[channelz] Shard indices (grpc#39306)

Previously channelz kept an index of all objects in a global std::map protected by a single mutex.

Instead: split the index in two. The first is a list of all base nodes, queryable on demand, sharded by pointer. These base nodes are available, but un-numbered. The id allocation then occurs whenever channelz is actually queried. At that point unnumbered nodes can be iterated and numbered sequentially, keeping that essential invariant for channelz pagination.

In doing so, most nodes never need to hit a singleton global resource at all: the most they do is take a sharded lock and do a list insert/removal pair.

(it turns out this code is a major source of contention in gRPC)

Closes grpc#39306

COPYBARA_INTEGRATE_REVIEW=grpc#39306 from ctiller:counting-on-it a7a04c3
PiperOrigin-RevId: 750342301

Author: ctiller

BUILD

@@ -1251,6 +1251,7 @@ grpc_cc_library(
     ],
     external_deps = [
         "absl/base:core_headers",
+        "absl/container:btree",
         "absl/log",
         "absl/log:check",
         "absl/status:statusor",
@@ -1273,6 +1274,7 @@ grpc_cc_library(
         "//src/core:per_cpu",
         "//src/core:ref_counted",
         "//src/core:resolved_address",
+        "//src/core:shared_bit_gen",
         "//src/core:slice",
         "//src/core:sync",
         "//src/core:time",

CMakeLists.txt

@@ -168,7 +168,9 @@ BaseNode::BaseNode(EntityType type, std::string name)
   ChannelzRegistry::Register(this);
 }
 
-BaseNode::~BaseNode() { ChannelzRegistry::Unregister(uuid_); }
+BaseNode::~BaseNode() { ChannelzRegistry::Unregister(this); }
+
+intptr_t BaseNode::UuidSlow() { return ChannelzRegistry::NumberNode(this); }
 
 std::string BaseNode::RenderJsonString() {
   Json json = RenderJson();

bazel/experiments.bzl

@@ -141,7 +141,11 @@ class BaseNode : public RefCounted<BaseNode> {
   std::string RenderJsonString();
 
   EntityType type() const { return type_; }
-  intptr_t uuid() const { return uuid_; }
+  intptr_t uuid() {
+    const intptr_t id = uuid_.load(std::memory_order_relaxed);
+    if (id > 0) return id;
+    return UuidSlow();
+  }
   const std::string& name() const { return name_; }
 
   void RunZTrace(absl::string_view name, Timestamp deadline,
@@ -155,16 +159,20 @@ class BaseNode : public RefCounted<BaseNode> {
   void PopulateJsonFromDataSources(Json::Object& json);
 
  private:
+  intptr_t UuidSlow();
+
   // to allow the ChannelzRegistry to set uuid_ under its lock.
   friend class ChannelzRegistry;
   // allow data source to register/unregister itself
   friend class DataSource;
   const EntityType type_;
-  intptr_t uuid_;
+  std::atomic<intptr_t> uuid_;
   std::string name_;
   Mutex data_sources_mu_;
   absl::InlinedVector<DataSource*, 3> data_sources_
       ABSL_GUARDED_BY(data_sources_mu_);
+  BaseNode* prev_;
+  BaseNode* next_;
 };
 
 class ZTrace {