rainbows 2.0.1 → 2.1.0

data/lib/rainbows/rev_thread_pool.rb

@@ -1,76 +1,22 @@
 # -*- encoding: binary -*-
-require 'rainbows/rev/thread'
-
-module Rainbows
-
-  # A combination of the Rev and ThreadPool models.  This allows Ruby
-  # Thread-based concurrency for application processing.  It DOES NOT
-  # expose a streamable "rack.input" for upload processing within the
-  # app.  DevFdResponse should be used with this class to proxy
-  # asynchronous responses.  All network I/O between the client and
-  # server are handled by the main thread and outside of the core
-  # application dispatch.
-  #
-  # Unlike ThreadPool, Rev makes this model highly suitable for
-  # slow clients and applications with medium-to-slow response times
-  # (I/O bound), but less suitable for sleepy applications.
-  #
-  # This concurrency model is designed for Ruby 1.9, and Ruby 1.8
-  # users are NOT advised to use this due to high CPU usage.
-
-  module RevThreadPool
-
-    # :stopdoc:
-    DEFAULTS = {
-      :pool_size => 20, # same default size as ThreadPool (w/o Rev)
-    }
-    #:startdoc:
-
-    def self.setup # :nodoc:
-      DEFAULTS.each { |k,v| O[k] ||= v }
-      Integer === O[:pool_size] && O[:pool_size] > 0 or
-        raise ArgumentError, "pool_size must a be an Integer > 0"
-    end
-
-    class PoolWatcher < ::Rev::TimerWatcher # :nodoc: all
-      def initialize(threads)
-        @threads = threads
-        super(G.server.timeout, true)
-      end
-
-      def on_timer
-        @threads.each { |t| t.join(0) and G.quit! }
-      end
-    end
-
-    class Client < Rainbows::Rev::ThreadClient # :nodoc:
-      def app_dispatch
-        QUEUE << self
-      end
-    end
-
-    include Rainbows::Rev::Core
-
-    def init_worker_threads(master, queue) # :nodoc:
-      O[:pool_size].times.map do
-        Thread.new do
-          begin
-            client = queue.pop
-            master << [ client, client.app_response ]
-          rescue => e
-            Error.listen_loop(e)
-          end while true
-        end
-      end
-    end
-
-    def init_worker_process(worker) # :nodoc:
-      super
-      master = Rev::Master.new(Queue.new).attach(::Rev::Loop.default)
-      queue = Client.const_set(:QUEUE, Queue.new)
-      threads = init_worker_threads(master, queue)
-      PoolWatcher.new(threads).attach(::Rev::Loop.default)
-      logger.info "RevThreadPool pool_size=#{O[:pool_size]}"
-    end
-  end
-end
+# :stopdoc:
+Rainbows.const_set(:RevThreadPool, Rainbows::CoolioThreadPool)
+# :startdoc:
+
+# CoolioThreadPool is the new version of this, use that instead.
+#
+# A combination of the Rev and ThreadPool models.  This allows Ruby
+# Thread-based concurrency for application processing.  It DOES NOT
+# expose a streamable "rack.input" for upload processing within the
+# app.  DevFdResponse should be used with this class to proxy
+# asynchronous responses.  All network I/O between the client and
+# server are handled by the main thread and outside of the core
+# application dispatch.
+#
+# Unlike ThreadPool, Rev makes this model highly suitable for
+# slow clients and applications with medium-to-slow response times
+# (I/O bound), but less suitable for sleepy applications.
+#
+# This concurrency model is designed for Ruby 1.9, and Ruby 1.8
+# users are NOT advised to use this due to high CPU usage.
+module Rainbows::RevThreadPool; end

data/lib/rainbows/rev_thread_spawn.rb

@@ -1,38 +1,20 @@
 # -*- encoding: binary -*-
-require 'rainbows/rev/thread'
+Rainbows.const_set(:RevThreadSpawn, Rainbows::CoolioThreadSpawn)
 
-module Rainbows
-
-  # A combination of the Rev and ThreadSpawn models.  This allows Ruby
-  # Thread-based concurrency for application processing.  It DOES NOT
-  # expose a streamable "rack.input" for upload processing within the
-  # app.  DevFdResponse should be used with this class to proxy
-  # asynchronous responses.  All network I/O between the client and
-  # server are handled by the main thread and outside of the core
-  # application dispatch.
-  #
-  # Unlike ThreadSpawn, Rev makes this model highly suitable for
-  # slow clients and applications with medium-to-slow response times
-  # (I/O bound), but less suitable for sleepy applications.
-  #
-  # Ruby 1.8 users are strongly advised to use Rev >= 0.3.2 to get
-  # usable performance.
-
-  module RevThreadSpawn
-
-    class Client < Rainbows::Rev::ThreadClient # :nodoc: all
-      def app_dispatch
-        Thread.new(self) { |client| MASTER << [ client, app_response ] }
-      end
-    end
-
-    include Rainbows::Rev::Core
-
-    def init_worker_process(worker) # :nodoc:
-      super
-      master = Rev::Master.new(Queue.new).attach(::Rev::Loop.default)
-      Client.const_set(:MASTER, master)
-    end
-
-  end
-end
+# CoolioThreadPool is the new version of this, use that instead.
+#
+# A combination of the Rev and ThreadSpawn models.  This allows Ruby
+# Thread-based concurrency for application processing.  It DOES NOT
+# expose a streamable "rack.input" for upload processing within the
+# app.  DevFdResponse should be used with this class to proxy
+# asynchronous responses.  All network I/O between the client and
+# server are handled by the main thread and outside of the core
+# application dispatch.
+#
+# Unlike ThreadSpawn, Rev makes this model highly suitable for
+# slow clients and applications with medium-to-slow response times
+# (I/O bound), but less suitable for sleepy applications.
+#
+# This concurrency model is designed for Ruby 1.9, and Ruby 1.8
+# users are NOT advised to use this due to high CPU usage.
+module Rainbows::RevThreadSpawn; end

data/lib/rainbows/revactor/body.rb

@@ -8,8 +8,9 @@ module Rainbows::Revactor::Body
 
   if IO.method_defined?(:sendfile_nonblock)
     def write_body_file(client, body, range)
+      body = body_to_io(body)
       sock = client.instance_variable_get(:@_io)
-      pfx = ::Revactor::TCP::Socket === client ? :tcp : :unix
+      pfx = Revactor::TCP::Socket === client ? :tcp : :unix
       write_complete = T[:"#{pfx}_write_complete", client]
       closed = T[:"#{pfx}_closed", client]
       offset, count = range ? range : [ 0, body.stat.size ]
@@ -29,6 +30,8 @@ module Rainbows::Revactor::Body
       rescue EOFError
         break
       end while (count -= n) > 0
+      ensure
+        close_if_private(body)
     end
   else
     ALIASES[:write_body] = :write_body_each

data/lib/rainbows/revactor/tee_socket.rb

@@ -0,0 +1,44 @@
+# -*- encoding: binary -*-
+# :enddoc:
+#
+# Revactor Sockets do not implement readpartial, so we emulate just
+# enough to avoid mucking with TeeInput internals.  Fortunately
+# this code is not heavily used so we can usually avoid the overhead
+# of adding a userspace buffer.
+class Rainbows::Revactor::TeeSocket
+  def initialize(socket)
+    # IO::Buffer is used internally by Rev which Revactor is based on
+    # so we'll always have it available
+    @socket, @rbuf = socket, IO::Buffer.new
+  end
+
+  def leftover
+    @rbuf.read
+  end
+
+  # Revactor socket reads always return an unspecified amount,
+  # sometimes too much
+  def kgio_read(length, dst = "")
+    return dst.replace("") if length == 0
+
+    # always check and return from the userspace buffer first
+    @rbuf.size > 0 and return dst.replace(@rbuf.read(length))
+
+    # read off the socket since there was nothing in rbuf
+    tmp = @socket.read
+
+    # we didn't read too much, good, just return it straight back
+    # to avoid needlessly wasting memory bandwidth
+    tmp.size <= length and return dst.replace(tmp)
+
+    # ugh, read returned too much
+    @rbuf << tmp[length, tmp.size]
+    dst.replace(tmp[0, length])
+    rescue EOFError
+  end
+
+  # just proxy any remaining methods TeeInput may use
+  def close
+    @socket.close
+  end
+end

data/lib/rainbows/revactor.rb

@@ -24,10 +24,11 @@ module Rainbows::Revactor
   RD_ARGS = {}
 
   autoload :Proxy, 'rainbows/revactor/proxy'
+  autoload :TeeSocket, 'rainbows/revactor/tee_socket'
 
   include Rainbows::Base
   LOCALHOST = Kgio::LOCALHOST
-  TCP = ::Revactor::TCP::Socket
+  TCP = Revactor::TCP::Socket
 
   # once a client is accepted, it is processed in its entirety here
   # in 3 easy steps: read request, call app, write app response
@@ -46,13 +47,14 @@ module Rainbows::Revactor
     alive = false
 
     begin
+      ts = nil
       until env = hp.parse
         buf << client.read(*rd_args)
       end
 
       env[CLIENT_IO] = client
       env[RACK_INPUT] = 0 == hp.content_length ?
-               NULL_IO : Unicorn::TeeInput.new(TeeSocket.new(client), hp)
+               NULL_IO : IC.new(ts = TeeSocket.new(client), hp)
       env[REMOTE_ADDR] = remote_addr
       status, headers, body = app.call(env.update(RACK_DEFAULTS))
 
@@ -68,10 +70,11 @@ module Rainbows::Revactor
         alive = hp.next? && G.alive && G.kato > 0
         headers[CONNECTION] = alive ? KEEP_ALIVE : CLOSE
         client.write(response_header(status, headers))
+        alive && ts and buf << ts.leftover
       end
       write_body(client, body, range)
     end while alive
-  rescue ::Revactor::TCP::ReadError
+  rescue Revactor::TCP::ReadError
   rescue => e
     Rainbows::Error.write(io, e)
   ensure
@@ -85,6 +88,7 @@ module Rainbows::Revactor
     init_worker_process(worker)
     require 'rainbows/revactor/body'
     self.class.__send__(:include, Rainbows::Revactor::Body)
+    self.class.const_set(:IC, Unicorn::HttpRequest.input_class)
     RD_ARGS[:timeout] = G.kato if G.kato > 0
     nr = 0
     limit = worker_connections
@@ -133,54 +137,15 @@ module Rainbows::Revactor
     LISTENERS.map do |s|
       case s
       when TCPServer
-        l = ::Revactor::TCP.listen(s, nil)
-        [ l, T[:tcp_closed, ::Revactor::TCP::Socket],
-          T[:tcp_connection, l, ::Revactor::TCP::Socket] ]
+        l = Revactor::TCP.listen(s, nil)
+        [ l, T[:tcp_closed, Revactor::TCP::Socket],
+          T[:tcp_connection, l, Revactor::TCP::Socket] ]
       when UNIXServer
-        l = ::Revactor::UNIX.listen(s)
-        [ l, T[:unix_closed, ::Revactor::UNIX::Socket ],
-          T[:unix_connection, l, ::Revactor::UNIX::Socket] ]
+        l = Revactor::UNIX.listen(s)
+        [ l, T[:unix_closed, Revactor::UNIX::Socket ],
+          T[:unix_connection, l, Revactor::UNIX::Socket] ]
       end
     end
   end
-
-  # Revactor Sockets do not implement readpartial, so we emulate just
-  # enough to avoid mucking with TeeInput internals.  Fortunately
-  # this code is not heavily used so we can usually avoid the overhead
-  # of adding a userspace buffer.
-  class TeeSocket
-    def initialize(socket)
-      # IO::Buffer is used internally by Rev which Revactor is based on
-      # so we'll always have it available
-      @socket, @rbuf = socket, IO::Buffer.new
-    end
-
-    # Revactor socket reads always return an unspecified amount,
-    # sometimes too much
-    def kgio_read(length, dst = "")
-      return dst.replace("") if length == 0
-
-      # always check and return from the userspace buffer first
-      @rbuf.size > 0 and return dst.replace(@rbuf.read(length))
-
-      # read off the socket since there was nothing in rbuf
-      tmp = @socket.read
-
-      # we didn't read too much, good, just return it straight back
-      # to avoid needlessly wasting memory bandwidth
-      tmp.size <= length and return dst.replace(tmp)
-
-      # ugh, read returned too much
-      @rbuf << tmp[length, tmp.size]
-      dst.replace(tmp[0, length])
-      rescue EOFError
-    end
-
-    # just proxy any remaining methods TeeInput may use
-    def close
-      @socket.close
-    end
-  end
-
   # :startdoc:
 end

data/lib/rainbows/socket_proxy.rb

@@ -0,0 +1,24 @@
+# -*- encoding: binary -*-
+# :enddoc:
+#
+module Rainbows::SocketProxy
+  def kgio_addr
+    to_io.kgio_addr
+  end
+
+  def kgio_read(size, buf = "")
+    to_io.kgio_read(size, buf)
+  end
+
+  def kgio_read!(size, buf = "")
+    to_io.kgio_read!(size, buf)
+  end
+
+  def kgio_trywrite(buf)
+    to_io.kgio_trywrite(buf)
+  end
+
+  def timed_read(buf)
+    to_io.timed_read(buf)
+  end
+end

data/lib/rainbows/sync_close.rb

@@ -0,0 +1,37 @@
+# -*- encoding: binary -*-
+# :enddoc:
+require 'thread'
+class Rainbows::SyncClose
+  def initialize(body)
+    @body = body
+    @mutex = Mutex.new
+    @cv = ConditionVariable.new
+    @mutex.synchronize do
+      yield self
+      @cv.wait(@mutex)
+    end
+  end
+
+  def respond_to?(m)
+    @body.respond_to?(m)
+  end
+
+  def to_path
+    @body.to_path
+  end
+
+  def each(&block)
+    @body.each(&block)
+  end
+
+  def to_io
+    @body.to_io
+  end
+
+  # called by the writer thread to wake up the original thread (in #initialize)
+  def close
+    @body.close
+    ensure
+      @mutex.synchronize { @cv.signal }
+  end
+end

data/lib/rainbows/thread_pool.rb

@@ -1,82 +1,78 @@
 # -*- encoding: binary -*-
 
-module Rainbows
+# Implements a worker thread pool model.  This is suited for platforms
+# like Ruby 1.9, where the cost of dynamically spawning a new thread
+# for every new client connection is higher than with the ThreadSpawn
+# model.
+#
+# This model should provide a high level of compatibility with all
+# Ruby implementations, and most libraries and applications.
+# Applications running under this model should be thread-safe
+# but not necessarily reentrant.
+#
+# Applications using this model are required to be thread-safe.
+# Threads are never spawned dynamically under this model.  If you're
+# connecting to external services and need to perform DNS lookups,
+# consider using the "resolv-replace" library which replaces parts of
+# the core Socket package with concurrent DNS lookup capabilities.
+#
+# This model probably less suited for many slow clients than the
+# others and thus a lower +worker_connections+ setting is recommended.
 
-  # Implements a worker thread pool model.  This is suited for platforms
-  # like Ruby 1.9, where the cost of dynamically spawning a new thread
-  # for every new client connection is higher than with the ThreadSpawn
-  # model.
-  #
-  # This model should provide a high level of compatibility with all
-  # Ruby implementations, and most libraries and applications.
-  # Applications running under this model should be thread-safe
-  # but not necessarily reentrant.
-  #
-  # Applications using this model are required to be thread-safe.
-  # Threads are never spawned dynamically under this model.  If you're
-  # connecting to external services and need to perform DNS lookups,
-  # consider using the "resolv-replace" library which replaces parts of
-  # the core Socket package with concurrent DNS lookup capabilities.
-  #
-  # This model probably less suited for many slow clients than the
-  # others and thus a lower +worker_connections+ setting is recommended.
+module Rainbows::ThreadPool
+  include Rainbows::Base
 
-  module ThreadPool
-    include Base
-
-    def worker_loop(worker) # :nodoc:
-      init_worker_process(worker)
-      pool = (1..worker_connections).map do
-        Thread.new { LISTENERS.size == 1 ? sync_worker : async_worker }
-      end
-
-      while G.alive
-        # if any worker dies, something is serious wrong, bail
-        pool.each do |thr|
-          G.tick or break
-          thr.join(1) and G.quit!
-        end
-      end
-      join_threads(pool)
+  def worker_loop(worker) # :nodoc:
+    init_worker_process(worker)
+    pool = (1..worker_connections).map do
+      Thread.new { LISTENERS.size == 1 ? sync_worker : async_worker }
     end
 
-    def sync_worker # :nodoc:
-      s = LISTENERS[0]
-      begin
-        c = s.kgio_accept and process_client(c)
-      rescue => e
-        Error.listen_loop(e)
-      end while G.alive
+    while G.alive
+      # if any worker dies, something is serious wrong, bail
+      pool.each do |thr|
+        G.tick or break
+        thr.join(1) and G.quit!
+      end
     end
+    join_threads(pool)
+  end
 
-    def async_worker # :nodoc:
-      begin
-        # TODO: check if select() or accept() is a problem on large
-        # SMP systems under Ruby 1.9.  Hundreds of native threads
-        # all working off the same socket could be a thundering herd
-        # problem.  On the other hand, a thundering herd may not
-        # even incur as much overhead as an extra Mutex#synchronize
-        ret = IO.select(LISTENERS, nil, nil, 1) and ret[0].each do |s|
-          s = s.kgio_tryaccept and process_client(s)
-        end
-      rescue Errno::EINTR
-      rescue => e
-        Error.listen_loop(e)
-      end while G.alive
-    end
+  def sync_worker # :nodoc:
+    s = LISTENERS[0]
+    begin
+      c = s.kgio_accept and process_client(c)
+    rescue => e
+      Rainbows::Error.listen_loop(e)
+    end while G.alive
+  end
 
-    def join_threads(threads) # :nodoc:
-      G.quit!
-      threads.delete_if do |thr|
-        G.tick
-        begin
-          thr.run
-          thr.join(0.01)
-        rescue
-          true
-        end
-      end until threads.empty?
-    end
+  def async_worker # :nodoc:
+    begin
+      # TODO: check if select() or accept() is a problem on large
+      # SMP systems under Ruby 1.9.  Hundreds of native threads
+      # all working off the same socket could be a thundering herd
+      # problem.  On the other hand, a thundering herd may not
+      # even incur as much overhead as an extra Mutex#synchronize
+      ret = select(LISTENERS) and ret[0].each do |s|
+        s = s.kgio_tryaccept and process_client(s)
+      end
+    rescue Errno::EINTR
+    rescue => e
+      Rainbows::Error.listen_loop(e)
+    end while G.alive
+  end
 
+  def join_threads(threads) # :nodoc:
+    G.quit!
+    threads.delete_if do |thr|
+      G.tick
+      begin
+        thr.run
+        thr.join(0.01)
+      rescue
+        true
+      end
+    end until threads.empty?
   end
 end

data/lib/rainbows/thread_spawn.rb

@@ -1,62 +1,52 @@
 # -*- encoding: binary -*-
 require 'thread'
-module Rainbows
 
-  # Spawns a new thread for every client connection we accept().  This
-  # model is recommended for platforms like Ruby 1.8 where spawning new
-  # threads is inexpensive.
-  #
-  # This model should provide a high level of compatibility with all
-  # Ruby implementations, and most libraries and applications.
-  # Applications running under this model should be thread-safe
-  # but not necessarily reentrant.
-  #
-  # If you're connecting to external services and need to perform DNS
-  # lookups, consider using the "resolv-replace" library which replaces
-  # parts of the core Socket package with concurrent DNS lookup
-  # capabilities
+# Spawns a new thread for every client connection we accept().  This
+# model is recommended for platforms like Ruby 1.8 where spawning new
+# threads is inexpensive.
+#
+# This model should provide a high level of compatibility with all
+# Ruby implementations, and most libraries and applications.
+# Applications running under this model should be thread-safe
+# but not necessarily reentrant.
+#
+# If you're connecting to external services and need to perform DNS
+# lookups, consider using the "resolv-replace" library which replaces
+# parts of the core Socket package with concurrent DNS lookup
+# capabilities
 
-  module ThreadSpawn
-    include Base
+module Rainbows::ThreadSpawn
+  include Rainbows::Base
+  include Rainbows::WorkerYield
 
-    def accept_loop(klass) #:nodoc:
-      lock = Mutex.new
-      limit = worker_connections
-      LISTENERS.each do |l|
-        klass.new(l) do |l|
-          begin
-            if lock.synchronize { G.cur >= limit }
-              # Sleep if we're busy, another less busy worker process may
-              # take it for us if we sleep. This is gross but other options
-              # still suck because they require expensive/complicated
-              # synchronization primitives for _every_ case, not just this
-              # unlikely one.  Since this case is (or should be) uncommon,
-              # just busy wait when we have to.
-              # We don't use Thread.pass because it needlessly spins the
-              # CPU during I/O wait, CPU cycles that can be better used
-              # by other worker _processes_.
-              sleep(0.01)
-            elsif c = l.kgio_accept
-              klass.new(c) do |c|
-                begin
-                  lock.synchronize { G.cur += 1 }
-                  process_client(c)
-                ensure
-                  lock.synchronize { G.cur -= 1 }
-                end
+  def accept_loop(klass) #:nodoc:
+    lock = Mutex.new
+    limit = worker_connections
+    LISTENERS.each do |l|
+      klass.new(l) do |l|
+        begin
+          if lock.synchronize { G.cur >= limit }
+            worker_yield
+          elsif c = l.kgio_accept
+            klass.new(c) do |c|
+              begin
+                lock.synchronize { G.cur += 1 }
+                process_client(c)
+              ensure
+                lock.synchronize { G.cur -= 1 }
               end
             end
-          rescue => e
-            Error.listen_loop(e)
-          end while G.alive
-        end
+          end
+        rescue => e
+          Rainbows::Error.listen_loop(e)
+        end while G.alive
       end
-      sleep 1 while G.tick || lock.synchronize { G.cur > 0 }
     end
+    sleep 1 while G.tick || lock.synchronize { G.cur > 0 }
+  end
 
-    def worker_loop(worker) #:nodoc:
-      init_worker_process(worker)
-      accept_loop(Thread)
-    end
+  def worker_loop(worker) #:nodoc:
+    init_worker_process(worker)
+    accept_loop(Thread)
   end
 end