DEBUG-2334 remaining dynamic instrumentation code

lib/datadog.rb

@@ -7,4 +7,7 @@
 # Load other products (must follow tracing)
 require_relative 'datadog/profiling'
 require_relative 'datadog/appsec'
+# Line probes will not work on Ruby < 2.6 because of lack of :script_compiled
+# trace point. Only load DI on supported Ruby versions.
+require_relative 'datadog/di' if RUBY_VERSION >= '2.6'
 require_relative 'datadog/kit'

lib/datadog/core/configuration.rb

@@ -258,8 +258,16 @@ def build_components(settings)
       def replace_components!(settings, old)
         components = Components.new(settings)
 
+        # Carry over state from existing components to the new ones.
+        # Currently, if we already started the remote component (which
+        # happens after a request goes through installed Rack middleware),
+        # we will start the new remote component as well.
+        old_state = {
+          remote_started: old.remote&.started?,
+        }
+
         old.shutdown!(components)
-        components.startup!(settings)
+        components.startup!(settings, old_state: old_state)
         components
       end
 

lib/datadog/core/configuration/components.rb

@@ -13,6 +13,7 @@
 require_relative '../../tracing/component'
 require_relative '../../profiling/component'
 require_relative '../../appsec/component'
+require_relative '../../di/component'
 require_relative '../crashtracking/component'
 
 module Datadog
@@ -83,6 +84,7 @@ def build_crashtracker(settings, agent_settings, logger:)
           :telemetry,
           :tracer,
           :crashtracker,
+          :dynamic_instrumentation,
           :appsec
 
         def initialize(settings)
@@ -110,12 +112,13 @@ def initialize(settings)
           @runtime_metrics = self.class.build_runtime_metrics_worker(settings)
           @health_metrics = self.class.build_health_metrics(settings)
           @appsec = Datadog::AppSec::Component.build_appsec_component(settings, telemetry: telemetry)
+          @dynamic_instrumentation = Datadog::DI::Component.build(settings, agent_settings, telemetry: telemetry)
 
           self.class.configure_tracing(settings)
         end
 
         # Starts up components
-        def startup!(settings)
+        def startup!(settings, old_state: nil)
           if settings.profiling.enabled
             if profiler
               profiler.start
@@ -126,6 +129,16 @@ def startup!(settings)
             end
           end
 
+          if settings.remote.enabled && old_state&.[](:remote_started)
+            # The library was reconfigured and previously it already started
+            # the remote component (i.e., it received at least one request
+            # through the installed Rack middleware which started the remote).
+            # If the new configuration also has remote enabled, start the
+            # new remote right away.
+            # remote should always be not nil here but steep doesn't know this.
+            remote&.start
+          end
+
           Core::Diagnostics::EnvironmentLogger.collect_and_log!(@environment_logger_extra)
         end
 
@@ -136,6 +149,9 @@ def shutdown!(replacement = nil)
           # Shutdown remote configuration
           remote.shutdown! if remote
 
+          # Shutdown DI after remote, since remote config triggers DI operations.
+          dynamic_instrumentation&.shutdown!
+
           # Decommission AppSec
           appsec.shutdown! if appsec
 

lib/datadog/core/remote/client.rb

@@ -24,93 +24,99 @@ def initialize(transport, capabilities, repository: Configuration::Repository.ne
           @dispatcher = Dispatcher.new(@capabilities.receivers)
         end
 
-        # rubocop:disable Metrics/AbcSize,Metrics/PerceivedComplexity,Metrics/MethodLength,Metrics/CyclomaticComplexity
         def sync
           # TODO: Skip sync if no capabilities are registered
           response = transport.send_config(payload)
 
           if response.ok?
-            # when response is completely empty, do nothing as in: leave as is
-            if response.empty?
-              Datadog.logger.debug { 'remote: empty response => NOOP' }
+            process_response(response)
+          elsif response.internal_error?
+            raise TransportError, response.to_s
+          end
+        end
 
-              return
-            end
+        private
 
-            begin
-              paths = response.client_configs.map do |path|
-                Configuration::Path.parse(path)
-              end
+        def process_response(response)
+          # when response is completely empty, do nothing as in: leave as is
+          if response.empty?
+            Datadog.logger.debug { 'remote: empty response => NOOP' }
 
-              targets = Configuration::TargetMap.parse(response.targets)
+            return
+          end
 
-              contents = Configuration::ContentList.parse(response.target_files)
-            rescue Remote::Configuration::Path::ParseError => e
-              raise SyncError, e.message
+          begin
+            paths = response.client_configs.map do |path|
+              Configuration::Path.parse(path)
             end
 
-            # To make sure steep does not complain
-            return unless paths && targets && contents
+            targets = Configuration::TargetMap.parse(response.targets)
+
+            contents = Configuration::ContentList.parse(response.target_files)
+          rescue Remote::Configuration::Path::ParseError => e
+            raise SyncError, e.message
+          end
 
-            # TODO: sometimes it can strangely be so that paths.empty?
-            # TODO: sometimes it can strangely be so that targets.empty?
+          # To make sure steep does not complain
+          return unless paths && targets && contents
 
-            changes = repository.transaction do |current, transaction|
-              # paths to be removed: previously applied paths minus ingress paths
-              (current.paths - paths).each { |p| transaction.delete(p) }
+          # TODO: sometimes it can strangely be so that paths.empty?
+          # TODO: sometimes it can strangely be so that targets.empty?
 
-              # go through each ingress path
-              paths.each do |path|
-                # match target with path
-                target = targets[path]
+          apply_config(paths, targets, contents)
+        end
 
-                # abort entirely if matching target not found
-                raise SyncError, "no target for path '#{path}'" if target.nil?
+        def apply_config(paths, targets, contents)
+          changes = repository.transaction do |current, transaction|
+            # paths to be removed: previously applied paths minus ingress paths
+            (current.paths - paths).each { |p| transaction.delete(p) }
 
-                # new paths are not in previously applied paths
-                new = !current.paths.include?(path)
+            # go through each ingress path
+            paths.each do |path|
+              # match target with path
+              target = targets[path]
 
-                # updated paths are in previously applied paths
-                # but the content hash changed
-                changed = current.paths.include?(path) && !current.contents.find_content(path, target)
+              # abort entirely if matching target not found
+              raise SyncError, "no target for path '#{path}'" if target.nil?
 
-                # skip if unchanged
-                same = !new && !changed
+              # new paths are not in previously applied paths
+              new = !current.paths.include?(path)
 
-                next if same
+              # updated paths are in previously applied paths
+              # but the content hash changed
+              changed = current.paths.include?(path) && !current.contents.find_content(path, target)
 
-                # match content with path and target
-                content = contents.find_content(path, target)
+              # skip if unchanged
+              same = !new && !changed
 
-                # abort entirely if matching content not found
-                raise SyncError, "no valid content for target at path '#{path}'" if content.nil?
+              next if same
 
-                # to be added or updated << config
-                # TODO: metadata (hash, version, etc...)
-                transaction.insert(path, target, content) if new
-                transaction.update(path, target, content) if changed
-              end
+              # match content with path and target
+              content = contents.find_content(path, target)
 
-              # save backend opaque backend state
-              transaction.set(opaque_backend_state: targets.opaque_backend_state)
-              transaction.set(targets_version: targets.version)
+              # abort entirely if matching content not found
+              raise SyncError, "no valid content for target at path '#{path}'" if content.nil?
 
-              # upon transaction end, new list of applied config + metadata (add, change, remove) will be saved
-              # TODO: also remove stale config (matching removed) from cache (client configs is exhaustive list of paths)
+              # to be added or updated << config
+              # TODO: metadata (hash, version, etc...)
+              transaction.insert(path, target, content) if new
+              transaction.update(path, target, content) if changed
             end
 
-            if changes.empty?
-              Datadog.logger.debug { 'remote: no changes' }
-            else
-              dispatcher.dispatch(changes, repository)
-            end
-          elsif response.internal_error?
-            raise TransportError, response.to_s
+            # save backend opaque backend state
+            transaction.set(opaque_backend_state: targets.opaque_backend_state)
+            transaction.set(targets_version: targets.version)
+
+            # upon transaction end, new list of applied config + metadata (add, change, remove) will be saved
+            # TODO: also remove stale config (matching removed) from cache (client configs is exhaustive list of paths)
           end
-        end
-        # rubocop:enable Metrics/AbcSize,Metrics/PerceivedComplexity,Metrics/MethodLength,Metrics/CyclomaticComplexity
 
-        private
+          if changes.empty?
+            Datadog.logger.debug { 'remote: no changes' }
+          else
+            dispatcher.dispatch(changes, repository)
+          end
+        end
 
         def payload # rubocop:disable Metrics/MethodLength
           state = repository.state

lib/datadog/core/remote/client/capabilities.rb

@@ -32,6 +32,12 @@ def register(settings)
               register_receivers(Datadog::AppSec::Remote.receivers(@telemetry))
             end
 
+            if settings.respond_to?(:dynamic_instrumentation) && settings.dynamic_instrumentation.enabled
+              register_capabilities(Datadog::DI::Remote.capabilities)
+              register_products(Datadog::DI::Remote.products)
+              register_receivers(Datadog::DI::Remote.receivers(@telemetry))
+            end
+
             register_capabilities(Datadog::Tracing::Remote.capabilities)
             register_products(Datadog::Tracing::Remote.products)
             register_receivers(Datadog::Tracing::Remote.receivers(@telemetry))

lib/datadog/di.rb

@@ -12,6 +12,7 @@
 require_relative 'di/probe_notification_builder'
 require_relative 'di/probe_notifier_worker'
 require_relative 'di/redactor'
+require_relative 'di/remote'
 require_relative 'di/serializer'
 require_relative 'di/transport'
 require_relative 'di/utils'
@@ -25,6 +26,9 @@
   # and AR should be loaded before any application code is loaded, being
   # part of Rails, therefore for now we should be OK to just require the
   # AR integration from here.
+  #
+  # TODO this require might need to be delayed via Rails post-initialization
+  # logic?
   require_relative 'di/contrib/active_record'
 end
 
@@ -42,6 +46,8 @@ def enabled?
     # Expose DI to global shared objects
     Extensions.activate!
 
+    LOCK = Mutex.new
+
     class << self
       attr_reader :code_tracker
 
@@ -58,6 +64,32 @@ def activate_tracking!
         (@code_tracker ||= CodeTracker.new).start
       end
 
+      # Activates code tracking if possible.
+      #
+      # This method does nothing if invoked in an environment that does not
+      # implement required trace points for code tracking (MRI Ruby < 2.6,
+      # JRuby) and rescues any exceptions that may be raised by downstream
+      # DI code.
+      def activate_tracking
+        # :script_compiled trace point was added in Ruby 2.6.
+        return unless RUBY_VERSION >= '2.6'
+
+        begin
+          # Activate code tracking by default because line trace points will not work
+          # without it.
+          Datadog::DI.activate_tracking!
+        rescue => exc
+          if defined?(Datadog.logger)
+            Datadog.logger.warn("Failed to activate code tracking for DI: #{exc.class}: #{exc}")
+          else
+            # We do not have Datadog logger potentially because DI code tracker is
+            # being loaded early in application boot process and the rest of datadog
+            # wasn't loaded yet. Output to standard error.
+            warn("Failed to activate code tracking for DI: #{exc.class}: #{exc}")
+          end
+        end
+      end
+
       # Deactivates code tracking. In normal usage of DI this method should
       # never be called, however it is used by DI's test suite to reset
       # state for individual tests.
@@ -76,30 +108,62 @@ def code_tracking_active?
         code_tracker&.active? || false
       end
 
+      # This method is called from DI Remote handler to issue DI operations
+      # to the probe manager (add or remove probes).
+      #
+      # When DI Remote is executing, Datadog.components should be initialized
+      # and we should be able to reference it to get to the DI component.
+      #
+      # Given that we need the current_component anyway for code tracker,
+      # perhaps we should delete the +component+ method and just use
+      # +current_component+ in all cases.
       def component
-        # TODO uncomment when remote is merged
-        #Datadog.send(:components).dynamic_instrumentation
+        Datadog.send(:components).dynamic_instrumentation
+      end
+
+      # DI code tracker is instantiated globally before the regular set of
+      # components is created, but the code tracker needs to call out to the
+      # "current" DI component to perform instrumentation when application
+      # code is loaded. Because this call may happen prior to Datadog
+      # components having been initialized, we maintain the "current component"
+      # which contains a reference to the most recently instantiated
+      # DI::Component. This way, if a DI component hasn't been instantiated,
+      # we do not try to reference Datadog.components.
+      def current_component
+        LOCK.synchronize do
+          @current_components&.last
+        end
+      end
+
+      # To avoid potential races with DI::Component being added and removed,
+      # we maintain a list of the components. Normally the list should contain
+      # either zero or one component depending on whether DI is enabled in
+      # Datadog configuration. However, if a new instance of DI::Component
+      # is created while the previous instance is still running, we are
+      # guaranteed to not end up with no component when one is running.
+      def add_current_component(component)
+        LOCK.synchronize do
+          @current_components ||= []
+          @current_components << component
+        end
+      end
+
+      def remove_current_component(component)
+        LOCK.synchronize do
+          @current_components&.delete(component)
+        end
       end
     end
   end
 end
 
-=begin not yet enabled
-# :script_compiled trace point was added in Ruby 2.6.
-if RUBY_VERSION >= '2.6'
-  begin
-    # Activate code tracking by default because line trace points will not work
-    # without it.
-    Datadog::DI.activate_tracking!
-  rescue => exc
-    if defined?(Datadog.logger)
-      Datadog.logger.warn("Failed to activate code tracking for DI: #{exc.class}: #{exc}")
-    else
-      # We do not have Datadog logger potentially because DI code tracker is
-      # being loaded early in application boot process and the rest of datadog
-      # wasn't loaded yet. Output to standard error.
-      warn("Failed to activate code tracking for DI: #{exc.class}: #{exc}")
-    end
-  end
+if %w(1 true).include?(ENV['DD_DYNAMIC_INSTRUMENTATION_ENABLED']) # steep:ignore
+  # For initial release of Dynamic Instrumentation, activate code tracking
+  # only if DI is explicitly requested in the environment.
+  # Code tracking is required for line probes to work; see the comments
+  # above for the implementation of the method.
+  #
+  # If DI is enabled programmatically, the application can (and must,
+  # for line probes to work) activate tracking in an initializer.
+  Datadog::DI.activate_tracking
 end
-=end