1 # -*- encoding: binary -*-
4 require 'unicorn/socket_helper'
7 # Unicorn module containing all of the classes (include C extensions) for running
8 # a Unicorn web server. It contains a minimalist HTTP server with just enough
9 # functionality to service web application requests fast as possible.
11 autoload :Const, 'unicorn/const'
12 autoload :HttpRequest, 'unicorn/http_request'
13 autoload :HttpResponse, 'unicorn/http_response'
14 autoload :Configurator, 'unicorn/configurator'
15 autoload :TeeInput, 'unicorn/tee_input'
16 autoload :Util, 'unicorn/util'
19 def run(app, options = {})
20 HttpServer.new(app, options).start.join
24 # This is the process manager of Unicorn. This manages worker
25 # processes which in turn handle the I/O and application process.
26 # Listener sockets are started in the master process and shared with
27 # forked worker children.
29 class HttpServer < Struct.new(:listener_opts, :timeout, :worker_processes,
30 :before_fork, :after_fork, :before_exec,
31 :logger, :pid, :app, :preload_app,
32 :reexec_pid, :orig_app, :init_listeners,
34 include ::Unicorn::SocketHelper
36 # prevents IO objects in here from being GC-ed
39 # all bound listener sockets
42 # This hash maps PIDs to Workers
45 # We use SELF_PIPE differently in the master and worker processes:
47 # * The master process never closes or reinitializes this once
48 # initialized. Signal handlers in the master process will write to
49 # it to wake up the master from IO.select in exactly the same manner
50 # djb describes in http://cr.yp.to/docs/selfpipe.html
52 # * The workers immediately close the pipe they inherit from the
53 # master and replace it with a new pipe after forking. This new
54 # pipe is also used to wakeup from IO.select from inside (worker)
55 # signal handlers. However, workers *close* the pipe descriptors in
56 # the signal handlers to raise EBADF in IO.select instead of writing
57 # like we do in the master. We cannot easily use the reader set for
58 # IO.select because LISTENERS is already that set, and it's extra
59 # work (and cycles) to distinguish the pipe FD from the reader set
60 # once IO.select returns. So we're lazy and just close the pipe when
61 # a (rare) signal arrives in the worker and reinitialize the pipe later.
64 # signal queue used for self-piping
67 # constant lookups are faster and we're single-threaded/non-reentrant
68 REQUEST = HttpRequest.new
70 # We populate this at startup so we can figure out how to reexecute
71 # and upgrade the currently running instance of Unicorn
72 # This Hash is considered a stable interface and changing its contents
73 # will allow you to switch between different installations of Unicorn
74 # or even different installations of the same applications without
75 # downtime. Keys of this constant Hash are described as follows:
77 # * 0 - the path to the unicorn/unicorn_rails executable
78 # * :argv - a deep copy of the ARGV array the executable originally saw
79 # * :cwd - the working directory of the application, this is where
80 # you originally started Unicorn.
82 # The following example may be used in your Unicorn config file to
83 # change your working directory during a config reload (HUP) without
84 # upgrading or restarting:
86 # Dir.chdir(Unicorn::HttpServer::START_CTX[:cwd] = path)
88 # To change your unicorn executable to a different path without downtime,
89 # you can set the following in your Unicorn config file, HUP and then
90 # continue with the traditional USR2 + QUIT upgrade steps:
92 # Unicorn::HttpServer::START_CTX[0] = "/home/bofh/1.9.2/bin/unicorn"
94 :argv => ARGV.map { |arg| arg.dup },
96 # favor ENV['PWD'] since it is (usually) symlink aware for
97 # Capistrano and like systems
99 a = File.stat(pwd = ENV['PWD'])
100 b = File.stat(Dir.pwd)
101 a.ino == b.ino && a.dev == b.dev ? pwd : Dir.pwd
109 # This class and its members can be considered a stable interface
110 # and will not change in a backwards-incompatible fashion between
111 # releases of Unicorn. You may need to access it in the
112 # before_fork/after_fork hooks. See the Unicorn::Configurator RDoc
114 class Worker < Struct.new(:nr, :tmp)
118 # worker objects may be compared to just plain numbers
123 # Changes the worker process to the specified +user+ and +group+
124 # This is only intended to be called from within the worker
125 # process from the +after_fork+ hook. This should be called in
126 # the +after_fork+ hook after any priviledged functions need to be
127 # run (e.g. to set per-worker CPU affinity, niceness, etc)
129 # Any and all errors raised within this method will be propagated
130 # directly back to the caller (usually the +after_fork+ hook.
131 # These errors commonly include ArgumentError for specifying an
132 # invalid user/group and Errno::EPERM for insufficient priviledges
133 def user(user, group = nil)
134 # we do not protect the caller, checking Process.euid == 0 is
135 # insufficient because modern systems have fine-grained
136 # capabilities. Let the caller handle any and all errors.
137 uid = Etc.getpwnam(user).uid
138 gid = Etc.getgrnam(group).gid if group
140 if gid && Process.egid != gid
141 Process.initgroups(user, gid)
142 Process::GID.change_privilege(gid)
144 Process.euid != uid and Process::UID.change_privilege(uid)
149 # Creates a working server on host:port (strange things happen if
150 # port isn't a Number). Use HttpServer::run to start the server and
151 # HttpServer.run.join to join the thread that's processing
152 # incoming requests on the socket.
153 def initialize(app, options = {})
156 self.init_listeners = options[:listeners] ? options[:listeners].dup : []
157 self.config = Configurator.new(options.merge(:use_defaults => true))
158 self.listener_opts = {}
160 # we try inheriting listeners first, so we bind them later.
161 # we don't write the pid file until we've bound listeners in case
162 # unicorn was started twice by mistake. Even though our #pid= method
163 # checks for stale/existing pid files, race conditions are still
164 # possible (and difficult/non-portable to avoid) and can be likely
165 # to clobber the pid if the second start was in quick succession
166 # after the first, so we rely on the listener binding to fail in
167 # that case. Some tests (in and outside of this source tree) and
168 # monitoring tools may also rely on pid files existing before we
169 # attempt to connect to the listener(s)
170 config.commit!(self, :skip => [:listeners, :pid])
174 # Runs the thing. Returns self so you can run join on it
176 BasicSocket.do_not_reverse_lookup = true
178 # inherit sockets from parents, they need to be plain Socket objects
179 # before they become UNIXServer or TCPServer
180 inherited = ENV['UNICORN_FD'].to_s.split(/,/).map do |fd|
181 io = Socket.for_fd(fd.to_i)
182 set_server_sockopt(io, listener_opts[sock_name(io)])
184 logger.info "inherited addr=#{sock_name(io)} fd=#{fd}"
188 config_listeners = config[:listeners].dup
189 LISTENERS.replace(inherited)
191 # we start out with generic Socket objects that get cast to either
192 # TCPServer or UNIXServer objects; but since the Socket objects
193 # share the same OS-level file descriptor as the higher-level *Server
194 # objects; we need to prevent Socket objects from being garbage-collected
195 config_listeners -= listener_names
196 if config_listeners.empty? && LISTENERS.empty?
197 config_listeners << Unicorn::Const::DEFAULT_LISTEN
198 init_listeners << Unicorn::Const::DEFAULT_LISTEN
199 START_CTX[:argv] << "-l#{Unicorn::Const::DEFAULT_LISTEN}"
201 config_listeners.each { |addr| listen(addr) }
202 raise ArgumentError, "no listeners" if LISTENERS.empty?
203 self.pid = config[:pid]
205 build_app! if preload_app
206 maintain_worker_count
210 # replaces current listener set with +listeners+. This will
211 # close the socket if it will not exist in the new listener set
212 def listeners=(listeners)
213 cur_names, dead_names = [], []
214 listener_names.each do |name|
216 # mark unlinked sockets as dead so we can rebind them
217 (File.socket?(name) ? cur_names : dead_names) << name
222 set_names = listener_names(listeners)
223 dead_names.concat(cur_names - set_names).uniq!
225 LISTENERS.delete_if do |io|
226 if dead_names.include?(sock_name(io))
227 IO_PURGATORY.delete_if do |pio|
228 pio.fileno == io.fileno && (pio.close rescue nil).nil? # true
230 (io.close rescue nil).nil? # true
232 set_server_sockopt(io, listener_opts[sock_name(io)])
237 (set_names - cur_names).each { |addr| listen(addr) }
240 def stdout_path=(path); redirect_io($stdout, path); end
241 def stderr_path=(path); redirect_io($stderr, path); end
243 alias_method :set_pid, :pid=
246 # sets the path for the PID file of the master process
249 if x = valid_pid?(path)
250 return path if pid && path == pid && x == $$
251 raise ArgumentError, "Already running on PID:#{x} " \
252 "(or pid=#{path} is stale)"
255 unlink_pid_safe(pid) if pid
259 tmp = "#{File.dirname(path)}/#{rand}.#$$"
260 File.open(tmp, File::RDWR|File::CREAT|File::EXCL, 0644)
265 File.rename(fp.path, path)
271 # add a given address to the +listeners+ set, idempotently
272 # Allows workers to add a private, per-process listener via the
273 # after_fork hook. Very useful for debugging and testing.
274 # +:tries+ may be specified as an option for the number of times
275 # to retry, and +:delay+ may be specified as the time in seconds
276 # to delay between retries.
277 # A negative value for +:tries+ indicates the listen will be
278 # retried indefinitely, this is useful when workers belonging to
279 # different masters are spawned during a transparent upgrade.
280 def listen(address, opt = {}.merge(listener_opts[address] || {}))
281 address = config.expand_addr(address)
282 return if String === address && listener_names.include?(address)
284 delay = opt[:delay] || 0.5
285 tries = opt[:tries] || 5
287 io = bind_listen(address, opt)
288 unless TCPServer === io || UNIXServer === io
292 logger.info "listening on addr=#{sock_name(io)} fd=#{io.fileno}"
295 rescue Errno::EADDRINUSE => err
296 logger.error "adding listener failed addr=#{address} (in use)"
297 raise err if tries == 0
299 logger.error "retrying in #{delay} seconds " \
300 "(#{tries < 0 ? 'infinite' : tries} tries left)"
306 # monitors children and receives signals forever
307 # (or until a termination signal is sent). This handles signals
308 # one-at-a-time time and we'll happily drop signals in case somebody
309 # is signalling us too often.
311 # this pipe is used to wake us up from select(2) in #join when signals
312 # are trapped. See trap_deferred
315 last_check = Time.now
317 QUEUE_SIGS.each { |sig| trap_deferred(sig) }
318 trap(:CHLD) { |sig_nr| awaken_master }
320 logger.info "master process ready" # test_exec.rb relies on this message
326 # avoid murdering workers after our master process (or the
327 # machine) comes out of suspend/hibernation
328 if (last_check + timeout) >= (last_check = Time.now)
331 maintain_worker_count if respawn
333 when :QUIT # graceful shutdown
335 when :TERM, :INT # immediate shutdown
338 when :USR1 # rotate logs
339 logger.info "master reopening logs..."
340 Unicorn::Util.reopen_logs
341 logger.info "master done reopening logs"
342 kill_each_worker(:USR1)
343 when :USR2 # exec binary, stay alive in case something went wrong
346 if Process.ppid == 1 || Process.getpgrp != $$
348 logger.info "gracefully stopping all workers"
349 kill_each_worker(:QUIT)
351 logger.info "SIGWINCH ignored because we're not daemonized"
354 self.worker_processes += 1
356 self.worker_processes -= 1 if self.worker_processes > 0
359 if config.config_file
361 redo # immediate reaping since we may have QUIT workers
362 else # exec binary and exit if there's no config file
363 logger.info "config_file not present, reexecuting binary"
372 logger.error "Unhandled master loop exception #{e.inspect}."
373 logger.error e.backtrace.join("\n")
376 stop # gracefully shutdown all workers on our way out
377 logger.info "master complete"
378 unlink_pid_safe(pid) if pid
381 # Terminates all workers, but does not exit master process
382 def stop(graceful = true)
384 limit = Time.now + timeout
385 until WORKERS.empty? || Time.now > limit
386 kill_each_worker(graceful ? :QUIT : :TERM)
390 kill_each_worker(:KILL)
395 # list of signals we care about and trap in master.
396 QUEUE_SIGS = [ :WINCH, :QUIT, :INT, :TERM, :USR1, :USR2, :HUP,
399 # defer a signal for later processing in #join (master process)
400 def trap_deferred(signal)
401 trap(signal) do |sig_nr|
402 if SIG_QUEUE.size < 5
406 logger.error "ignoring SIG#{signal}, queue=#{SIG_QUEUE.inspect}"
411 # wait for a signal hander to wake us up and then consume the pipe
412 # Wake up every second anyways to run murder_lazy_workers
415 ready = IO.select([SELF_PIPE.first], nil, nil, 1) or return
416 ready.first && ready.first.first or return
417 loop { SELF_PIPE.first.read_nonblock(Const::CHUNK_SIZE) }
418 rescue Errno::EAGAIN, Errno::EINTR
424 SELF_PIPE.last.write_nonblock('.') # wakeup master process from select
425 rescue Errno::EAGAIN, Errno::EINTR
426 # pipe is full, master should wake up anyways
431 # reaps all unreaped workers
435 wpid, status = Process.waitpid2(-1, Process::WNOHANG)
437 if reexec_pid == wpid
438 logger.error "reaped #{status.inspect} exec()-ed"
440 self.pid = pid.chomp('.oldbin') if pid
443 worker = WORKERS.delete(wpid) and worker.tmp.close rescue nil
444 logger.info "reaped #{status.inspect} " \
445 "worker=#{worker.nr rescue 'unknown'}"
452 # reexecutes the START_CTX with a new binary
456 Process.kill(0, reexec_pid)
457 logger.error "reexec-ed child already running PID:#{reexec_pid}"
465 old_pid = "#{pid}.oldbin"
468 self.pid = old_pid # clear the path for a new pid file
470 logger.error "old PID:#{valid_pid?(old_pid)} running with " \
471 "existing pid=#{old_pid}, refusing rexec"
474 logger.error "error writing pid=#{old_pid} #{e.class} #{e.message}"
479 self.reexec_pid = fork do
480 listener_fds = LISTENERS.map { |sock| sock.fileno }
481 ENV['UNICORN_FD'] = listener_fds.join(',')
482 Dir.chdir(START_CTX[:cwd])
483 cmd = [ START_CTX[0] ].concat(START_CTX[:argv])
485 # avoid leaking FDs we don't know about, but let before_exec
486 # unset FD_CLOEXEC, if anything else in the app eventually
487 # relies on FD inheritence.
488 (3..1024).each do |io|
489 next if listener_fds.include?(io)
490 io = IO.for_fd(io) rescue nil
493 io.fcntl(Fcntl::F_SETFD, Fcntl::FD_CLOEXEC)
495 logger.info "executing #{cmd.inspect} (in #{Dir.pwd})"
496 before_exec.call(self)
499 proc_name 'master (old)'
502 # forcibly terminate all workers that haven't checked in in timeout
503 # seconds. The timeout is implemented using an unlinked File
504 # shared between the parent process and each worker. The worker
505 # runs File#chmod to modify the ctime of the File. If the ctime
506 # is stale for >timeout seconds, then we'll kill the corresponding
508 def murder_lazy_workers
509 WORKERS.dup.each_pair do |wpid, worker|
510 (diff = (Time.now - worker.tmp.stat.ctime)) <= timeout and next
511 logger.error "worker=#{worker.nr} PID:#{wpid} timeout " \
512 "(#{diff}s > #{timeout}s), killing"
513 kill_worker(:KILL, wpid) # take no prisoners for timeout violations
517 def spawn_missing_workers
518 (0...worker_processes).each do |worker_nr|
519 WORKERS.values.include?(worker_nr) and next
520 worker = Worker.new(worker_nr, Unicorn::Util.tmpio)
521 before_fork.call(self, worker)
522 WORKERS[fork { worker_loop(worker) }] = worker
526 def maintain_worker_count
527 (off = WORKERS.size - worker_processes) == 0 and return
528 off < 0 and return spawn_missing_workers
529 WORKERS.dup.each_pair { |wpid,w|
530 w.nr >= worker_processes and kill_worker(:QUIT, wpid) rescue nil
534 # if we get any error, try to write something back to the client
535 # assuming we haven't closed the socket, but don't get hung up
536 # if the socket is already closed or broken. We'll always ensure
537 # the socket is closed at the end of this function
538 def handle_error(client, e)
540 when EOFError,Errno::ECONNRESET,Errno::EPIPE,Errno::EINVAL,Errno::EBADF
541 Const::ERROR_500_RESPONSE
542 when HttpParserError # try to tell the client they're bad
543 Const::ERROR_400_RESPONSE
545 logger.error "Read error: #{e.inspect}"
546 logger.error e.backtrace.join("\n")
547 Const::ERROR_500_RESPONSE
549 client.write_nonblock(msg)
555 # once a client is accepted, it is processed in its entirety here
556 # in 3 easy steps: read request, call app, write app response
557 def process_client(client)
558 client.fcntl(Fcntl::F_SETFD, Fcntl::FD_CLOEXEC)
559 response = app.call(env = REQUEST.read(client))
561 if 100 == response.first.to_i
562 client.write(Const::EXPECT_100_RESPONSE)
563 env.delete(Const::HTTP_EXPECT)
564 response = app.call(env)
566 HttpResponse.write(client, response, HttpRequest::PARSER.headers?)
568 handle_error(client, e)
571 # gets rid of stuff the worker has no business keeping track of
572 # to free some resources and drops all sig handlers.
573 # traps for USR1, USR2, and HUP may be set in the after_fork Proc
575 def init_worker_process(worker)
576 QUEUE_SIGS.each { |sig| trap(sig, nil) }
577 trap(:CHLD, 'DEFAULT')
579 proc_name "worker[#{worker.nr}]"
582 WORKERS.values.each { |other| other.tmp.close rescue nil }
584 LISTENERS.each { |sock| sock.fcntl(Fcntl::F_SETFD, Fcntl::FD_CLOEXEC) }
585 worker.tmp.fcntl(Fcntl::F_SETFD, Fcntl::FD_CLOEXEC)
586 after_fork.call(self, worker) # can drop perms
587 self.timeout /= 2.0 # halve it for select()
588 build_app! unless preload_app
591 def reopen_worker_logs(worker_nr)
592 logger.info "worker=#{worker_nr} reopening logs..."
593 Unicorn::Util.reopen_logs
594 logger.info "worker=#{worker_nr} done reopening logs"
598 # runs inside each forked worker, this sits around and waits
599 # for connections and doesn't die until the parent dies (or is
600 # given a INT, QUIT, or TERM signal)
601 def worker_loop(worker)
603 init_worker_process(worker)
604 nr = 0 # this becomes negative if we need to reopen logs
605 alive = worker.tmp # tmp is our lifeline to the master process
608 # closing anything we IO.select on will raise EBADF
609 trap(:USR1) { nr = -65536; SELF_PIPE.first.close rescue nil }
610 trap(:QUIT) { alive = nil; LISTENERS.each { |s| s.close rescue nil } }
611 [:TERM, :INT].each { |sig| trap(sig) { exit!(0) } } # instant shutdown
612 logger.info "worker=#{worker.nr} ready"
616 nr < 0 and reopen_worker_logs(worker.nr)
619 # we're a goner in timeout seconds anyways if alive.chmod
620 # breaks, so don't trap the exception. Using fchmod() since
621 # futimes() is not available in base Ruby and I very strongly
622 # prefer temporary files to be unlinked for security,
623 # performance and reliability reasons, so utime is out. No-op
624 # changes with chmod doesn't update ctime on all filesystems; so
625 # we change our counter each and every time (after process_client
626 # and before IO.select).
627 alive.chmod(m = 0 == m ? 1 : 0)
631 process_client(sock.accept_nonblock)
633 alive.chmod(m = 0 == m ? 1 : 0)
634 rescue Errno::EAGAIN, Errno::ECONNABORTED
639 # make the following bet: if we accepted clients this round,
640 # we're probably reasonably busy, so avoid calling select()
641 # and do a speculative accept_nonblock on ready listeners
642 # before we sleep again in select().
643 redo unless nr == 0 # (nr < 0) => reopen logs
645 ppid == Process.ppid or return
646 alive.chmod(m = 0 == m ? 1 : 0)
648 # timeout used so we can detect parent death:
649 ret = IO.select(LISTENERS, nil, SELF_PIPE, timeout) or redo
658 logger.error "Unhandled listen loop exception #{e.inspect}."
659 logger.error e.backtrace.join("\n")
664 # delivers a signal to a worker and fails gracefully if the worker
665 # is no longer running.
666 def kill_worker(signal, wpid)
668 Process.kill(signal, wpid)
670 worker = WORKERS.delete(wpid) and worker.tmp.close rescue nil
674 # delivers a signal to each worker
675 def kill_each_worker(signal)
676 WORKERS.keys.each { |wpid| kill_worker(signal, wpid) }
679 # unlinks a PID file at given +path+ if it contains the current PID
680 # still potentially racy without locking the directory (which is
681 # non-portable and may interact badly with other programs), but the
682 # window for hitting the race condition is small
683 def unlink_pid_safe(path)
684 (File.read(path).to_i == $$ and File.unlink(path)) rescue nil
687 # returns a PID if a given path contains a non-stale PID file,
690 wpid = File.read(path).to_i
691 wpid <= 0 and return nil
693 Process.kill(0, wpid)
696 # don't unlink stale pid files, racy without non-portable locking...
703 logger.info "reloading config_file=#{config.config_file}"
704 config[:listeners].replace(init_listeners)
707 kill_each_worker(:QUIT)
708 Unicorn::Util.reopen_logs
710 build_app! if preload_app
711 logger.info "done reloading config_file=#{config.config_file}"
713 logger.error "error reloading config_file=#{config.config_file}: " \
714 "#{e.class} #{e.message}"
718 # returns an array of string names for the given listener array
719 def listener_names(listeners = LISTENERS)
720 listeners.map { |io| sock_name(io) }
724 if app.respond_to?(:arity) && app.arity == 0
725 if defined?(Gem) && Gem.respond_to?(:refresh)
726 logger.info "Refreshing Gem list"
734 $0 = ([ File.basename(START_CTX[0]), tag
735 ]).concat(START_CTX[:argv]).join(' ')
738 def redirect_io(io, path)
739 File.open(path, 'ab') { |fp| io.reopen(fp) } if path
744 SELF_PIPE.each { |io| io.close rescue nil }
745 SELF_PIPE.replace(IO.pipe)
746 SELF_PIPE.each { |io| io.fcntl(Fcntl::F_SETFD, Fcntl::FD_CLOEXEC) }