ddtrace 0.47.0 → 0.48.0

data/lib/ddtrace/profiling/encoding/profile.rb

@@ -0,0 +1,31 @@
+require 'set'
+
+require 'ddtrace/profiling/flush'
+require 'ddtrace/profiling/pprof/template'
+
+module Datadog
+  module Profiling
+    module Encoding
+      module Profile
+        # Encodes gathered data into the pprof format
+        module Protobuf
+          module_function
+
+          def encode(flush)
+            return unless flush
+
+            # Create a pprof template from the list of event types
+            event_classes = flush.event_groups.collect(&:event_class).uniq
+            template = Pprof::Template.for_event_classes(event_classes)
+
+            # Add all events to the pprof
+            flush.event_groups.each { |event_group| template.add_events!(event_group.event_class, event_group.events) }
+
+            # Build the profile and encode it
+            template.to_pprof
+          end
+        end
+      end
+    end
+  end
+end

data/lib/ddtrace/profiling/event.rb

@@ -0,0 +1,13 @@
+module Datadog
+  module Profiling
+    # Describes a sample of some data obtained from the runtime.
+    class Event
+      attr_reader \
+        :timestamp
+
+      def initialize(timestamp = nil)
+        @timestamp = timestamp || Time.now.utc.to_f
+      end
+    end
+  end
+end

data/lib/ddtrace/profiling/events/stack.rb

@@ -0,0 +1,102 @@
+require 'ddtrace/profiling/event'
+
+module Datadog
+  module Profiling
+    module Events
+      # Describes a stack profiling event
+      class Stack < Event
+        attr_reader \
+          :frames,
+          :hash,
+          :span_id,
+          :thread_id,
+          :total_frame_count,
+          :trace_id
+
+        def initialize(
+          timestamp,
+          frames,
+          total_frame_count,
+          thread_id,
+          trace_id,
+          span_id
+        )
+          super(timestamp)
+
+          @frames = frames
+          @total_frame_count = total_frame_count
+          @thread_id = thread_id
+          @trace_id = trace_id
+          @span_id = span_id
+
+          @hash = [
+            thread_id,
+            trace_id,
+            span_id,
+            [
+              frames.collect(&:hash),
+              total_frame_count
+            ]
+          ].hash
+        end
+      end
+
+      # Describes a stack sample
+      class StackSample < Stack
+        attr_reader \
+          :cpu_time_interval_ns,
+          :wall_time_interval_ns
+
+        def initialize(
+          timestamp,
+          frames,
+          total_frame_count,
+          thread_id,
+          trace_id,
+          span_id,
+          cpu_time_interval_ns,
+          wall_time_interval_ns
+        )
+          super(
+            timestamp,
+            frames,
+            total_frame_count,
+            thread_id,
+            trace_id,
+            span_id
+          )
+
+          @cpu_time_interval_ns = cpu_time_interval_ns
+          @wall_time_interval_ns = wall_time_interval_ns
+        end
+      end
+
+      # Describes a stack sample with exception
+      class StackExceptionSample < Stack
+        attr_reader \
+          :exception
+
+        def initialize(
+          timestamp,
+          frames,
+          total_frame_count,
+          thread_id,
+          trace_id,
+          span_id,
+          exception
+        )
+          super(
+            timestamp,
+            frames,
+            total_frame_count,
+            thread_id,
+            trace_id,
+            span_id
+          )
+
+          @exception = exception
+        end
+      end
+    end
+  end
+end

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 +1,54 @@
+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.send(:prepend, Profiling::Ext::CThread)
+          ::Thread.singleton_class.send(: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::Version.new(RUBY_VERSION) < Gem::Version.new('2.1')
+            'Ruby >= 2.1 is required'
+          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.'
+          end
+        end
+      end
+    end
+  end
+end

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

@@ -0,0 +1,134 @@
+require 'ffi'
+
+module Datadog
+  module Profiling
+    module Ext
+      # C-struct for retrieving clock ID from pthread
+      class CClockId < FFI::Struct
+        layout :value, :int
+      end
+
+      # Extension used to enable CPU-time profiling via use of Pthread's `getcpuclockid`.
+      module CThread
+        extend FFI::Library
+        ffi_lib ['pthread', 'libpthread.so.0']
+        attach_function :pthread_self, [], :ulong
+        attach_function :pthread_getcpuclockid, [:ulong, CClockId], :int
+
+        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
+
+        attr_reader \
+          :native_thread_id
+
+        def initialize(*args)
+          @pid = ::Process.pid
+          @native_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 clock_id
+          update_native_ids if forked?
+          defined?(@clock_id) && @clock_id
+        end
+
+        def cpu_time(unit = :float_second)
+          return unless clock_id && ::Process.respond_to?(:clock_gettime)
+
+          ::Process.clock_gettime(clock_id, unit)
+        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
+
+        # Retrieves number of classes from runtime
+        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
+          @native_thread_id = get_native_thread_id
+          @clock_id = get_clock_id(@native_thread_id)
+        end
+
+        def get_native_thread_id
+          return unless ::Thread.current == self
+
+          # NOTE: Only returns thread ID for thread that evaluates this call.
+          #       a.k.a. evaluating `thread_a.get_native_thread_id` from within
+          #       `thread_b` will return `thread_b`'s thread ID, not `thread_a`'s.
+          pthread_self
+        end
+
+        def get_clock_id(pthread_id)
+          return unless pthread_id && alive?
+
+          # Build a struct, pass it to Pthread's getcpuclockid function.
+          clock = CClockId.new
+          clock[:value] = 0
+          pthread_getcpuclockid(pthread_id, clock).zero? ? clock[:value] : nil
+        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|
+            if mod.class <= Module
+              mod.singleton_class.class_eval do
+                prepend Kernel
+              end
+            else
+              mod.class.send(:prepend, Kernel)
+            end
+          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