google-protobuf 3.21.12 → 3.24.2

data/ext/google/protobuf_c/protobuf.c

@@ -171,6 +171,8 @@ void StringBuilder_PrintMsgval(StringBuilder *b, upb_MessageValue val,
 
 typedef struct {
   upb_Arena *arena;
+  // IMPORTANT: WB_PROTECTED objects must only use the RB_OBJ_WRITE()
+  // macro to update VALUE references, as to trigger write barriers.
   VALUE pinned_objs;
 } Arena;
 
@@ -190,10 +192,11 @@ static VALUE cArena;
 const rb_data_type_t Arena_type = {
     "Google::Protobuf::Internal::Arena",
     {Arena_mark, Arena_free, NULL},
-    .flags = RUBY_TYPED_FREE_IMMEDIATELY,
+    .flags = RUBY_TYPED_FREE_IMMEDIATELY | RUBY_TYPED_WB_PROTECTED,
 };
 
-static void* ruby_upb_allocfunc(upb_alloc* alloc, void* ptr, size_t oldsize, size_t size) {
+static void *ruby_upb_allocfunc(upb_alloc *alloc, void *ptr, size_t oldsize,
+                                size_t size) {
   if (size == 0) {
     xfree(ptr);
     return NULL;
@@ -233,7 +236,7 @@ void Arena_Pin(VALUE _arena, VALUE obj) {
   Arena *arena;
   TypedData_Get_Struct(_arena, Arena, &Arena_type, arena);
   if (arena->pinned_objs == Qnil) {
-    arena->pinned_objs = rb_ary_new();
+    RB_OBJ_WRITE(_arena, &arena->pinned_objs, rb_ary_new());
   }
   rb_ary_push(arena->pinned_objs, obj);
 }
@@ -250,164 +253,48 @@ void Arena_register(VALUE module) {
 // Object Cache
 // -----------------------------------------------------------------------------
 
-// A pointer -> Ruby Object cache that keeps references to Ruby wrapper
-// objects.  This allows us to look up any Ruby wrapper object by the address
-// of the object it is wrapping. That way we can avoid ever creating two
-// different wrapper objects for the same C object, which saves memory and
-// preserves object identity.
-//
-// We use WeakMap for the cache. For Ruby <2.7 we also need a secondary Hash
-// to store WeakMap keys because Ruby <2.7 WeakMap doesn't allow non-finalizable
-// keys.
-//
-// We also need the secondary Hash if sizeof(long) < sizeof(VALUE), because this
-// means it may not be possible to fit a pointer into a Fixnum. Keys are
-// pointers, and if they fit into a Fixnum, Ruby doesn't collect them, but if
-// they overflow and require allocating a Bignum, they could get collected
-// prematurely, thus removing the cache entry. This happens on 64-bit Windows,
-// on which pointers are 64 bits but longs are 32 bits. In this case, we enable
-// the secondary Hash to hold the keys and prevent them from being collected.
-
-#if RUBY_API_VERSION_CODE >= 20700 && SIZEOF_LONG >= SIZEOF_VALUE
-#define USE_SECONDARY_MAP 0
-#else
-#define USE_SECONDARY_MAP 1
-#endif
-
-#if USE_SECONDARY_MAP
-
-// Maps Numeric -> Object. The object is then used as a key into the WeakMap.
-// This is needed for Ruby <2.7 where a number cannot be a key to WeakMap.
-// The object is used only for its identity; it does not contain any data.
-VALUE secondary_map = Qnil;
-
-// Mutations to the map are under a mutex, because SeconaryMap_MaybeGC()
-// iterates over the map which cannot happen in parallel with insertions, or
-// Ruby will throw:
-//   can't add a new key into hash during iteration (RuntimeError)
-VALUE secondary_map_mutex = Qnil;
-
-// Lambda that will GC entries from the secondary map that are no longer present
-// in the primary map.
-VALUE gc_secondary_map_lambda = Qnil;
-ID length;
-
-extern VALUE weak_obj_cache;
-
-static void SecondaryMap_Init() {
-  rb_gc_register_address(&secondary_map);
-  rb_gc_register_address(&gc_secondary_map_lambda);
-  rb_gc_register_address(&secondary_map_mutex);
-  secondary_map = rb_hash_new();
-  gc_secondary_map_lambda = rb_eval_string(
-      "->(secondary, weak) {\n"
-      "  secondary.delete_if { |k, v| !weak.key?(v) }\n"
-      "}\n");
-  secondary_map_mutex = rb_mutex_new();
-  length = rb_intern("length");
-}
-
-// The secondary map is a regular Hash, and will never shrink on its own.
-// The main object cache is a WeakMap that will automatically remove entries
-// when the target object is no longer reachable, but unless we manually
-// remove the corresponding entries from the secondary map, it will grow
-// without bound.
-//
-// To avoid this unbounded growth we periodically remove entries from the
-// secondary map that are no longer present in the WeakMap. The logic of
-// how often to perform this GC is an artbirary tuning parameter that
-// represents a straightforward CPU/memory tradeoff.
-//
-// Requires: secondary_map_mutex is held.
-static void SecondaryMap_MaybeGC() {
-  PBRUBY_ASSERT(rb_mutex_locked_p(secondary_map_mutex) == Qtrue);
-  size_t weak_len = NUM2ULL(rb_funcall(weak_obj_cache, length, 0));
-  size_t secondary_len = RHASH_SIZE(secondary_map);
-  if (secondary_len < weak_len) {
-    // Logically this case should not be possible: a valid entry cannot exist in
-    // the weak table unless there is a corresponding entry in the secondary
-    // table. It should *always* be the case that secondary_len >= weak_len.
-    //
-    // However ObjectSpace::WeakMap#length (and therefore weak_len) is
-    // unreliable: it overreports its true length by including non-live objects.
-    // However these non-live objects are not yielded in iteration, so we may
-    // have previously deleted them from the secondary map in a previous
-    // invocation of SecondaryMap_MaybeGC().
-    //
-    // In this case, we can't measure any waste, so we just return.
-    return;
-  }
-  size_t waste = secondary_len - weak_len;
-  // GC if we could remove at least 2000 entries or 20% of the table size
-  // (whichever is greater).  Since the cost of the GC pass is O(N), we
-  // want to make sure that we condition this on overall table size, to
-  // avoid O(N^2) CPU costs.
-  size_t threshold = PBRUBY_MAX(secondary_len * 0.2, 2000);
-  if (waste > threshold) {
-    rb_funcall(gc_secondary_map_lambda, rb_intern("call"), 2, secondary_map,
-               weak_obj_cache);
-  }
-}
-
-// Requires: secondary_map_mutex is held by this thread iff create == true.
-static VALUE SecondaryMap_Get(VALUE key, bool create) {
-  PBRUBY_ASSERT(!create || rb_mutex_locked_p(secondary_map_mutex) == Qtrue);
-  VALUE ret = rb_hash_lookup(secondary_map, key);
-  if (ret == Qnil && create) {
-    SecondaryMap_MaybeGC();
-    ret = rb_class_new_instance(0, NULL, rb_cObject);
-    rb_hash_aset(secondary_map, key, ret);
-  }
-  return ret;
-}
-
-#endif
-
-// Requires: secondary_map_mutex is held by this thread iff create == true.
-static VALUE ObjectCache_GetKey(const void *key, bool create) {
-  VALUE key_val = (VALUE)key;
-  PBRUBY_ASSERT((key_val & 3) == 0);
-  VALUE ret = LL2NUM(key_val >> 2);
-#if USE_SECONDARY_MAP
-  ret = SecondaryMap_Get(ret, create);
-#endif
-  return ret;
-}
-
 // Public ObjectCache API.
 
 VALUE weak_obj_cache = Qnil;
 ID item_get;
-ID item_set;
+ID item_try_add;
+
+static void ObjectCache_Init(VALUE protobuf) {
+  item_get = rb_intern("get");
+  item_try_add = rb_intern("try_add");
 
-static void ObjectCache_Init() {
   rb_gc_register_address(&weak_obj_cache);
-  VALUE klass = rb_eval_string("ObjectSpace::WeakMap");
-  weak_obj_cache = rb_class_new_instance(0, NULL, klass);
-  item_get = rb_intern("[]");
-  item_set = rb_intern("[]=");
-#if USE_SECONDARY_MAP
-  SecondaryMap_Init();
+#if SIZEOF_LONG >= SIZEOF_VALUE
+  VALUE cache_class = rb_const_get(protobuf, rb_intern("ObjectCache"));
+#else
+  VALUE cache_class = rb_const_get(protobuf, rb_intern("LegacyObjectCache"));
 #endif
+
+  weak_obj_cache = rb_class_new_instance(0, NULL, cache_class);
+  rb_const_set(protobuf, rb_intern("OBJECT_CACHE"), weak_obj_cache);
+  rb_const_set(protobuf, rb_intern("SIZEOF_LONG"), INT2NUM(SIZEOF_LONG));
+  rb_const_set(protobuf, rb_intern("SIZEOF_VALUE"), INT2NUM(SIZEOF_VALUE));
 }
 
-void ObjectCache_Add(const void *key, VALUE val) {
-  PBRUBY_ASSERT(ObjectCache_Get(key) == Qnil);
-#if USE_SECONDARY_MAP
-  rb_mutex_lock(secondary_map_mutex);
-#endif
-  VALUE key_rb = ObjectCache_GetKey(key, true);
-  rb_funcall(weak_obj_cache, item_set, 2, key_rb, val);
-#if USE_SECONDARY_MAP
-  rb_mutex_unlock(secondary_map_mutex);
-#endif
-  PBRUBY_ASSERT(ObjectCache_Get(key) == val);
+static VALUE ObjectCache_GetKey(const void *key) {
+  VALUE key_val = (VALUE)key;
+  PBRUBY_ASSERT((key_val & 3) == 0);
+  // Ensure the key can be stored as a Fixnum since 1 bit is needed for
+  // FIXNUM_FLAG and 1 bit is needed for the sign bit.
+  VALUE new_key = LL2NUM(key_val >> 2);
+  PBRUBY_ASSERT(FIXNUM_P(new_key));
+  return new_key;
+}
+
+VALUE ObjectCache_TryAdd(const void *key, VALUE val) {
+  VALUE key_val = ObjectCache_GetKey(key);
+  return rb_funcall(weak_obj_cache, item_try_add, 2, key_val, val);
 }
 
 // Returns the cached object for this key, if any. Otherwise returns Qnil.
 VALUE ObjectCache_Get(const void *key) {
-  VALUE key_rb = ObjectCache_GetKey(key, false);
-  return rb_funcall(weak_obj_cache, item_get, 1, key_rb);
+  VALUE key_val = ObjectCache_GetKey(key);
+  return rb_funcall(weak_obj_cache, item_get, 1, key_val);
 }
 
 /*
@@ -457,11 +344,10 @@ VALUE Google_Protobuf_deep_copy(VALUE self, VALUE obj) {
 // This must be named "Init_protobuf_c" because the Ruby module is named
 // "protobuf_c" -- the VM looks for this symbol in our .so.
 __attribute__((visibility("default"))) void Init_protobuf_c() {
-  ObjectCache_Init();
-
   VALUE google = rb_define_module("Google");
   VALUE protobuf = rb_define_module_under(google, "Protobuf");
 
+  ObjectCache_Init(protobuf);
   Arena_register(protobuf);
   Defs_register(protobuf);
   RepeatedField_register(protobuf);

data/ext/google/protobuf_c/protobuf.h

@@ -31,8 +31,22 @@
 #ifndef __GOOGLE_PROTOBUF_RUBY_PROTOBUF_H__
 #define __GOOGLE_PROTOBUF_RUBY_PROTOBUF_H__
 
+// Ruby 3+ defines NDEBUG itself, see: https://bugs.ruby-lang.org/issues/18777
+#ifdef NDEBUG
+#include <ruby.h>
+#else
+#include <ruby.h>
+#undef NDEBUG
+#endif
+
+#include <ruby/version.h>
+
+#if RUBY_API_VERSION_CODE < 20700
+#error Protobuf requires Ruby >= 2.7
+#endif
+
+#include <assert.h>  // Must be included after the NDEBUG logic above.
 #include <ruby/encoding.h>
-#include <ruby/ruby.h>
 #include <ruby/vm.h>
 
 #include "defs.h"
@@ -76,10 +90,9 @@ void Arena_Pin(VALUE arena, VALUE obj);
 // being collected (though in Ruby <2.7 is it effectively strong, due to
 // implementation limitations).
 
-// Adds an entry to the cache. The "arena" parameter must give the arena that
-// "key" was allocated from.  In Ruby <2.7.0, it will be used to remove the key
-// from the cache when the arena is destroyed.
-void ObjectCache_Add(const void* key, VALUE val);
+// Tries to add a new entry to the cache, returning the newly installed value or
+// the pre-existing entry.
+VALUE ObjectCache_TryAdd(const void* key, VALUE val);
 
 // Returns the cached object for this key, if any. Otherwise returns Qnil.
 VALUE ObjectCache_Get(const void* key);
@@ -110,7 +123,9 @@ extern VALUE cTypeError;
   do {                      \
   } while (false && (expr))
 #else
-#define PBRUBY_ASSERT(expr) assert(expr)
+#define PBRUBY_ASSERT(expr) \
+  if (!(expr))              \
+  rb_bug("Assertion failed at %s:%d, expr: %s", __FILE__, __LINE__, #expr)
 #endif
 
 #define PBRUBY_MAX(x, y) (((x) > (y)) ? (x) : (y))

data/ext/google/protobuf_c/repeated_field.c

@@ -87,7 +87,6 @@ VALUE RepeatedField_GetRubyWrapper(upb_Array* array, TypeInfo type_info,
   if (val == Qnil) {
     val = RepeatedField_alloc(cRepeatedField);
     RepeatedField* self;
-    ObjectCache_Add(array, val);
     TypedData_Get_Struct(val, RepeatedField, &RepeatedField_type, self);
     self->array = array;
     self->arena = arena;
@@ -95,11 +94,14 @@ VALUE RepeatedField_GetRubyWrapper(upb_Array* array, TypeInfo type_info,
     if (self->type_info.type == kUpb_CType_Message) {
       self->type_class = Descriptor_DefToClass(type_info.def.msgdef);
     }
+    val = ObjectCache_TryAdd(array, val);
   }
 
   PBRUBY_ASSERT(ruby_to_RepeatedField(val)->type_info.type == type_info.type);
   PBRUBY_ASSERT(ruby_to_RepeatedField(val)->type_info.def.msgdef ==
                 type_info.def.msgdef);
+  PBRUBY_ASSERT(ruby_to_RepeatedField(val)->array == array);
+
   return val;
 }
 
@@ -613,7 +615,8 @@ VALUE RepeatedField_init(int argc, VALUE* argv, VALUE _self) {
 
   self->type_info = TypeInfo_FromClass(argc, argv, 0, &self->type_class, &ary);
   self->array = upb_Array_New(arena, self->type_info.type);
-  ObjectCache_Add(self->array, _self);
+  VALUE stored_val = ObjectCache_TryAdd(self->array, _self);
+  PBRUBY_ASSERT(stored_val == _self);
 
   if (ary != Qnil) {
     if (!RB_TYPE_P(ary, T_ARRAY)) {

data/ext/google/protobuf_c/repeated_field.h

@@ -31,8 +31,6 @@
 #ifndef RUBY_PROTOBUF_REPEATED_FIELD_H_
 #define RUBY_PROTOBUF_REPEATED_FIELD_H_
 
-#include <ruby/ruby.h>
-
 #include "protobuf.h"
 #include "ruby-upb.h"