puma 2.1.1 → 7.2.1

data/lib/puma/server.rb

@@ -1,102 +1,195 @@
-require 'rack'
+# frozen_string_literal: true
+
 require 'stringio'
 
-require 'puma/thread_pool'
-require 'puma/const'
-require 'puma/events'
-require 'puma/null_io'
-require 'puma/compat'
-require 'puma/reactor'
-require 'puma/client'
-require 'puma/binder'
-require 'puma/delegation'
-require 'puma/accept_nonblock'
-require 'puma/util'
-
-require 'puma/puma_http11'
-
-unless Puma.const_defined? "IOBuffer"
-  require 'puma/io_buffer'
-end
+require_relative 'thread_pool'
+require_relative 'const'
+require_relative 'log_writer'
+require_relative 'events'
+require_relative 'null_io'
+require_relative 'reactor'
+require_relative 'client'
+require_relative 'binder'
+require_relative 'util'
+require_relative 'request'
+require_relative 'configuration'
+require_relative 'cluster_accept_loop_delay'
 
 require 'socket'
+require 'io/wait' unless Puma::HAS_NATIVE_IO_WAIT
 
 module Puma
-
   # The HTTP Server itself. Serves out a single Rack app.
+  #
+  # This class is used by the `Puma::Single` and `Puma::Cluster` classes
+  # to generate one or more `Puma::Server` instances capable of handling requests.
+  # Each Puma process will contain one `Puma::Server` instance.
+  #
+  # The `Puma::Server` instance pulls requests from the socket, adds them to a
+  # `Puma::Reactor` where they get eventually passed to a `Puma::ThreadPool`.
+  #
+  # Each `Puma::Server` will have one reactor and one thread pool.
   class Server
+    module FiberPerRequest
+      def handle_request(client, requests)
+        Fiber.new do
+          super
+        end.resume
+      end
+    end
 
     include Puma::Const
-    extend  Puma::Delegation
+    include Request
 
+    attr_reader :options
     attr_reader :thread
+    attr_reader :log_writer
     attr_reader :events
-    attr_accessor :app
+    attr_reader :min_threads, :max_threads  # for #stats
+    attr_reader :requests_count             # @version 5.0.0
 
-    attr_accessor :min_threads
-    attr_accessor :max_threads
-    attr_accessor :persistent_timeout
-    attr_accessor :auto_trim_time
+    # @todo the following may be deprecated in the future
+    attr_reader :auto_trim_time, :early_hints, :first_data_timeout,
+      :leak_stack_on_error,
+      :persistent_timeout, :reaping_time
+
+    attr_accessor :app
+    attr_accessor :binder
 
     # Create a server for the rack app +app+.
     #
-    # +events+ is an object which will be called when certain error events occur
-    # to be handled. See Puma::Events for the list of current methods to implement.
+    # +log_writer+ is a Puma::LogWriter object used to log info and error messages.
+    #
+    # +events+ is a Puma::Events object used to notify application status events.
     #
     # Server#run returns a thread that you can join on to wait for the server
-    # to do it's work.
+    # to do its work.
+    #
+    # @note Several instance variables exist so they are available for testing,
+    #   and have default values set via +fetch+.  Normally the values are set via
+    #   `::Puma::Configuration.puma_default_options`.
+    #
+    # @note The `events` parameter is set to nil, and set to `Events.new` in code.
+    #   Often `options` needs to be passed, but `events` does not.  Using nil allows
+    #   calling code to not require events.rb.
     #
-    def initialize(app, events=Events::DEFAULT)
+    def initialize(app, events = nil, options = {})
       @app = app
-      @events = events
-
-      @check, @notify = Puma::Util.pipe
+      @events = events || Events.new
 
+      @check, @notify = nil
       @status = :stop
 
-      @min_threads = 0
-      @max_threads = 16
-      @auto_trim_time = 1
-
       @thread = nil
       @thread_pool = nil
+      @reactor = nil
+
+      @env_set_http_version = nil
 
-      @persistent_timeout = PERSISTENT_TIMEOUT
+      @options = if options.is_a?(UserFileDefaultOptions)
+        options
+      else
+        UserFileDefaultOptions.new(options, Configuration::DEFAULTS)
+      end
+
+      @clustered                 = (@options.fetch :workers, 0) > 0
+      @worker_write              = @options[:worker_write]
+      @log_writer                = @options.fetch :log_writer, LogWriter.stdio
+      @early_hints               = @options[:early_hints]
+      @first_data_timeout        = @options[:first_data_timeout]
+      @persistent_timeout        = @options[:persistent_timeout]
+      @idle_timeout              = @options[:idle_timeout]
+      @min_threads               = @options[:min_threads]
+      @max_threads               = @options[:max_threads]
+      @queue_requests            = @options[:queue_requests]
+      @max_keep_alive            = @options[:max_keep_alive]
+      @enable_keep_alives        = @options[:enable_keep_alives]
+      @enable_keep_alives      &&= @queue_requests
+      @io_selector_backend       = @options[:io_selector_backend]
+      @http_content_length_limit = @options[:http_content_length_limit]
+      @cluster_accept_loop_delay = ClusterAcceptLoopDelay.new(
+        workers: @options[:workers],
+        max_delay: @options[:wait_for_less_busy_worker] || 0 # Real default is in Configuration::DEFAULTS, this is for unit testing
+      )
+
+      if @options[:fiber_per_request]
+        singleton_class.prepend(FiberPerRequest)
+      end
+
+      # make this a hash, since we prefer `key?` over `include?`
+      @supported_http_methods =
+        if @options[:supported_http_methods] == :any
+          :any
+        else
+          if (ary = @options[:supported_http_methods])
+            ary
+          else
+            SUPPORTED_HTTP_METHODS
+          end.sort.product([nil]).to_h.freeze
+        end
 
-      @binder = Binder.new(events)
-      @own_binder = true
+      temp = !!(@options[:environment] =~ /\A(development|test)\z/)
+      @leak_stack_on_error = @options[:environment] ? temp : true
 
-      @first_data_timeout = FIRST_DATA_TIMEOUT
+      @binder = Binder.new(log_writer, @options)
 
       ENV['RACK_ENV'] ||= "development"
-    end
 
-    attr_accessor :binder
+      @mode = :http
 
-    forward :add_tcp_listener,  :@binder
-    forward :add_ssl_listener,  :@binder
-    forward :add_unix_listener, :@binder
+      @precheck_closing = true
+
+      @requests_count = 0
+
+      @idle_timeout_reached = false
+    end
 
     def inherit_binder(bind)
       @binder = bind
-      @own_binder = false
+    end
+
+    class << self
+      # @!attribute [r] current
+      def current
+        Thread.current.puma_server
+      end
+
+      # :nodoc:
+      # @version 5.0.0
+      def tcp_cork_supported?
+        Socket.const_defined?(:TCP_CORK) && Socket.const_defined?(:IPPROTO_TCP)
+      end
+
+      # :nodoc:
+      # @version 5.0.0
+      def closed_socket_supported?
+        Socket.const_defined?(:TCP_INFO) && Socket.const_defined?(:IPPROTO_TCP)
+      end
+      private :tcp_cork_supported?
+      private :closed_socket_supported?
     end
 
     # On Linux, use TCP_CORK to better control how the TCP stack
     # packetizes our stream. This improves both latency and throughput.
+    # socket parameter may be an MiniSSL::Socket, so use to_io
     #
-    if RUBY_PLATFORM =~ /linux/
+    if tcp_cork_supported?
       # 6 == Socket::IPPROTO_TCP
       # 3 == TCP_CORK
       # 1/0 == turn on/off
       def cork_socket(socket)
-        socket.setsockopt(6, 3, 1) if socket.kind_of? TCPSocket
+        skt = socket.to_io
+        begin
+          skt.setsockopt(Socket::IPPROTO_TCP, Socket::TCP_CORK, 1) if skt.kind_of? TCPSocket
+        rescue IOError, SystemCallError
+        end
       end
 
       def uncork_socket(socket)
+        skt = socket.to_io
         begin
-          socket.setsockopt(6, 3, 0) if socket.kind_of? TCPSocket
-        rescue IOError
+          skt.setsockopt(Socket::IPPROTO_TCP, Socket::TCP_CORK, 0) if skt.kind_of? TCPSocket
+        rescue IOError, SystemCallError
         end
       end
     else
@@ -107,12 +200,50 @@ module Puma
       end
     end
 
+    if closed_socket_supported?
+      UNPACK_TCP_STATE_FROM_TCP_INFO = "C".freeze
+
+      def closed_socket?(socket)
+        skt = socket.to_io
+        return false unless skt.kind_of?(TCPSocket) && @precheck_closing
+
+        begin
+          tcp_info = skt.getsockopt(Socket::IPPROTO_TCP, Socket::TCP_INFO)
+        rescue IOError, SystemCallError
+          @precheck_closing = false
+          false
+        else
+          state = tcp_info.unpack(UNPACK_TCP_STATE_FROM_TCP_INFO)[0]
+          # TIME_WAIT: 6, CLOSE: 7, CLOSE_WAIT: 8, LAST_ACK: 9, CLOSING: 11
+          (state >= 6 && state <= 9) || state == 11
+        end
+      end
+    else
+      def closed_socket?(socket)
+        false
+      end
+    end
+
+    # @!attribute [r] backlog
     def backlog
-      @thread_pool and @thread_pool.backlog
+      @thread_pool&.backlog
     end
 
+    # @!attribute [r] running
     def running
-      @thread_pool and @thread_pool.spawned
+      @thread_pool&.spawned
+    end
+
+    # This number represents the number of requests that
+    # the server is capable of taking right now.
+    #
+    # For example if the number is 5 then it means
+    # there are 5 threads sitting idle ready to take
+    # a request. If one request comes in, then the
+    # value would be 4 until it finishes processing.
+    # @!attribute [r] pool_capacity
+    def pool_capacity
+      @thread_pool&.pool_capacity
     end
 
     # Runs the server.
@@ -121,101 +252,205 @@ module Puma
     # up in the background to handle requests. Otherwise requests
     # are handled synchronously.
     #
-    def run(background=true)
+    def run(background=true, thread_name: 'srv')
       BasicSocket.do_not_reverse_lookup = true
 
+      @events.fire :state, :booting
+
       @status = :run
 
-      @thread_pool = ThreadPool.new(@min_threads,
-                                    @max_threads,
-                                    IOBuffer) do |client, buffer|
-        process_now = false
+      @thread_pool = ThreadPool.new(thread_name, options, server: self) { |client| process_client client }
 
-        begin
-          process_now = client.eagerly_finish
-        rescue HttpParserError => e
-          client.write_400
-          client.close
-
-          @events.parse_error self, client.env, e
-        rescue ConnectionError
-          client.close
-        else
-          if process_now
-            process_client client, buffer
-          else
-            client.set_timeout @first_data_timeout
-            @reactor.add client
-          end
-        end
+      if @queue_requests
+        @reactor = Reactor.new(@io_selector_backend) { |c|
+          # Inversion of control, the reactor is calling a method on the server when it
+          # is done buffering a request or receives a new request from a keepalive connection.
+          self.reactor_wakeup(c)
+        }
+        @reactor.run
       end
 
-      @reactor = Reactor.new self, @thread_pool
+      @thread_pool.auto_reap! if options[:reaping_time]
+      @thread_pool.auto_trim! if @min_threads != @max_threads && options[:auto_trim_time]
 
-      @reactor.run_in_thread
+      @check, @notify = Puma::Util.pipe unless @notify
 
-      if @auto_trim_time
-        @thread_pool.auto_trim!(@auto_trim_time)
-      end
+      @events.fire :state, :running
 
       if background
-        @thread = Thread.new { handle_servers }
+        @thread = Thread.new do
+          Puma.set_thread_name thread_name
+          handle_servers
+        end
         return @thread
       else
         handle_servers
       end
     end
 
+    # This method is called from the Reactor thread when a queued Client receives data,
+    # times out, or when the Reactor is shutting down.
+    #
+    # While the code lives in the Server, the logic is executed on the reactor thread, independently
+    # from the server.
+    #
+    # It is responsible for ensuring that a request has been completely received
+    # before it starts to be processed by the ThreadPool. This may be known as read buffering.
+    # If read buffering is not done, and no other read buffering is performed (such as by an application server
+    # such as nginx) then the application would be subject to a slow client attack.
+    #
+    # For a graphical representation of how the request buffer works see [architecture.md](https://github.com/puma/puma/blob/main/docs/architecture.md).
+    #
+    # The method checks to see if it has the full header and body with
+    # the `Puma::Client#try_to_finish` method. If the full request has been sent,
+    # then the request is passed to the ThreadPool (`@thread_pool << client`)
+    # so that a "worker thread" can pick up the request and begin to execute application logic.
+    # The Client is then removed from the reactor (return `true`).
+    #
+    # If a client object times out, a 408 response is written, its connection is closed,
+    # and the object is removed from the reactor (return `true`).
+    #
+    # If the Reactor is shutting down, all Clients are either timed out or passed to the
+    # ThreadPool, depending on their current state (#can_close?).
+    #
+    # Otherwise, if the full request is not ready then the client will remain in the reactor
+    # (return `false`). When the client sends more data to the socket the `Puma::Client` object
+    # will wake up and again be checked to see if it's ready to be passed to the thread pool.
+    def reactor_wakeup(client)
+      shutdown = !@queue_requests
+      if client.try_to_finish || (shutdown && !client.can_close?)
+        @thread_pool << client
+      elsif shutdown || client.timeout == 0
+        client.timeout!
+      else
+        client.set_timeout(@first_data_timeout)
+        false
+      end
+    rescue StandardError => e
+      client_error(e, client)
+      close_client_safely(client)
+      true
+    end
+
     def handle_servers
+      @env_set_http_version = Object.const_defined?(:Rack) && ::Rack.respond_to?(:release) &&
+        Gem::Version.new(::Rack.release) < Gem::Version.new('3.1.0')
+
       begin
         check = @check
         sockets = [check] + @binder.ios
         pool = @thread_pool
+        queue_requests = @queue_requests
+        drain = options[:drain_on_shutdown] ? 0 : nil
+
+        addr_send_name, addr_value = case options[:remote_address]
+        when :value
+          [:peerip=, options[:remote_address_value]]
+        when :header
+          [:remote_addr_header=, options[:remote_address_header]]
+        when :proxy_protocol
+          [:expect_proxy_proto=, options[:remote_address_proxy_protocol]]
+        else
+          [nil, nil]
+        end
 
-        while @status == :run
+        while @status == :run || (drain && shutting_down?)
           begin
-            ios = IO.select sockets
+            ios = IO.select sockets, nil, nil, (shutting_down? ? 0 : @idle_timeout)
+            unless ios
+              unless shutting_down?
+                @idle_timeout_reached = true
+
+                if @clustered
+                  @worker_write << "#{PipeRequest::PIPE_IDLE}#{Process.pid}\n" rescue nil
+                  next
+                else
+                  @log_writer.log "- Idle timeout reached"
+                  @status = :stop
+                end
+              end
+
+              break
+            end
+
+            if @idle_timeout_reached && @clustered
+              @idle_timeout_reached = false
+              @worker_write << "#{PipeRequest::PIPE_IDLE}#{Process.pid}\n" rescue nil
+            end
+
             ios.first.each do |sock|
               if sock == check
                 break if handle_check
               else
-                begin
-                  if io = sock.accept_nonblock
-                    c = Client.new io, @binder.env(sock)
-                    pool << c
-                  end
-                rescue SystemCallError
+                # if ThreadPool out_of_band code is running, we don't want to add
+                # clients until the code is finished.
+                pool.wait_while_out_of_band_running
+
+                # A well rested herd (cluster) runs faster
+                if @cluster_accept_loop_delay.on? && (busy_threads_plus_todo = pool.busy_threads) > 0
+                  delay = @cluster_accept_loop_delay.calculate(
+                    max_threads: @max_threads,
+                    busy_threads_plus_todo: busy_threads_plus_todo
+                  )
+                  sleep(delay)
                 end
+
+                io = begin
+                  sock.accept_nonblock
+                rescue IO::WaitReadable
+                  next
+                end
+                drain += 1 if shutting_down?
+                client = new_client(io, sock)
+                client.send(addr_send_name, addr_value) if addr_value
+                pool << client
               end
             end
-          rescue Errno::ECONNABORTED
-            # client closed the socket even before accept
-            client.close rescue nil
-          rescue Object => e
-            @events.unknown_error self, e, "Listen loop"
+          rescue IOError, Errno::EBADF
+            # In the case that any of the sockets are unexpectedly close.
+            raise
+          rescue StandardError => e
+            @log_writer.unknown_error e, nil, "Listen loop"
           end
         end
 
-        graceful_shutdown if @status == :stop || @status == :restart
-        @reactor.clear! if @status == :restart
+        @log_writer.debug "Drained #{drain} additional connections." if drain
+        @events.fire :state, @status
 
-        @reactor.shutdown
+        if queue_requests
+          @queue_requests = false
+          @reactor.shutdown
+        end
+
+        graceful_shutdown if @status == :stop || @status == :restart
       rescue Exception => e
-        STDERR.puts "Exception handling servers: #{e.message} (#{e.class})"
-        STDERR.puts e.backtrace
+        @log_writer.unknown_error e, nil, "Exception handling servers"
       ensure
-        @check.close
-        @notify.close
-
-        if @status != :restart and @own_binder
-          @binder.close
+        # Errno::EBADF is infrequently raised
+        [@check, @notify].each do |io|
+          begin
+            io.close unless io.closed?
+          rescue Errno::EBADF
+          end
         end
+        @notify = nil
+        @check = nil
       end
+
+      @events.fire :state, :done
+    end
+
+    # :nodoc:
+    def new_client(io, sock)
+      client = Client.new(io, @binder.env(sock))
+      client.listener = sock
+      client.http_content_length_limit = @http_content_length_limit
+      client
     end
 
     # :nodoc:
     def handle_check
-      cmd = @check.read(1) 
+      cmd = @check.read(1)
 
       case cmd
       when STOP_COMMAND
@@ -229,402 +464,275 @@ module Puma
         return true
       end
 
-      return false
+      false
     end
 
-    # Given a connection on +client+, handle the incoming requests.
+    # Given a connection on +client+, handle the incoming requests,
+    # or queue the connection in the Reactor if no request is available.
     #
-    # This method support HTTP Keep-Alive so it may, depending on if the client
+    # This method is called from a ThreadPool worker thread.
+    #
+    # This method supports HTTP Keep-Alive so it may, depending on if the client
     # indicates that it supports keep alive, wait for another request before
     # returning.
     #
-    def process_client(client, buffer)
+    # Return true if one or more requests were processed.
+    def process_client(client)
+      close_socket = true
+
+      requests = 0
+
       begin
-        close_socket = true
+        if @queue_requests && !client.eagerly_finish
 
-        while true
-          case handle_request(client, buffer)
-          when false
-            return
-          when :async
+          client.set_timeout(@first_data_timeout)
+          if @reactor.add client
             close_socket = false
-            return
-          when true
-            buffer.reset
-
-            unless client.reset(@status == :run)
-              close_socket = false
-              client.set_timeout @persistent_timeout
-              @reactor.add client
-              return
-            end
+            return false
           end
         end
 
-      # The client disconnected while we were reading data
-      rescue ConnectionError
-        # Swallow them. The ensure tries to close +client+ down
+        with_force_shutdown(client) do
+          client.finish(@first_data_timeout)
+        end
 
-      # The client doesn't know HTTP well
-      rescue HttpParserError => e
-        client.write_400
+        can_loop = true
+        while can_loop
+          can_loop = false
+          @requests_count += 1
+          case handle_request(client, requests + 1)
+          when :close
+          when :async
+            close_socket = false
+          when :keep_alive
+            requests += 1
+
+            client.reset
 
-        @events.parse_error self, client.env, e
+            # This indicates data exists in the client read buffer and there may be
+            # additional requests on it, so process them
+            next_request_ready = if client.has_back_to_back_requests?
+              with_force_shutdown(client) { client.process_back_to_back_requests }
+            else
+              with_force_shutdown(client) { client.eagerly_finish }
+            end
 
-      # Server error
+            if next_request_ready
+              # When Puma has spare threads, allow this one to be monopolized
+              # Perf optimization for https://github.com/puma/puma/issues/3788
+              if @thread_pool.waiting > 0
+                can_loop = true
+              else
+                @thread_pool << client
+                close_socket = false
+              end
+            elsif @queue_requests
+              client.set_timeout @persistent_timeout
+              if @reactor.add client
+                close_socket = false
+              end
+            end
+          end
+        end
+        true
       rescue StandardError => e
-        client.write_500
+        client_error(e, client, requests)
+        # The ensure tries to close +client+ down
+        requests > 0
+      ensure
+        client.io_buffer.reset
 
-        @events.unknown_error self, e, "Read"
+        close_client_safely(client) if close_socket
+      end
+    end
 
-      ensure
-        buffer.reset
+    # :nodoc:
+    def close_client_safely(client)
+      client.close
+    rescue IOError, SystemCallError
+      # Already closed
+    rescue MiniSSL::SSLError => e
+      @log_writer.ssl_error e, client.io
+    rescue StandardError => e
+      @log_writer.unknown_error e, nil, "Client"
+    end
 
-        begin
-          client.close if close_socket
-        rescue IOError, SystemCallError
-          # Already closed
-        rescue StandardError => e
-          @events.unknown_error self, e, "Client"
-        end
+    # Triggers a client timeout if the thread-pool shuts down
+    # during execution of the provided block.
+    def with_force_shutdown(client, &block)
+      @thread_pool.with_force_shutdown(&block)
+    rescue ThreadPool::ForceShutdown
+      client.timeout!
+    end
+
+    # :nocov:
+
+    # Handle various error types thrown by Client I/O operations.
+    def client_error(e, client, requests = 1)
+      # Swallow, do not log
+      return if [ConnectionError, EOFError].include?(e.class)
+
+      case e
+      when MiniSSL::SSLError
+        lowlevel_error(e, client.env)
+        @log_writer.ssl_error e, client.io
+      when HttpParserError
+        response_to_error(client, requests, e, 400)
+        @log_writer.parse_error e, client
+      when HttpParserError501
+        response_to_error(client, requests, e, 501)
+        @log_writer.parse_error e, client
+      else
+        response_to_error(client, requests, e, 500)
+        @log_writer.unknown_error e, nil, "Read"
       end
     end
 
-    # Given a Hash +env+ for the request read from +client+, add
-    # and fixup keys to comply with Rack's env guidelines.
+    # A fallback rack response if +@app+ raises as exception.
     #
-    def normalize_env(env, client)
-      if host = env[HTTP_HOST]
-        if colon = host.index(":")
-          env[SERVER_NAME] = host[0, colon]
-          env[SERVER_PORT] = host[colon+1, host.bytesize]
+    def lowlevel_error(e, env, status=500)
+      if handler = options[:lowlevel_error_handler]
+        if handler.arity == 1
+          return handler.call(e)
+        elsif handler.arity == 2
+          return handler.call(e, env)
         else
-          env[SERVER_NAME] = host
-          env[SERVER_PORT] = default_server_port(env)
+          return handler.call(e, env, status)
         end
-      else
-        env[SERVER_NAME] = LOCALHOST
-        env[SERVER_PORT] = default_server_port(env)
       end
 
-      unless env[REQUEST_PATH]
-        # it might be a dumbass full host request header
-        uri = URI.parse(env[REQUEST_URI])
-        env[REQUEST_PATH] = uri.path
-
-        raise "No REQUEST PATH" unless env[REQUEST_PATH]
+      if @leak_stack_on_error
+        backtrace = e.backtrace.nil? ? '<no backtrace available>' : e.backtrace.join("\n")
+        [status, {}, ["Puma caught this error: #{e.message} (#{e.class})\n#{backtrace}"]]
+      else
+        [status, {}, [""]]
       end
-
-      env[PATH_INFO] = env[REQUEST_PATH]
-
-      # From http://www.ietf.org/rfc/rfc3875 :
-      # "Script authors should be aware that the REMOTE_ADDR and
-      # REMOTE_HOST meta-variables (see sections 4.1.8 and 4.1.9)
-      # may not identify the ultimate source of the request.
-      # They identify the client for the immediate request to the
-      # server; that client may be a proxy, gateway, or other
-      # intermediary acting on behalf of the actual source client."
-      #
-      env[REMOTE_ADDR] = client.peeraddr.last
     end
 
-    def default_server_port(env)
-      env['HTTP_X_FORWARDED_PROTO'] == 'https' ? PORT_443 : PORT_80
+    def response_to_error(client, requests, err, status_code)
+      status, headers, res_body = lowlevel_error(err, client.env, status_code)
+      prepare_response(status, headers, res_body, requests, client)
     end
+    private :response_to_error
 
-    # Given the request +env+ from +client+ and a partial request body
-    # in +body+, finish reading the body if there is one and invoke
-    # the rack app. Then construct the response and write it back to
-    # +client+
-    #
-    # +cl+ is the previously fetched Content-Length header if there
-    # was one. This is an optimization to keep from having to look
-    # it up again.
+    # Wait for all outstanding requests to finish.
     #
-    def handle_request(req, lines)
-      env = req.env
-      client = req.io
-
-      normalize_env env, client
-
-      env[PUMA_SOCKET] = client
-
-      env[HIJACK_P] = true
-      env[HIJACK] = req
-
-      body = req.body
-
-      head = env[REQUEST_METHOD] == HEAD
-
-      env[RACK_INPUT] = body
-      env[RACK_URL_SCHEME] =  env[HTTPS_KEY] ? HTTPS : HTTP
-
-      # A rack extension. If the app writes #call'ables to this
-      # array, we will invoke them when the request is done.
-      #
-      after_reply = env[RACK_AFTER_REPLY] = []
-
-      begin
-        begin
-          status, headers, res_body = @app.call(env)
-
-          return :async if req.hijacked
-
-          status = status.to_i
-
-          if status == -1
-            unless headers.empty? and res_body == []
-              raise "async response must have empty headers and body"
-            end
-
-            return :async
-          end
-        rescue => e
-          @events.unknown_error self, e, "Rack app"
+    def graceful_shutdown
+      if options[:shutdown_debug]
+        threads = Thread.list
+        total = threads.size
 
-          status, headers, res_body = lowlevel_error(e)
-        end
+        pid = Process.pid
 
-        content_length = nil
-        no_body = head
+        $stdout.syswrite "#{pid}: === Begin thread backtrace dump ===\n"
 
-        if res_body.kind_of? Array and res_body.size == 1
-          content_length = res_body[0].bytesize
+        threads.each_with_index do |t,i|
+          $stdout.syswrite "#{pid}: Thread #{i+1}/#{total}: #{t.inspect}\n"
+          $stdout.syswrite "#{pid}: #{t.backtrace.join("\n#{pid}: ")}\n\n"
         end
+        $stdout.syswrite "#{pid}: === End thread backtrace dump ===\n"
+      end
 
-        cork_socket client
-
-        line_ending = LINE_END
-        colon = COLON
-
-        if env[HTTP_VERSION] == HTTP_11
-          allow_chunked = true
-          keep_alive = env[HTTP_CONNECTION] != CLOSE
-          include_keepalive_header = false
-
-          # An optimization. The most common response is 200, so we can
-          # reply with the proper 200 status without having to compute
-          # the response header.
-          #
-          if status == 200
-            lines << HTTP_11_200
-          else
-            lines.append "HTTP/1.1 ", status.to_s, " ",
-                         HTTP_STATUS_CODES[status], line_ending
+      if @status != :restart
+        @binder.close
+      end
 
-            no_body ||= status < 200 || STATUS_WITH_NO_ENTITY_BODY[status]
-          end
+      if @thread_pool
+        if timeout = options[:force_shutdown_after]
+          @thread_pool.shutdown timeout.to_f
         else
-          allow_chunked = false
-          keep_alive = env[HTTP_CONNECTION] == KEEP_ALIVE
-          include_keepalive_header = keep_alive
-
-          # Same optimization as above for HTTP/1.1
-          #
-          if status == 200
-            lines << HTTP_10_200
-          else
-            lines.append "HTTP/1.0 ", status.to_s, " ",
-                         HTTP_STATUS_CODES[status], line_ending
-
-            no_body ||= status < 200 || STATUS_WITH_NO_ENTITY_BODY[status]
-          end
-        end
-
-        response_hijack = nil
-
-        headers.each do |k, vs|
-          case k
-          when CONTENT_LENGTH2
-            content_length = vs
-            next
-          when TRANSFER_ENCODING
-            allow_chunked = false
-            content_length = nil
-          when CONTENT_TYPE
-            next if no_body
-          when HIJACK
-            response_hijack = vs
-            next
-          end
-
-          vs.split(NEWLINE).each do |v|
-            lines.append k, colon, v, line_ending
-          end
-        end
-
-        if no_body
-          lines << line_ending
-          fast_write client, lines.to_s
-          return keep_alive
-        end
-
-        if include_keepalive_header
-          lines << CONNECTION_KEEP_ALIVE
-        elsif !keep_alive
-          lines << CONNECTION_CLOSE
-        end
-
-        unless response_hijack
-          if content_length
-            lines.append CONTENT_LENGTH_S, content_length.to_s, line_ending
-            chunked = false
-          elsif allow_chunked
-            lines << TRANSFER_ENCODING_CHUNKED
-            chunked = true
-          end
-        end
-
-        lines << line_ending
-
-        fast_write client, lines.to_s
-
-        if response_hijack
-          response_hijack.call client
-          return :async
-        end
-
-        res_body.each do |part|
-          if chunked
-            client.syswrite part.bytesize.to_s(16)
-            client.syswrite line_ending
-            fast_write client, part
-            client.syswrite line_ending
-          else
-            fast_write client, part
-          end
-
-          client.flush
-        end
-
-        if chunked
-          client.syswrite CLOSE_CHUNKED
-          client.flush
+          @thread_pool.shutdown
         end
-
-      ensure
-        uncork_socket client
-
-        body.close
-        res_body.close if res_body.respond_to? :close
-
-        after_reply.each { |o| o.call }
       end
-
-      return keep_alive
     end
 
-    # Given the requset +env+ from +client+ and the partial body +body+
-    # plus a potential Content-Length value +cl+, finish reading
-    # the body and return it.
-    #
-    # If the body is larger than MAX_BODY, a Tempfile object is used
-    # for the body, otherwise a StringIO is used.
-    #
-    def read_body(env, client, body, cl)
-      content_length = cl.to_i
-
-      remain = content_length - body.bytesize
-
-      return StringIO.new(body) if remain <= 0
-
-      # Use a Tempfile if there is a lot of data left
-      if remain > MAX_BODY
-        stream = Tempfile.new(Const::PUMA_TMP_BASE)
-        stream.binmode
+    def notify_safely(message)
+      @notify << message
+    rescue IOError, NoMethodError, Errno::EPIPE, Errno::EBADF
+      # The server, in another thread, is shutting down
+    rescue RuntimeError => e
+      # Temporary workaround for https://bugs.ruby-lang.org/issues/13239
+      if e.message.include?('IOError')
+        # ignore
       else
-        # The body[0,0] trick is to get an empty string in the same
-        # encoding as body.
-        stream = StringIO.new body[0,0]
-      end
-
-      stream.write body
-
-      # Read an odd sized chunk so we can read even sized ones
-      # after this
-      chunk = client.readpartial(remain % CHUNK_SIZE)
-
-      # No chunk means a closed socket
-      unless chunk
-        stream.close
-        return nil
+        raise e
       end
+    end
+    private :notify_safely
 
-      remain -= stream.write(chunk)
-
-      # Raed the rest of the chunks
-      while remain > 0
-        chunk = client.readpartial(CHUNK_SIZE)
-        unless chunk
-          stream.close
-          return nil
-        end
-
-        remain -= stream.write(chunk)
-      end
+    # Stops the acceptor thread and then causes the worker threads to finish
+    # off the request queue before finally exiting.
 
-      stream.rewind
+    def stop(sync=false)
+      notify_safely(STOP_COMMAND)
+      @thread.join if @thread && sync
+    end
 
-      return stream
+    def halt(sync=false)
+      notify_safely(HALT_COMMAND)
+      @thread.join if @thread && sync
     end
 
-    # A fallback rack response if +@app+ raises as exception.
-    #
-    def lowlevel_error(e)
-      [500, {}, ["Puma caught this error: #{e}\n#{e.backtrace.join("\n")}"]]
+    def begin_restart(sync=false)
+      notify_safely(RESTART_COMMAND)
+      @thread.join if @thread && sync
     end
 
-    # Wait for all outstanding requests to finish.
-    #
-    def graceful_shutdown
-      @thread_pool.shutdown if @thread_pool
+    def shutting_down?
+      @status == :stop || @status == :restart
     end
 
-    # Stops the acceptor thread and then causes the worker threads to finish
-    # off the request queue before finally exiting.
-    #
-    def stop(sync=false)
-      begin
-        @notify << STOP_COMMAND
-      rescue IOError
-        # The server, in another thread, is shutting down
-      end
+    # List of methods invoked by #stats.
+    # @version 5.0.0
+    STAT_METHODS = [
+      :backlog,
+      :running,
+      :pool_capacity,
+      :busy_threads,
+      :backlog_max,
+      :max_threads,
+      :requests_count,
+      :reactor_max,
+    ].freeze
+
+    # Returns a hash of stats about the running server for reporting purposes.
+    # @version 5.0.0
+    # @!attribute [r] stats
+    # @return [Hash] hash containing stat info from `Server` and `ThreadPool`
+    def stats
+      stats = @thread_pool&.stats || {}
+      stats[:max_threads]    = @max_threads
+      stats[:requests_count] = @requests_count
+      stats[:reactor_max] = @reactor.reactor_max if @reactor
+      reset_max
+      stats
+    end
 
-      @thread.join if @thread && sync
+    def reset_max
+      @reactor.reactor_max = 0 if @reactor
+      @thread_pool&.reset_max
     end
 
-    def halt(sync=false)
-      begin
-        @notify << HALT_COMMAND
-      rescue IOError
-        # The server, in another thread, is shutting down
-      end
+    # below are 'delegations' to binder
+    # remove in Puma 7?
 
-      @thread.join if @thread && sync
+
+    def add_tcp_listener(host, port, optimize_for_latency = true, backlog = 1024)
+      @binder.add_tcp_listener host, port, optimize_for_latency, backlog
     end
 
-    def begin_restart
-      begin
-        @notify << RESTART_COMMAND
-      rescue IOError
-        # The server, in another thread, is shutting down
-      end
+    def add_ssl_listener(host, port, ctx, optimize_for_latency = true,
+                         backlog = 1024)
+      @binder.add_ssl_listener host, port, ctx, optimize_for_latency, backlog
     end
 
-    def fast_write(io, str)
-      n = 0
-      while true
-        begin
-          n = io.syswrite str
-        rescue Errno::EAGAIN, Errno::EWOULDBLOCK
-          IO.select(nil, [io], nil, 1)
-          retry
-        end
+    def add_unix_listener(path, umask = nil, mode = nil, backlog = 1024)
+      @binder.add_unix_listener path, umask, mode, backlog
+    end
 
-        return if n == str.bytesize
-        str = str.byteslice(n..-1)
-      end
+    # @!attribute [r] connected_ports
+    def connected_ports
+      @binder.connected_ports
     end
-    private :fast_write
   end
 end