ddtrace 0.54.2 → 1.10.1

data/lib/datadog/profiling/exporter.rb

@@ -0,0 +1,83 @@
+require_relative 'ext'
+require_relative 'tag_builder'
+
+module Datadog
+  module Profiling
+    # Exports profiling data gathered by the multiple recorders in a `Flush`.
+    #
+    # @ivoanjo: Note that the recorder that gathers pprof data is special, since we use its start/finish/empty? to
+    # decide if there's data to flush, as well as the timestamp for that data.
+    # I could've made the whole design more generic, but I'm unsure if we'll ever have more than a handful of
+    # recorders, so I've decided to make it specific until we actually need to support more recorders.
+    #
+    class Exporter
+      # Profiles with duration less than this will not be reported
+      PROFILE_DURATION_THRESHOLD_SECONDS = 1
+
+      private
+
+      attr_reader \
+        :pprof_recorder,
+        :code_provenance_collector, # The code provenance collector acts both as collector and as a recorder
+        :minimum_duration_seconds,
+        :time_provider,
+        :last_flush_finish_at,
+        :created_at
+
+      public
+
+      def initialize(
+        pprof_recorder:,
+        code_provenance_collector:,
+        minimum_duration_seconds: PROFILE_DURATION_THRESHOLD_SECONDS,
+        time_provider: Time
+      )
+        @pprof_recorder = pprof_recorder
+        @code_provenance_collector = code_provenance_collector
+        @minimum_duration_seconds = minimum_duration_seconds
+        @time_provider = time_provider
+        @last_flush_finish_at = nil
+        @created_at = time_provider.now.utc
+      end
+
+      def flush
+        start, finish, uncompressed_pprof = pprof_recorder.serialize
+        @last_flush_finish_at = finish
+
+        return if uncompressed_pprof.nil? # We don't want to report empty profiles
+
+        if duration_below_threshold?(start, finish)
+          Datadog.logger.debug('Skipped exporting profiling events as profile duration is below minimum')
+          return
+        end
+
+        uncompressed_code_provenance = code_provenance_collector.refresh.generate_json if code_provenance_collector
+
+        Flush.new(
+          start: start,
+          finish: finish,
+          pprof_file_name: Datadog::Profiling::Ext::Transport::HTTP::PPROF_DEFAULT_FILENAME,
+          pprof_data: uncompressed_pprof.to_s,
+          code_provenance_file_name: Datadog::Profiling::Ext::Transport::HTTP::CODE_PROVENANCE_FILENAME,
+          code_provenance_data: uncompressed_code_provenance,
+          tags_as_array: Datadog::Profiling::TagBuilder.call(settings: Datadog.configuration).to_a,
+        )
+      end
+
+      def can_flush?
+        !duration_below_threshold?(last_flush_finish_at || created_at, time_provider.now.utc)
+      end
+
+      def reset_after_fork
+        @last_flush_finish_at = time_provider.now.utc
+        nil
+      end
+
+      private
+
+      def duration_below_threshold?(start, finish)
+        (finish - start) < minimum_duration_seconds
+      end
+    end
+  end
+end

data/lib/datadog/profiling/ext/forking.rb

@@ -0,0 +1,96 @@
+module Datadog
+  module Profiling
+    module Ext
+      # Monkey patches `Kernel#fork`, adding a `Kernel#at_fork` callback mechanism which is used to restore
+      # profiling abilities after the VM forks.
+      #
+      # Known limitations: Does not handle `BasicObject`s that include `Kernel` directly; e.g.
+      # `Class.new(BasicObject) { include(::Kernel); def call; fork { }; end }.new.call`.
+      #
+      # This will be fixed once we moved to hooking into `Process._fork`
+      module Forking
+        def self.supported?
+          Process.respond_to?(:fork)
+        end
+
+        def self.apply!
+          return false unless supported?
+
+          [
+            ::Process.singleton_class, # Process.fork
+            ::Kernel.singleton_class,  # Kernel.fork
+            ::Object,                  # fork without explicit receiver (it's defined as a method in ::Kernel)
+            # Note: Modifying Object as we do here is irreversible. During tests, this
+            # change will stick around even if we otherwise stub `Process` and `Kernel`
+          ].each { |target| target.prepend(Kernel) }
+
+          ::Process.singleton_class.prepend(ProcessDaemonPatch)
+        end
+
+        # Extensions for kernel
+        #
+        # TODO: Consider hooking into `Process._fork` on Ruby 3.1+ instead, see
+        #       https://github.com/ruby/ruby/pull/5017 and https://bugs.ruby-lang.org/issues/17795
+        module Kernel
+          def fork
+            # If a block is provided, it must be wrapped to trigger callbacks.
+            child_block = if block_given?
+                            proc do
+                              # Trigger :child callback
+                              ddtrace_at_fork_blocks[:child].each(&:call) if ddtrace_at_fork_blocks.key?(:child)
+
+                              # Invoke original block
+                              yield
+                            end
+                          end
+
+            # Start fork
+            # If a block is provided, use the wrapped version.
+            result = child_block.nil? ? super : super(&child_block)
+
+            # Trigger correct callbacks depending on whether we're in the parent or child.
+            # If we're in the fork, result = nil: trigger child callbacks.
+            # If we're in the parent, result = fork PID: trigger parent callbacks.
+            ddtrace_at_fork_blocks[:child].each(&:call) if result.nil? && ddtrace_at_fork_blocks.key?(:child)
+
+            # Return PID from #fork
+            result
+          end
+
+          def at_fork(stage, &block)
+            raise ArgumentError, 'Bad \'stage\' for ::at_fork' unless stage == :child
+
+            ddtrace_at_fork_blocks[stage] = [] unless ddtrace_at_fork_blocks.key?(stage)
+            ddtrace_at_fork_blocks[stage] << block
+          end
+
+          module_function
+
+          def ddtrace_at_fork_blocks
+            # Blocks should be shared across all users of this module,
+            # e.g. Process#fork, Kernel#fork, etc. should all invoke the same callbacks.
+            # rubocop:disable Style/ClassVars
+            @@ddtrace_at_fork_blocks ||= {}
+            # rubocop:enable Style/ClassVars
+          end
+        end
+
+        # A call to Process.daemon ( https://rubyapi.org/3.1/o/process#method-c-daemon ) forks the current process and
+        # keeps executing code in the child process, killing off the parent, thus effectively replacing it.
+        #
+        # This monkey patch makes the `Kernel#at_fork` mechanism defined above also work in this situation.
+        module ProcessDaemonPatch
+          def daemon(*args)
+            ddtrace_at_fork_blocks = Datadog::Profiling::Ext::Forking::Kernel.ddtrace_at_fork_blocks
+
+            result = super
+
+            ddtrace_at_fork_blocks[:child].each(&:call) if ddtrace_at_fork_blocks.key?(:child)
+
+            result
+          end
+        end
+      end
+    end
+  end
+end

data/lib/datadog/profiling/ext.rb

@@ -0,0 +1,43 @@
+module Datadog
+  module Profiling
+    module Ext
+      ENV_ENABLED = 'DD_PROFILING_ENABLED'.freeze
+      ENV_UPLOAD_TIMEOUT = 'DD_PROFILING_UPLOAD_TIMEOUT'.freeze
+      ENV_MAX_FRAMES = 'DD_PROFILING_MAX_FRAMES'.freeze
+      ENV_AGENTLESS = 'DD_PROFILING_AGENTLESS'.freeze
+      ENV_ENDPOINT_COLLECTION_ENABLED = 'DD_PROFILING_ENDPOINT_COLLECTION_ENABLED'.freeze
+
+      # TODO: Consider removing this once the Ruby-based pprof encoding is removed and replaced by libdatadog
+      module Pprof
+        LABEL_KEY_LOCAL_ROOT_SPAN_ID = 'local root span id'.freeze
+        LABEL_KEY_SPAN_ID = 'span id'.freeze
+        LABEL_KEY_THREAD_ID = 'thread id'.freeze
+        LABEL_KEY_TRACE_ENDPOINT = 'trace endpoint'.freeze
+        SAMPLE_VALUE_NO_VALUE = 0
+        VALUE_TYPE_CPU = 'cpu-time'.freeze
+        VALUE_TYPE_WALL = 'wall-time'.freeze
+        VALUE_UNIT_NANOSECONDS = 'nanoseconds'.freeze
+      end
+
+      module Transport
+        module HTTP
+          FORM_FIELD_TAG_ENV = 'env'.freeze
+          FORM_FIELD_TAG_HOST = 'host'.freeze
+          FORM_FIELD_TAG_LANGUAGE = 'language'.freeze
+          FORM_FIELD_TAG_PID = 'process_id'.freeze
+          FORM_FIELD_TAG_PROFILER_VERSION = 'profiler_version'.freeze
+          FORM_FIELD_TAG_RUNTIME = 'runtime'.freeze
+          FORM_FIELD_TAG_RUNTIME_ENGINE = 'runtime_engine'.freeze
+          FORM_FIELD_TAG_RUNTIME_ID = 'runtime-id'.freeze
+          FORM_FIELD_TAG_RUNTIME_PLATFORM = 'runtime_platform'.freeze
+          FORM_FIELD_TAG_RUNTIME_VERSION = 'runtime_version'.freeze
+          FORM_FIELD_TAG_SERVICE = 'service'.freeze
+          FORM_FIELD_TAG_VERSION = 'version'.freeze
+
+          PPROF_DEFAULT_FILENAME = 'rubyprofile.pprof'.freeze
+          CODE_PROVENANCE_FILENAME = 'code-provenance.json'.freeze
+        end
+      end
+    end
+  end
+end

data/lib/datadog/profiling/flush.rb

@@ -0,0 +1,36 @@
+module Datadog
+  module Profiling
+    # Represents a collection of events of a specific type being flushed.
+    EventGroup = Struct.new(:event_class, :events)
+
+    # Entity class used to represent metadata for a given profile
+    class Flush
+      attr_reader \
+        :start,
+        :finish,
+        :pprof_file_name,
+        :pprof_data, # gzipped pprof bytes
+        :code_provenance_file_name,
+        :code_provenance_data, # gzipped json bytes
+        :tags_as_array
+
+      def initialize(
+        start:,
+        finish:,
+        pprof_file_name:,
+        pprof_data:,
+        code_provenance_file_name:,
+        code_provenance_data:,
+        tags_as_array:
+      )
+        @start = start
+        @finish = finish
+        @pprof_file_name = pprof_file_name
+        @pprof_data = pprof_data
+        @code_provenance_file_name = code_provenance_file_name
+        @code_provenance_data = code_provenance_data
+        @tags_as_array = tags_as_array
+      end
+    end
+  end
+end

data/lib/datadog/profiling/http_transport.rb

@@ -0,0 +1,130 @@
+module Datadog
+  module Profiling
+    # Used to report profiling data to Datadog.
+    # Methods prefixed with _native_ are implemented in `http_transport.c`
+    class HttpTransport
+      def initialize(agent_settings:, site:, api_key:, upload_timeout_seconds:)
+        @upload_timeout_milliseconds = (upload_timeout_seconds * 1_000).to_i
+
+        validate_agent_settings(agent_settings)
+
+        @exporter_configuration =
+          if agentless?(site, api_key)
+            [:agentless, site, api_key]
+          else
+            [:agent, base_url_from(agent_settings)]
+          end
+
+        status, result = validate_exporter(@exporter_configuration)
+
+        raise(ArgumentError, "Failed to initialize transport: #{result}") if status == :error
+      end
+
+      def export(flush)
+        status, result = do_export(
+          exporter_configuration: @exporter_configuration,
+          upload_timeout_milliseconds: @upload_timeout_milliseconds,
+
+          # why "timespec"?
+          # libdatadog represents time using POSIX's struct timespec, see
+          # https://www.gnu.org/software/libc/manual/html_node/Time-Types.html
+          # aka it represents the seconds part separate from the nanoseconds part
+          start_timespec_seconds: flush.start.tv_sec,
+          start_timespec_nanoseconds: flush.start.tv_nsec,
+          finish_timespec_seconds: flush.finish.tv_sec,
+          finish_timespec_nanoseconds: flush.finish.tv_nsec,
+
+          pprof_file_name: flush.pprof_file_name,
+          pprof_data: flush.pprof_data,
+          code_provenance_file_name: flush.code_provenance_file_name,
+          code_provenance_data: flush.code_provenance_data,
+
+          tags_as_array: flush.tags_as_array,
+        )
+
+        if status == :ok
+          if (200..299).cover?(result)
+            Datadog.logger.debug('Successfully reported profiling data')
+            true
+          else
+            Datadog.logger.error("Failed to report profiling data: server returned unexpected HTTP #{result} status code")
+            false
+          end
+        else
+          Datadog.logger.error("Failed to report profiling data: #{result}")
+          false
+        end
+      end
+
+      # Used to log soft failures in `ddog_Vec_tag_push` (e.g. we still report the profile in these cases)
+      # Called from native code
+      def self.log_failure_to_process_tag(failure_details)
+        Datadog.logger.warn("Failed to add tag to profiling request: #{failure_details}")
+      end
+
+      private
+
+      def base_url_from(agent_settings)
+        case agent_settings.adapter
+        when Datadog::Transport::Ext::HTTP::ADAPTER
+          "#{agent_settings.ssl ? 'https' : 'http'}://#{agent_settings.hostname}:#{agent_settings.port}/"
+        when Datadog::Transport::Ext::UnixSocket::ADAPTER
+          "unix://#{agent_settings.uds_path}"
+        else
+          raise ArgumentError, "Unexpected adapter: #{agent_settings.adapter}"
+        end
+      end
+
+      def validate_agent_settings(agent_settings)
+        supported_adapters = [Datadog::Transport::Ext::HTTP::ADAPTER, Datadog::Transport::Ext::UnixSocket::ADAPTER]
+        unless supported_adapters.include?(agent_settings.adapter)
+          raise ArgumentError,
+            "Unsupported transport configuration for profiling: Adapter #{agent_settings.adapter} " \
+                        ' is not supported'
+        end
+
+        if agent_settings.deprecated_for_removal_transport_configuration_proc
+          Datadog.logger.warn(
+            'Ignoring custom c.tracing.transport_options setting as the profiler does not support it.'
+          )
+        end
+      end
+
+      def agentless?(site, api_key)
+        site && api_key && Core::Environment::VariableHelpers.env_to_bool(Profiling::Ext::ENV_AGENTLESS, false)
+      end
+
+      def validate_exporter(exporter_configuration)
+        self.class._native_validate_exporter(exporter_configuration)
+      end
+
+      def do_export(
+        exporter_configuration:,
+        upload_timeout_milliseconds:,
+        start_timespec_seconds:,
+        start_timespec_nanoseconds:,
+        finish_timespec_seconds:,
+        finish_timespec_nanoseconds:,
+        pprof_file_name:,
+        pprof_data:,
+        code_provenance_file_name:,
+        code_provenance_data:,
+        tags_as_array:
+      )
+        self.class._native_do_export(
+          exporter_configuration,
+          upload_timeout_milliseconds,
+          start_timespec_seconds,
+          start_timespec_nanoseconds,
+          finish_timespec_seconds,
+          finish_timespec_nanoseconds,
+          pprof_file_name,
+          pprof_data,
+          code_provenance_file_name,
+          code_provenance_data,
+          tags_as_array,
+        )
+      end
+    end
+  end
+end

data/lib/datadog/profiling/load_native_extension.rb

@@ -0,0 +1,20 @@
+# This file is used to load the profiling native extension. It works in two steps:
+#
+# 1. Load the ddtrace_profiling_loader extension. This extension will be used to load the actual extension, but in
+#    a special way that avoids exposing native-level code symbols. See `ddtrace_profiling_loader.c` for more details.
+#
+# 2. Use the Datadog::Profiling::Loader exposed by the ddtrace_profiling_loader extension to load the actual
+#    profiling native extension.
+#
+# All code on this file is on-purpose at the top-level; this makes it so this file is executed only once,
+# the first time it gets required, to avoid any issues with the native extension being initialized more than once.
+
+require "ddtrace_profiling_loader.#{RUBY_VERSION}_#{RUBY_PLATFORM}"
+
+extension_name = "ddtrace_profiling_native_extension.#{RUBY_VERSION}_#{RUBY_PLATFORM}"
+full_file_path = "#{__dir__}/../../#{extension_name}.#{RbConfig::CONFIG['DLEXT']}"
+init_function_name = "Init_#{extension_name.split('.').first}"
+
+status, result = Datadog::Profiling::Loader._native_load(full_file_path, init_function_name)
+
+raise "Failure to load #{extension_name} due to #{result}" if status == :error

data/lib/datadog/profiling/native_extension.rb

@@ -0,0 +1,39 @@
+module Datadog
+  module Profiling
+    # This module contains classes and methods which are implemented using native code in the
+    # ext/ddtrace_profiling_native_extension folder, as well as some Ruby-level utilities that don't make sense to
+    # write using C
+    module NativeExtension
+      private_class_method def self.working?
+        native_working?
+      end
+
+      unless singleton_class.private_method_defined?(:native_working?)
+        private_class_method def self.native_working?
+          false
+        end
+      end
+
+      unless singleton_class.method_defined?(:clock_id_for)
+        def self.clock_id_for(_)
+          nil
+        end
+      end
+
+      def self.cpu_time_ns_for(thread)
+        clock_id =
+          begin
+            clock_id_for(thread)
+          rescue Errno::ESRCH
+            nil
+          end
+
+        begin
+          ::Process.clock_gettime(clock_id, :nanosecond) if clock_id
+        rescue Errno::EINVAL
+          nil
+        end
+      end
+    end
+  end
+end

data/lib/datadog/profiling/old_recorder.rb

@@ -0,0 +1,107 @@
+require_relative 'buffer'
+require_relative 'encoding/profile'
+
+module Datadog
+  module Profiling
+    # Stores profiling events gathered by the `Stack` collector
+    class OldRecorder
+      attr_reader :max_size
+
+      def initialize(
+        event_classes,
+        max_size,
+        last_flush_time: Time.now.utc
+      )
+        @buffers = {}
+        @last_flush_time = last_flush_time
+        @max_size = max_size
+
+        # Add a buffer for each class
+        event_classes.each do |event_class|
+          @buffers[event_class] = Profiling::Buffer.new(max_size)
+        end
+
+        # Event classes can only be added ahead of time
+        @buffers.freeze
+      end
+
+      def [](event_class)
+        @buffers[event_class]
+      end
+
+      def push(events)
+        if events.is_a?(Array)
+          # Push multiple events
+          event_class = events.first.class
+          raise UnknownEventError, event_class unless @buffers.key?(event_class)
+
+          @buffers[event_class].concat(events)
+        else
+          # Push single event
+          event_class = events.class
+          raise UnknownEventError, event_class unless @buffers.key?(event_class)
+
+          @buffers[event_class].push(events)
+        end
+      end
+
+      def serialize
+        event_count = 0
+
+        event_groups, start, finish = update_time do
+          @buffers.collect do |event_class, buffer|
+            events = buffer.pop
+            next if events.empty?
+
+            event_count += events.length
+            EventGroup.new(event_class, events)
+          end.compact
+        end
+
+        return if event_count.zero? # We don't want to report empty profiles
+
+        encoded_pprof =
+          Datadog::Profiling::Encoding::Profile::Protobuf.encode(
+            event_count: event_count,
+            event_groups: event_groups,
+            start: start,
+            finish: finish,
+          )
+
+        [start, finish, encoded_pprof]
+      end
+
+      def reset_after_fork
+        Datadog.logger.debug('Resetting OldRecorder in child process after fork')
+
+        # NOTE: A bit of a heavy-handed approach, but it doesn't happen often and this class will be removed soon anyway
+        serialize
+        nil
+      end
+
+      # Error when event of an unknown type is used with the OldRecorder
+      class UnknownEventError < StandardError
+        attr_reader :event_class
+
+        def initialize(event_class)
+          @event_class = event_class
+        end
+
+        def message
+          @message ||= "Unknown event class '#{event_class}' for profiling recorder."
+        end
+      end
+
+      private
+
+      def update_time
+        start = @last_flush_time
+        result = yield
+        @last_flush_time = Time.now.utc
+
+        # Return event groups, start time, finish time
+        [result, start, @last_flush_time]
+      end
+    end
+  end
+end

data/lib/datadog/profiling/pprof/builder.rb

@@ -0,0 +1,125 @@
+# frozen_string_literal: true
+
+require_relative '../flush'
+require_relative 'message_set'
+require_relative 'string_table'
+require_relative '../../core/utils/time'
+
+module Datadog
+  module Profiling
+    module Pprof
+      # Accumulates profile data and produces a Perftools::Profiles::Profile
+      class Builder
+        DEFAULT_ENCODING = 'UTF-8'
+        DESC_FRAME_OMITTED = 'frame omitted'
+        DESC_FRAMES_OMITTED = 'frames omitted'
+
+        attr_reader \
+          :functions,
+          :locations,
+          :mappings,
+          :sample_types,
+          :samples,
+          :string_table
+
+        def initialize
+          @functions = MessageSet.new(1)
+          @locations = initialize_locations_hash
+          @mappings = MessageSet.new(1)
+          @sample_types = MessageSet.new
+          @samples = []
+          @string_table = StringTable.new
+
+          # Cache this proc, since it's pretty expensive to keep recreating it
+          @build_function = method(:build_function).to_proc
+        end
+
+        # The locations hash maps unique BacktraceLocation instances to their corresponding pprof Location objects;
+        # there's a 1:1 correspondence, since BacktraceLocations were already deduped
+        def initialize_locations_hash
+          sequence = Core::Utils::Sequence.new(1)
+          Hash.new do |locations_hash, backtrace_location|
+            locations_hash[backtrace_location] = build_location(sequence.next, backtrace_location)
+          end
+        end
+
+        def encode_profile(profile)
+          Perftools::Profiles::Profile.encode(profile).force_encoding(DEFAULT_ENCODING)
+        end
+
+        def build_profile(start:, finish:)
+          start_ns = Core::Utils::Time.as_utc_epoch_ns(start)
+          finish_ns = Core::Utils::Time.as_utc_epoch_ns(finish)
+
+          Perftools::Profiles::Profile.new(
+            sample_type: @sample_types.messages,
+            sample: @samples,
+            mapping: @mappings.messages,
+            location: @locations.values,
+            function: @functions.messages,
+            string_table: @string_table.strings,
+            time_nanos: start_ns,
+            duration_nanos: finish_ns - start_ns,
+          )
+        end
+
+        def build_value_type(type, unit)
+          Perftools::Profiles::ValueType.new(
+            type: @string_table.fetch(type),
+            unit: @string_table.fetch(unit)
+          )
+        end
+
+        def build_locations(backtrace_locations, length)
+          locations = backtrace_locations.collect { |backtrace_location| @locations[backtrace_location] }
+
+          omitted = length - backtrace_locations.length
+
+          # Add placeholder stack frame if frames were truncated
+          if omitted > 0
+            desc = omitted == 1 ? DESC_FRAME_OMITTED : DESC_FRAMES_OMITTED
+            locations << @locations[Profiling::BacktraceLocation.new('', 0, "#{omitted} #{desc}")]
+          end
+
+          locations
+        end
+
+        def build_location(id, backtrace_location)
+          Perftools::Profiles::Location.new(
+            id: id,
+            line: [build_line(
+              @functions.fetch(
+                backtrace_location.path,
+                backtrace_location.base_label,
+                &@build_function
+              ).id,
+              backtrace_location.lineno
+            )]
+          )
+        end
+
+        def build_line(function_id, line_number)
+          Perftools::Profiles::Line.new(
+            function_id: function_id,
+            line: line_number
+          )
+        end
+
+        def build_function(id, filename, function_name)
+          Perftools::Profiles::Function.new(
+            id: id,
+            name: @string_table.fetch(function_name),
+            filename: @string_table.fetch(filename)
+          )
+        end
+
+        def build_mapping(id, filename)
+          Perftools::Profiles::Mapping.new(
+            id: id,
+            filename: @string_table.fetch(filename)
+          )
+        end
+      end
+    end
+  end
+end