_slot(profile_slot *slot, ddog_Timespec start_timestamp) {
   ddog_prof_Profile_Result reset_result = ddog_prof_Profile_reset(&slot->profile);
   if (reset_result.tag == DDOG_PROF_PROFILE_RESULT_ERR) {
@@ -1046,3 +1037,109 @@ static VALUE _native_recorder_after_gc_step(DDTRACE_UNUSED VALUE _self, VALUE re
   recorder_after_gc_step(recorder_instance);
   return Qtrue;
 }
+
+static VALUE _native_benchmark_intern(DDTRACE_UNUSED VALUE _self, VALUE recorder_instance, VALUE string, VALUE times, VALUE use_all) {
+  ENFORCE_TYPE(string, T_STRING);
+  ENFORCE_TYPE(times, T_FIXNUM);
+  ENFORCE_BOOLEAN(use_all);
+
+  stack_recorder_state *state;
+  TypedData_Get_Struct(recorder_instance, stack_recorder_state, &stack_recorder_typed_data, state);
+
+  heap_recorder_testonly_benchmark_intern(state->heap_recorder, char_slice_from_ruby_string(string), FIX2INT(times), use_all == Qtrue);
+
+  return Qtrue;
+}
+
+// See comments in rspec test for details on what we're testing here.
+static VALUE _native_test_managed_string_storage_produces_valid_profiles(DDTRACE_UNUSED VALUE _self) {
+  ddog_prof_ManagedStringStorageNewResult string_storage = ddog_prof_ManagedStringStorage_new();
+
+  if (string_storage.tag == DDOG_PROF_MANAGED_STRING_STORAGE_NEW_RESULT_ERR) {
+    rb_raise(rb_eRuntimeError, "Failed to initialize string storage: %"PRIsVALUE, get_error_details_and_drop(&string_storage.err));
+  }
+
+  ddog_prof_Slice_ValueType sample_types = {.ptr = all_value_types, .len = ALL_VALUE_TYPES_COUNT};
+  ddog_prof_Profile_NewResult profile = ddog_prof_Profile_with_string_storage(sample_types, NULL, string_storage.ok);
+
+  if (profile.tag == DDOG_PROF_PROFILE_NEW_RESULT_ERR) {
+    rb_raise(rb_eRuntimeError, "Failed to initialize profile: %"PRIsVALUE, get_error_details_and_drop(&profile.err));
+  }
+
+  ddog_prof_ManagedStringId hello = intern_or_raise(string_storage.ok, DDOG_CHARSLICE_C("hello"));
+  ddog_prof_ManagedStringId world = intern_or_raise(string_storage.ok, DDOG_CHARSLICE_C("world"));
+  ddog_prof_ManagedStringId key   = intern_or_raise(string_storage.ok, DDOG_CHARSLICE_C("key"));
+
+  int64_t metric_values[] = {1, 2, 3, 4, 5, 6, 7, 8};
+  ddog_prof_Label labels[] = {{.key_id = key, .str_id = key}};
+
+  ddog_prof_Location locations[] = {
+    (ddog_prof_Location) {
+      .mapping = {.filename = DDOG_CHARSLICE_C(""), .build_id = DDOG_CHARSLICE_C(""), .build_id_id = {}},
+      .function = {
+        .name = DDOG_CHARSLICE_C(""),
+        .name_id = hello,
+        .filename = DDOG_CHARSLICE_C(""),
+        .filename_id = world,
+      },
+      .line = 1,
+    }
+  };
+
+  ddog_prof_Profile_Result result = ddog_prof_Profile_add(
+    &profile.ok,
+    (ddog_prof_Sample) {
+      .locations = (ddog_prof_Slice_Location) { .ptr = locations, .len = 1},
+      .values = (ddog_Slice_I64) {.ptr = metric_values, .len = 8},
+      .labels = (ddog_prof_Slice_Label) { .ptr = labels, .len = 1 }
+    },
+    0
+  );
+
+  if (result.tag == DDOG_PROF_PROFILE_RESULT_ERR) {
+    rb_raise(rb_eArgError, "Failed to record sample: %"PRIsVALUE, get_error_details_and_drop(&result.err));
+  }
+
+  ddog_Timespec finish_timestamp = system_epoch_now_timespec();
+  ddog_Timespec start_timestamp = {.seconds = finish_timestamp.seconds - 60};
+  ddog_prof_Profile_SerializeResult serialize_result = ddog_prof_Profile_serialize(&profile.ok, &start_timestamp, &finish_timestamp);
+
+  if (serialize_result.tag == DDOG_PROF_PROFILE_SERIALIZE_RESULT_ERR) {
+    rb_raise(rb_eRuntimeError, "Failed to serialize: %"PRIsVALUE, get_error_details_and_drop(&serialize_result.err));
+  }
+
+  ddog_prof_MaybeError advance_gen_result = ddog_prof_ManagedStringStorage_advance_gen(string_storage.ok);
+
+  if (advance_gen_result.tag == DDOG_PROF_OPTION_ERROR_SOME_ERROR) {
+    rb_raise(rb_eRuntimeError, "Failed to advance string storage gen: %"PRIsVALUE, get_error_details_and_drop(&advance_gen_result.some));
+  }
+
+  VALUE encoded_pprof_1 = from_ddog_prof_EncodedProfile(serialize_result.ok);
+
+  result = ddog_prof_Profile_add(
+    &profile.ok,
+    (ddog_prof_Sample) {
+      .locations = (ddog_prof_Slice_Location) { .ptr = locations, .len = 1},
+      .values = (ddog_Slice_I64) {.ptr = metric_values, .len = 8},
+      .labels = (ddog_prof_Slice_Label) { .ptr = labels, .len = 1 }
+    },
+    0
+  );
+
+  if (result.tag == DDOG_PROF_PROFILE_RESULT_ERR) {
+    rb_raise(rb_eArgError, "Failed to record sample: %"PRIsVALUE, get_error_details_and_drop(&result.err));
+  }
+
+  serialize_result = ddog_prof_Profile_serialize(&profile.ok, &start_timestamp, &finish_timestamp);
+
+  if (serialize_result.tag == DDOG_PROF_PROFILE_SERIALIZE_RESULT_ERR) {
+    rb_raise(rb_eArgError, "Failed to serialize: %"PRIsVALUE, get_error_details_and_drop(&serialize_result.err));
+  }
+
+  VALUE encoded_pprof_2 = from_ddog_prof_EncodedProfile(serialize_result.ok);
+
+  ddog_prof_Profile_drop(&profile.ok);
+  ddog_prof_ManagedStringStorage_drop(string_storage.ok);
+
+  return rb_ary_new_from_args(2, encoded_pprof_1, encoded_pprof_2);
+}

data/ext/libdatadog_api/extconf.rb

@@ -72,8 +72,8 @@ if ENV['DDTRACE_DEBUG'] == 'true'
 end
 
 # If we got here, libdatadog is available and loaded
-ENV['PKG_CONFIG_PATH'] = "#{ENV['PKG_CONFIG_PATH']}:#{Libdatadog.pkgconfig_folder}"
-Logging.message("[datadog] PKG_CONFIG_PATH set to #{ENV['PKG_CONFIG_PATH'].inspect}\n")
+ENV['PKG_CONFIG_PATH'] = "#{ENV["PKG_CONFIG_PATH"]}:#{Libdatadog.pkgconfig_folder}"
+Logging.message("[datadog] PKG_CONFIG_PATH set to #{ENV["PKG_CONFIG_PATH"].inspect}\n")
 $stderr.puts("Using libdatadog #{Libdatadog::VERSION} from #{Libdatadog.pkgconfig_folder}")
 
 unless pkg_config('datadog_profiling_with_rpath')