ddtrace 0.46.0 → 0.51.0

data/lib/ddtrace/profiling/exporter.rb

@@ -0,0 +1,23 @@
+require 'ddtrace/profiling/transport/io/client'
+
+module Datadog
+  module Profiling
+    # Writes profiling data to a given transport
+    class Exporter
+      attr_reader \
+        :transport
+
+      def initialize(transport)
+        unless transport.is_a?(Profiling::Transport::Client)
+          raise ArgumentError, 'Unsupported transport for profiling exporter.'
+        end
+
+        @transport = transport
+      end
+
+      def export(flush)
+        transport.send_profiling_flush(flush)
+      end
+    end
+  end
+end

data/lib/ddtrace/profiling/ext/cpu.rb

@@ -0,0 +1,67 @@
+module Datadog
+  module Profiling
+    module Ext
+      # Monkey patches Ruby's `Thread` with our `Ext::CThread` to enable CPU-time profiling
+      module CPU
+        # We cannot apply our CPU extension if a broken rollbar is around because that can cause customer apps to fail
+        # with a SystemStackError: stack level too deep.
+        #
+        # This occurs whenever our extensions to Thread are applied BEFORE rollbar applies its own. This happens
+        # because a loop forms: our extension tries to call Thread#initialize, but it's intercepted by rollbar, which
+        # then tries to call the original Thread#initialize as well, but instead alls our extension, leading to stack
+        # exhaustion.
+        #
+        # See https://github.com/rollbar/rollbar-gem/pull/1018 for more details on the issue
+        ROLLBAR_INCOMPATIBLE_VERSIONS = Gem::Requirement.new('<= 3.1.1')
+
+        def self.supported?
+          unsupported_reason.nil?
+        end
+
+        def self.apply!
+          return false unless supported?
+
+          # Applying CThread to Thread will ensure any new threads
+          # will provide a thread/clock ID for CPU timing.
+          require 'ddtrace/profiling/ext/cthread'
+          ::Thread.prepend(Profiling::Ext::CThread)
+          ::Thread.singleton_class.prepend(Datadog::Profiling::Ext::WrapThreadStartFork)
+        end
+
+        def self.unsupported_reason
+          # NOTE: Only the first matching reason is returned, so try to keep a nice order on reasons -- e.g. tell users
+          # first that they can't use this on macOS before telling them that they have the wrong ffi version
+
+          if RUBY_ENGINE == 'jruby'
+            'JRuby is not supported'
+          elsif RUBY_PLATFORM.include?('darwin')
+            'Feature requires Linux; macOS is not supported'
+          elsif RUBY_PLATFORM =~ /(mswin|mingw)/
+            'Feature requires Linux; Windows is not supported'
+          elsif !RUBY_PLATFORM.include?('linux')
+            "Feature requires Linux; #{RUBY_PLATFORM} is not supported"
+          elsif Gem::Specification.find_all_by_name('rollbar', ROLLBAR_INCOMPATIBLE_VERSIONS).any?
+            'You have an incompatible rollbar gem version installed; ensure that you have rollbar >= 3.1.2 by ' \
+            "adding `gem 'rollbar', '>= 3.1.2'` to your Gemfile or gems.rb file. " \
+            'See https://github.com/rollbar/rollbar-gem/pull/1018 for details'
+          elsif Gem::Specification.find_all_by_name('logging').any? && logging_inherit_context_enabled?
+            'The `logging` gem is installed and its thread inherit context feature is enabled. ' \
+            "Please add LOGGING_INHERIT_CONTEXT=false to your application's environment variables to disable the " \
+            'conflicting `logging` gem feature. ' \
+            'See https://github.com/TwP/logging/pull/230 for details'
+          end
+        end
+
+        private_class_method def self.logging_inherit_context_enabled?
+          # The logging gem provides a mechanism to disable the conflicting behavior, see
+          # https://github.com/TwP/logging/blob/ae9872d093833b2a5a34cbe1faa4e895a81f6845/lib/logging/diagnostic_context.rb#L418
+          # Here we check if the behavior is enabled
+          inherit_context_configuration = ENV['LOGGING_INHERIT_CONTEXT']
+
+          inherit_context_configuration.nil? ||
+          (inherit_context_configuration && !%w[false no 0].include?(inherit_context_configuration.downcase))
+        end
+      end
+    end
+  end
+end

data/lib/ddtrace/profiling/ext/cthread.rb

@@ -0,0 +1,155 @@
+require 'ffi'
+
+module Datadog
+  module Profiling
+    module Ext
+      # C-struct for retrieving clock ID from pthread
+      class CClockId < FFI::Struct
+        layout :value, :int
+      end
+
+      # Enables interfacing with pthread via FFI
+      module NativePthread
+        extend FFI::Library
+        ffi_lib ['pthread', 'libpthread.so.0']
+        attach_function :pthread_self, [], :ulong
+        attach_function :pthread_getcpuclockid, [:ulong, CClockId], :int
+
+        # NOTE: Only returns thread ID for thread that evaluates this call.
+        #       a.k.a. evaluating `get_pthread_thread_id(thread_a)` from within
+        #       `thread_b` will return `thread_b`'s thread ID, not `thread_a`'s.
+        def self.get_pthread_thread_id(thread)
+          return unless ::Thread.current == thread
+
+          pthread_self
+        end
+
+        def self.get_clock_id(thread, pthread_id)
+          return unless ::Thread.current == thread && pthread_id
+
+          clock = CClockId.new
+          clock[:value] = 0
+          pthread_getcpuclockid(pthread_id, clock).zero? ? clock[:value] : nil
+        end
+      end
+
+      # Extension used to enable CPU-time profiling via use of pthread's `getcpuclockid`.
+      module CThread
+        def self.prepended(base)
+          # Threads that have already been created, will not have resolved
+          # a thread/clock ID. This is because these IDs can only be resolved
+          # from within the thread's execution context, which we do not control.
+          #
+          # We can mitigate this for the current thread via #update_native_ids,
+          # since we are currently running within its execution context. We cannot
+          # do this for any other threads that may have been created already.
+          # (This is why it's important that CThread is applied before anything else runs.)
+          base.current.send(:update_native_ids) if base.current.is_a?(CThread)
+        end
+
+        # Process::Waiter crash workaround:
+        #
+        # This is a workaround for a Ruby VM segfault (usually something like
+        # "[BUG] Segmentation fault at 0x0000000000000008") in the affected Ruby versions.
+        # See https://bugs.ruby-lang.org/issues/17807 and the regression tests added to this module's specs for details.
+        #
+        # In those Ruby versions, there's a very special subclass of `Thread` called `Process::Waiter` that causes VM
+        # crashes whenever something tries to read its instance variables. This subclass of thread only shows up when
+        # the `Process.detach` API gets used.
+        # In this module's specs you can find crash regression tests that include a way of reproducing it.
+        #
+        # The workaround is to use `defined?` to check first if the instance variable exists. This seems to be fine
+        # with Ruby.
+        # Note that this crash doesn't affect `@foo ||=` nor instance variable writes (after the first write ever of any
+        # instance variable on a `Process::Waiter`, then further reads and writes to that or any other instance are OK;
+        # it looks like there's some lazily-created structure that is missing and did not get created).
+        if Gem::Version.new(RUBY_VERSION) < Gem::Version.new('2.3') &&
+           Gem::Version.new(RUBY_VERSION) >= Gem::Version.new('2.7')
+          attr_reader :pthread_thread_id
+        else
+          def pthread_thread_id
+            defined?(@pthread_thread_id) && @pthread_thread_id
+          end
+        end
+
+        def initialize(*args)
+          @pid = ::Process.pid
+          @pthread_thread_id = nil
+          @clock_id = nil
+
+          # Wrap the work block with our own
+          # so we can retrieve the native thread ID within the thread's context.
+          wrapped_block = proc do |*t_args|
+            # Set native thread ID & clock ID
+            update_native_ids
+            yield(*t_args)
+          end
+          wrapped_block.ruby2_keywords if wrapped_block.respond_to?(:ruby2_keywords, true)
+
+          super(*args, &wrapped_block)
+        end
+        ruby2_keywords :initialize if respond_to?(:ruby2_keywords, true)
+
+        def cpu_time(unit = :float_second)
+          ::Process.clock_gettime(clock_id, unit) if clock_id
+        end
+
+        def cpu_time_instrumentation_installed?
+          # If this thread was started before this module was added to Thread OR if something caused the initialize
+          # method above not to be properly called on new threads, this instance variable is never defined (never set to
+          # any value at all, including nil).
+          #
+          # Thus, we can use @clock_id as a canary to detect a thread that has missing instrumentation, because we
+          # know that in initialize above we always set this variable to nil.
+          defined?(@clock_id) != nil
+        end
+
+        private
+
+        def clock_id
+          update_native_ids if forked?
+          defined?(@clock_id) && @clock_id
+        end
+
+        def forked?
+          ::Process.pid != (@pid ||= nil)
+        end
+
+        def update_native_ids
+          # Can only resolve if invoked from same thread
+          return unless ::Thread.current == self
+
+          @pid = ::Process.pid
+          @pthread_thread_id = NativePthread.get_pthread_thread_id(self)
+          @clock_id = NativePthread.get_clock_id(self, @pthread_thread_id)
+        end
+      end
+
+      # Threads in Ruby can be started by creating a new instance of `Thread` (or a subclass) OR by calling
+      # `start`/`fork` on `Thread` (or a subclass).
+      #
+      # This module intercepts calls to `start`/`fork`, ensuring that the `update_native_ids` operation is correctly
+      # called once the new thread starts.
+      #
+      # Note that unlike CThread above, this module should be prepended to the `Thread`'s singleton class, not to
+      # the class.
+      module WrapThreadStartFork
+        def start(*args)
+          # Wrap the work block with our own
+          # so we can retrieve the native thread ID within the thread's context.
+          wrapped_block = proc do |*t_args|
+            # Set native thread ID & clock ID
+            ::Thread.current.send(:update_native_ids)
+            yield(*t_args)
+          end
+          wrapped_block.ruby2_keywords if wrapped_block.respond_to?(:ruby2_keywords, true)
+
+          super(*args, &wrapped_block)
+        end
+        ruby2_keywords :start if respond_to?(:ruby2_keywords, true)
+
+        alias fork start
+      end
+    end
+  end
+end

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

@@ -0,0 +1,97 @@
+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.
+      module Forking
+        def self.supported?
+          Process.respond_to?(:fork)
+        end
+
+        def self.apply!
+          return false unless supported?
+
+          modules = [::Process, ::Kernel]
+          # TODO: Ruby < 2.3 doesn't support Binding#receiver.
+          #       Remove "else #eval" clause when Ruby < 2.3 support is dropped.
+          # NOTE: Modifying the "main" object as we do here is (as far as I know) irreversible. During tests, this change
+          #       will stick around even if we otherwise stub `Process` and `Kernel`.
+          modules << (TOPLEVEL_BINDING.respond_to?(:receiver) ? TOPLEVEL_BINDING.receiver : TOPLEVEL_BINDING.eval('self'))
+
+          # Patch top-level binding, Kernel, Process.
+          # NOTE: We could instead do Kernel.module_eval { def fork; ... end }
+          #       however, this method rewrite is more invasive and irreversible.
+          #       It could also have collisions with other libraries that patch.
+          #       Opt to modify the inheritance of each relevant target instead.
+          modules.each do |mod|
+            clazz = if mod.class <= Module
+                      mod.singleton_class
+                    else
+                      mod.class
+                    end
+
+            clazz.prepend(Kernel)
+          end
+        end
+
+        # Extensions for kernel
+        module Kernel
+          FORK_STAGES = [:prepare, :parent, :child].freeze
+
+          def fork
+            # If a block is provided, it must be wrapped to trigger callbacks.
+            child_block = if block_given?
+                            proc do
+                              # Trigger :child callback
+                              at_fork_blocks[:child].each(&:call) if at_fork_blocks.key?(:child)
+
+                              # Invoke original block
+                              yield
+                            end
+                          end
+
+            # Trigger :prepare callback
+            at_fork_blocks[:prepare].each(&:call) if at_fork_blocks.key?(:prepare)
+
+            # 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.
+            # rubocop:disable Style/IfInsideElse
+            if result.nil?
+              # Trigger :child callback
+              at_fork_blocks[:child].each(&:call) if at_fork_blocks.key?(:child)
+            else
+              # Trigger :parent callback
+              at_fork_blocks[:parent].each(&:call) if at_fork_blocks.key?(:parent)
+            end
+            # rubocop:enable Style/IfInsideElse
+
+            # Return PID from #fork
+            result
+          end
+
+          def at_fork(stage = :prepare, &block)
+            raise ArgumentError, 'Bad \'stage\' for ::at_fork' unless FORK_STAGES.include?(stage)
+
+            at_fork_blocks[stage] = [] unless at_fork_blocks.key?(stage)
+            at_fork_blocks[stage] << block
+          end
+
+          module_function
+
+          def 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
+            @@at_fork_blocks ||= {}
+            # rubocop:enable Style/ClassVars
+          end
+        end
+      end
+    end
+  end
+end

data/lib/ddtrace/profiling/flush.rb

@@ -0,0 +1,43 @@
+require 'datadog/core/environment/identity'
+require 'datadog/core/environment/socket'
+
+module Datadog
+  module Profiling
+    # Entity class used to represent metadata for a given profile
+    Flush = Struct.new(
+      :start,
+      :finish,
+      :event_groups,
+      :event_count,
+      :runtime_id,
+      :service,
+      :env,
+      :version,
+      :host,
+      :language,
+      :runtime_engine,
+      :runtime_platform,
+      :runtime_version,
+      :profiler_version,
+      :tags
+    ) do
+      def initialize(*args)
+        super
+        self.runtime_id = runtime_id || Datadog::Core::Environment::Identity.id
+        self.service = service || Datadog.configuration.service
+        self.env = env || Datadog.configuration.env
+        self.version = version || Datadog.configuration.version
+        self.host = host || Datadog::Core::Environment::Socket.hostname
+        self.language = language || Datadog::Core::Environment::Identity.lang
+        self.runtime_engine = runtime_engine || Datadog::Core::Environment::Identity.lang_engine
+        self.runtime_platform = runtime_platform || Datadog::Core::Environment::Identity.lang_platform
+        self.runtime_version = runtime_version || Datadog::Core::Environment::Identity.lang_version
+        self.profiler_version = profiler_version || Datadog::Core::Environment::Identity.tracer_version
+        self.tags = tags || Datadog.configuration.tags
+      end
+    end
+
+    # Represents a collection of events of a specific type being flushed.
+    EventGroup = Struct.new(:event_class, :events).freeze
+  end
+end

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

@@ -0,0 +1,119 @@
+# frozen_string_literal: true
+
+require 'ddtrace/profiling/flush'
+require 'ddtrace/profiling/pprof/message_set'
+require 'ddtrace/profiling/pprof/string_table'
+
+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 = 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
+          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
+          )
+        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