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.
12 # raised inside TeeInput when a client closes the socket inside the
13 # application dispatch. This is always raised with an empty backtrace
14 # since there is nothing in the application stack that is responsible
15 # for client shutdowns/disconnects.
16 class ClientShutdown < EOFError
19 autoload :Const, 'unicorn/const'
20 autoload :HttpRequest, 'unicorn/http_request'
21 autoload :HttpResponse, 'unicorn/http_response'
22 autoload :Configurator, 'unicorn/configurator'
23 autoload :TeeInput, 'unicorn/tee_input'
24 autoload :Util, 'unicorn/util'
27 def run(app, options = {})
28 ready_pipe = options.delete(:ready_pipe)
29 HttpServer.new(app, options).start.join(ready_pipe)
33 # This is the process manager of Unicorn. This manages worker
34 # processes which in turn handle the I/O and application process.
35 # Listener sockets are started in the master process and shared with
36 # forked worker children.
38 class HttpServer < Struct.new(:listener_opts, :timeout, :worker_processes,
39 :before_fork, :after_fork, :before_exec,
40 :logger, :pid, :app, :preload_app,
41 :reexec_pid, :orig_app, :init_listeners,
43 include ::Unicorn::SocketHelper
45 # prevents IO objects in here from being GC-ed
48 # all bound listener sockets
51 # This hash maps PIDs to Workers
54 # We use SELF_PIPE differently in the master and worker processes:
56 # * The master process never closes or reinitializes this once
57 # initialized. Signal handlers in the master process will write to
58 # it to wake up the master from IO.select in exactly the same manner
59 # djb describes in http://cr.yp.to/docs/selfpipe.html
61 # * The workers immediately close the pipe they inherit from the
62 # master and replace it with a new pipe after forking. This new
63 # pipe is also used to wakeup from IO.select from inside (worker)
64 # signal handlers. However, workers *close* the pipe descriptors in
65 # the signal handlers to raise EBADF in IO.select instead of writing
66 # like we do in the master. We cannot easily use the reader set for
67 # IO.select because LISTENERS is already that set, and it's extra
68 # work (and cycles) to distinguish the pipe FD from the reader set
69 # once IO.select returns. So we're lazy and just close the pipe when
70 # a (rare) signal arrives in the worker and reinitialize the pipe later.
73 # signal queue used for self-piping
76 # constant lookups are faster and we're single-threaded/non-reentrant
77 REQUEST = HttpRequest.new
79 # We populate this at startup so we can figure out how to reexecute
80 # and upgrade the currently running instance of Unicorn
81 # This Hash is considered a stable interface and changing its contents
82 # will allow you to switch between different installations of Unicorn
83 # or even different installations of the same applications without
84 # downtime. Keys of this constant Hash are described as follows:
86 # * 0 - the path to the unicorn/unicorn_rails executable
87 # * :argv - a deep copy of the ARGV array the executable originally saw
88 # * :cwd - the working directory of the application, this is where
89 # you originally started Unicorn.
91 # The following example may be used in your Unicorn config file to
92 # change your working directory during a config reload (HUP) without
93 # upgrading or restarting:
95 # Dir.chdir(Unicorn::HttpServer::START_CTX[:cwd] = path)
97 # To change your unicorn executable to a different path without downtime,
98 # you can set the following in your Unicorn config file, HUP and then
99 # continue with the traditional USR2 + QUIT upgrade steps:
101 # Unicorn::HttpServer::START_CTX[0] = "/home/bofh/1.9.2/bin/unicorn"
103 :argv => ARGV.map { |arg| arg.dup },
105 # favor ENV['PWD'] since it is (usually) symlink aware for
106 # Capistrano and like systems
108 a = File.stat(pwd = ENV['PWD'])
109 b = File.stat(Dir.pwd)
110 a.ino == b.ino && a.dev == b.dev ? pwd : Dir.pwd
118 # This class and its members can be considered a stable interface
119 # and will not change in a backwards-incompatible fashion between
120 # releases of Unicorn. You may need to access it in the
121 # before_fork/after_fork hooks. See the Unicorn::Configurator RDoc
123 class Worker < Struct.new(:nr, :tmp)
127 # worker objects may be compared to just plain numbers
132 # Changes the worker process to the specified +user+ and +group+
133 # This is only intended to be called from within the worker
134 # process from the +after_fork+ hook. This should be called in
135 # the +after_fork+ hook after any priviledged functions need to be
136 # run (e.g. to set per-worker CPU affinity, niceness, etc)
138 # Any and all errors raised within this method will be propagated
139 # directly back to the caller (usually the +after_fork+ hook.
140 # These errors commonly include ArgumentError for specifying an
141 # invalid user/group and Errno::EPERM for insufficient priviledges
142 def user(user, group = nil)
143 # we do not protect the caller, checking Process.euid == 0 is
144 # insufficient because modern systems have fine-grained
145 # capabilities. Let the caller handle any and all errors.
146 uid = Etc.getpwnam(user).uid
147 gid = Etc.getgrnam(group).gid if group
148 Unicorn::Util.chown_logs(uid, gid)
150 if gid && Process.egid != gid
151 Process.initgroups(user, gid)
152 Process::GID.change_privilege(gid)
154 Process.euid != uid and Process::UID.change_privilege(uid)
159 # Creates a working server on host:port (strange things happen if
160 # port isn't a Number). Use HttpServer::run to start the server and
161 # HttpServer.run.join to join the thread that's processing
162 # incoming requests on the socket.
163 def initialize(app, options = {})
166 self.init_listeners = options[:listeners] ? options[:listeners].dup : []
167 self.config = Configurator.new(options.merge(:use_defaults => true))
168 self.listener_opts = {}
170 # we try inheriting listeners first, so we bind them later.
171 # we don't write the pid file until we've bound listeners in case
172 # unicorn was started twice by mistake. Even though our #pid= method
173 # checks for stale/existing pid files, race conditions are still
174 # possible (and difficult/non-portable to avoid) and can be likely
175 # to clobber the pid if the second start was in quick succession
176 # after the first, so we rely on the listener binding to fail in
177 # that case. Some tests (in and outside of this source tree) and
178 # monitoring tools may also rely on pid files existing before we
179 # attempt to connect to the listener(s)
180 config.commit!(self, :skip => [:listeners, :pid])
184 # Runs the thing. Returns self so you can run join on it
186 BasicSocket.do_not_reverse_lookup = true
188 # inherit sockets from parents, they need to be plain Socket objects
189 # before they become UNIXServer or TCPServer
190 inherited = ENV['UNICORN_FD'].to_s.split(/,/).map do |fd|
191 io = Socket.for_fd(fd.to_i)
192 set_server_sockopt(io, listener_opts[sock_name(io)])
194 logger.info "inherited addr=#{sock_name(io)} fd=#{fd}"
198 config_listeners = config[:listeners].dup
199 LISTENERS.replace(inherited)
201 # we start out with generic Socket objects that get cast to either
202 # TCPServer or UNIXServer objects; but since the Socket objects
203 # share the same OS-level file descriptor as the higher-level *Server
204 # objects; we need to prevent Socket objects from being garbage-collected
205 config_listeners -= listener_names
206 if config_listeners.empty? && LISTENERS.empty?
207 config_listeners << Unicorn::Const::DEFAULT_LISTEN
208 init_listeners << Unicorn::Const::DEFAULT_LISTEN
209 START_CTX[:argv] << "-l#{Unicorn::Const::DEFAULT_LISTEN}"
211 config_listeners.each { |addr| listen(addr) }
212 raise ArgumentError, "no listeners" if LISTENERS.empty?
213 self.pid = config[:pid]
215 build_app! if preload_app
216 maintain_worker_count
220 # replaces current listener set with +listeners+. This will
221 # close the socket if it will not exist in the new listener set
222 def listeners=(listeners)
223 cur_names, dead_names = [], []
224 listener_names.each do |name|
226 # mark unlinked sockets as dead so we can rebind them
227 (File.socket?(name) ? cur_names : dead_names) << name
232 set_names = listener_names(listeners)
233 dead_names.concat(cur_names - set_names).uniq!
235 LISTENERS.delete_if do |io|
236 if dead_names.include?(sock_name(io))
237 IO_PURGATORY.delete_if do |pio|
238 pio.fileno == io.fileno && (pio.close rescue nil).nil? # true
240 (io.close rescue nil).nil? # true
242 set_server_sockopt(io, listener_opts[sock_name(io)])
247 (set_names - cur_names).each { |addr| listen(addr) }
250 def stdout_path=(path); redirect_io($stdout, path); end
251 def stderr_path=(path); redirect_io($stderr, path); end
253 # sets the path for the PID file of the master process
256 if x = valid_pid?(path)
257 return path if pid && path == pid && x == $$
258 raise ArgumentError, "Already running on PID:#{x} " \
259 "(or pid=#{path} is stale)"
262 unlink_pid_safe(pid) if pid
266 tmp = "#{File.dirname(path)}/#{rand}.#$$"
267 File.open(tmp, File::RDWR|File::CREAT|File::EXCL, 0644)
272 File.rename(fp.path, path)
278 # add a given address to the +listeners+ set, idempotently
279 # Allows workers to add a private, per-process listener via the
280 # after_fork hook. Very useful for debugging and testing.
281 # +:tries+ may be specified as an option for the number of times
282 # to retry, and +:delay+ may be specified as the time in seconds
283 # to delay between retries.
284 # A negative value for +:tries+ indicates the listen will be
285 # retried indefinitely, this is useful when workers belonging to
286 # different masters are spawned during a transparent upgrade.
287 def listen(address, opt = {}.merge(listener_opts[address] || {}))
288 address = config.expand_addr(address)
289 return if String === address && listener_names.include?(address)
291 delay = opt[:delay] || 0.5
292 tries = opt[:tries] || 5
294 io = bind_listen(address, opt)
295 unless TCPServer === io || UNIXServer === io
299 logger.info "listening on addr=#{sock_name(io)} fd=#{io.fileno}"
302 rescue Errno::EADDRINUSE => err
303 logger.error "adding listener failed addr=#{address} (in use)"
304 raise err if tries == 0
306 logger.error "retrying in #{delay} seconds " \
307 "(#{tries < 0 ? 'infinite' : tries} tries left)"
313 # monitors children and receives signals forever
314 # (or until a termination signal is sent). This handles signals
315 # one-at-a-time time and we'll happily drop signals in case somebody
316 # is signalling us too often.
317 def join(ready_pipe = nil)
318 # this pipe is used to wake us up from select(2) in #join when signals
319 # are trapped. See trap_deferred
322 last_check = Time.now
324 QUEUE_SIGS.each { |sig| trap_deferred(sig) }
325 trap(:CHLD) { |sig_nr| awaken_master }
327 logger.info "master process ready" # test_exec.rb relies on this message
329 ready_pipe.syswrite($$.to_s)
337 # avoid murdering workers after our master process (or the
338 # machine) comes out of suspend/hibernation
339 if (last_check + timeout) >= (last_check = Time.now)
342 maintain_worker_count if respawn
344 when :QUIT # graceful shutdown
346 when :TERM, :INT # immediate shutdown
349 when :USR1 # rotate logs
350 logger.info "master reopening logs..."
351 Unicorn::Util.reopen_logs
352 logger.info "master done reopening logs"
353 kill_each_worker(:USR1)
354 when :USR2 # exec binary, stay alive in case something went wrong
357 if Process.ppid == 1 || Process.getpgrp != $$
359 logger.info "gracefully stopping all workers"
360 kill_each_worker(:QUIT)
362 logger.info "SIGWINCH ignored because we're not daemonized"
365 self.worker_processes += 1
367 self.worker_processes -= 1 if self.worker_processes > 0
370 if config.config_file
372 redo # immediate reaping since we may have QUIT workers
373 else # exec binary and exit if there's no config file
374 logger.info "config_file not present, reexecuting binary"
383 logger.error "Unhandled master loop exception #{e.inspect}."
384 logger.error e.backtrace.join("\n")
387 stop # gracefully shutdown all workers on our way out
388 logger.info "master complete"
389 unlink_pid_safe(pid) if pid
392 # Terminates all workers, but does not exit master process
393 def stop(graceful = true)
395 limit = Time.now + timeout
396 until WORKERS.empty? || Time.now > limit
397 kill_each_worker(graceful ? :QUIT : :TERM)
401 kill_each_worker(:KILL)
406 # list of signals we care about and trap in master.
407 QUEUE_SIGS = [ :WINCH, :QUIT, :INT, :TERM, :USR1, :USR2, :HUP,
410 # defer a signal for later processing in #join (master process)
411 def trap_deferred(signal)
412 trap(signal) do |sig_nr|
413 if SIG_QUEUE.size < 5
417 logger.error "ignoring SIG#{signal}, queue=#{SIG_QUEUE.inspect}"
422 # wait for a signal hander to wake us up and then consume the pipe
423 # Wake up every second anyways to run murder_lazy_workers
426 ready = IO.select([SELF_PIPE.first], nil, nil, 1) or return
427 ready.first && ready.first.first or return
428 loop { SELF_PIPE.first.read_nonblock(Const::CHUNK_SIZE) }
429 rescue Errno::EAGAIN, Errno::EINTR
435 SELF_PIPE.last.write_nonblock('.') # wakeup master process from select
436 rescue Errno::EAGAIN, Errno::EINTR
437 # pipe is full, master should wake up anyways
442 # reaps all unreaped workers
446 wpid, status = Process.waitpid2(-1, Process::WNOHANG)
448 if reexec_pid == wpid
449 logger.error "reaped #{status.inspect} exec()-ed"
451 self.pid = pid.chomp('.oldbin') if pid
454 worker = WORKERS.delete(wpid) and worker.tmp.close rescue nil
455 logger.info "reaped #{status.inspect} " \
456 "worker=#{worker.nr rescue 'unknown'}"
463 # reexecutes the START_CTX with a new binary
467 Process.kill(0, reexec_pid)
468 logger.error "reexec-ed child already running PID:#{reexec_pid}"
476 old_pid = "#{pid}.oldbin"
479 self.pid = old_pid # clear the path for a new pid file
481 logger.error "old PID:#{valid_pid?(old_pid)} running with " \
482 "existing pid=#{old_pid}, refusing rexec"
485 logger.error "error writing pid=#{old_pid} #{e.class} #{e.message}"
490 self.reexec_pid = fork do
491 listener_fds = LISTENERS.map { |sock| sock.fileno }
492 ENV['UNICORN_FD'] = listener_fds.join(',')
493 Dir.chdir(START_CTX[:cwd])
494 cmd = [ START_CTX[0] ].concat(START_CTX[:argv])
496 # avoid leaking FDs we don't know about, but let before_exec
497 # unset FD_CLOEXEC, if anything else in the app eventually
498 # relies on FD inheritence.
499 (3..1024).each do |io|
500 next if listener_fds.include?(io)
501 io = IO.for_fd(io) rescue nil
504 io.fcntl(Fcntl::F_SETFD, Fcntl::FD_CLOEXEC)
506 logger.info "executing #{cmd.inspect} (in #{Dir.pwd})"
507 before_exec.call(self)
510 proc_name 'master (old)'
513 # forcibly terminate all workers that haven't checked in in timeout
514 # seconds. The timeout is implemented using an unlinked File
515 # shared between the parent process and each worker. The worker
516 # runs File#chmod to modify the ctime of the File. If the ctime
517 # is stale for >timeout seconds, then we'll kill the corresponding
519 def murder_lazy_workers
520 WORKERS.dup.each_pair do |wpid, worker|
521 stat = worker.tmp.stat
522 # skip workers that disable fchmod or have never fchmod-ed
523 stat.mode == 0100600 and next
524 (diff = (Time.now - stat.ctime)) <= timeout and next
525 logger.error "worker=#{worker.nr} PID:#{wpid} timeout " \
526 "(#{diff}s > #{timeout}s), killing"
527 kill_worker(:KILL, wpid) # take no prisoners for timeout violations
531 def spawn_missing_workers
532 (0...worker_processes).each do |worker_nr|
533 WORKERS.values.include?(worker_nr) and next
534 worker = Worker.new(worker_nr, Unicorn::Util.tmpio)
535 before_fork.call(self, worker)
536 WORKERS[fork { worker_loop(worker) }] = worker
540 def maintain_worker_count
541 (off = WORKERS.size - worker_processes) == 0 and return
542 off < 0 and return spawn_missing_workers
543 WORKERS.dup.each_pair { |wpid,w|
544 w.nr >= worker_processes and kill_worker(:QUIT, wpid) rescue nil
548 # if we get any error, try to write something back to the client
549 # assuming we haven't closed the socket, but don't get hung up
550 # if the socket is already closed or broken. We'll always ensure
551 # the socket is closed at the end of this function
552 def handle_error(client, e)
554 when EOFError,Errno::ECONNRESET,Errno::EPIPE,Errno::EINVAL,Errno::EBADF
555 Const::ERROR_500_RESPONSE
556 when HttpParserError # try to tell the client they're bad
557 Const::ERROR_400_RESPONSE
559 logger.error "Read error: #{e.inspect}"
560 logger.error e.backtrace.join("\n")
561 Const::ERROR_500_RESPONSE
563 client.write_nonblock(msg)
569 # once a client is accepted, it is processed in its entirety here
570 # in 3 easy steps: read request, call app, write app response
571 def process_client(client)
572 client.fcntl(Fcntl::F_SETFD, Fcntl::FD_CLOEXEC)
573 response = app.call(env = REQUEST.read(client))
575 if 100 == response.first.to_i
576 client.write(Const::EXPECT_100_RESPONSE)
577 env.delete(Const::HTTP_EXPECT)
578 response = app.call(env)
580 HttpResponse.write(client, response, HttpRequest::PARSER.headers?)
582 handle_error(client, e)
585 # gets rid of stuff the worker has no business keeping track of
586 # to free some resources and drops all sig handlers.
587 # traps for USR1, USR2, and HUP may be set in the after_fork Proc
589 def init_worker_process(worker)
590 QUEUE_SIGS.each { |sig| trap(sig, nil) }
591 trap(:CHLD, 'DEFAULT')
593 proc_name "worker[#{worker.nr}]"
596 WORKERS.values.each { |other| other.tmp.close rescue nil }
598 LISTENERS.each { |sock| sock.fcntl(Fcntl::F_SETFD, Fcntl::FD_CLOEXEC) }
599 worker.tmp.fcntl(Fcntl::F_SETFD, Fcntl::FD_CLOEXEC)
600 after_fork.call(self, worker) # can drop perms
601 self.timeout /= 2.0 # halve it for select()
602 build_app! unless preload_app
605 def reopen_worker_logs(worker_nr)
606 logger.info "worker=#{worker_nr} reopening logs..."
607 Unicorn::Util.reopen_logs
608 logger.info "worker=#{worker_nr} done reopening logs"
612 # runs inside each forked worker, this sits around and waits
613 # for connections and doesn't die until the parent dies (or is
614 # given a INT, QUIT, or TERM signal)
615 def worker_loop(worker)
617 init_worker_process(worker)
618 nr = 0 # this becomes negative if we need to reopen logs
619 alive = worker.tmp # tmp is our lifeline to the master process
622 # closing anything we IO.select on will raise EBADF
623 trap(:USR1) { nr = -65536; SELF_PIPE.first.close rescue nil }
624 trap(:QUIT) { alive = nil; LISTENERS.each { |s| s.close rescue nil } }
625 [:TERM, :INT].each { |sig| trap(sig) { exit!(0) } } # instant shutdown
626 logger.info "worker=#{worker.nr} ready"
630 nr < 0 and reopen_worker_logs(worker.nr)
633 # we're a goner in timeout seconds anyways if alive.chmod
634 # breaks, so don't trap the exception. Using fchmod() since
635 # futimes() is not available in base Ruby and I very strongly
636 # prefer temporary files to be unlinked for security,
637 # performance and reliability reasons, so utime is out. No-op
638 # changes with chmod doesn't update ctime on all filesystems; so
639 # we change our counter each and every time (after process_client
640 # and before IO.select).
641 alive.chmod(m = 0 == m ? 1 : 0)
645 process_client(sock.accept_nonblock)
647 alive.chmod(m = 0 == m ? 1 : 0)
648 rescue Errno::EAGAIN, Errno::ECONNABORTED
653 # make the following bet: if we accepted clients this round,
654 # we're probably reasonably busy, so avoid calling select()
655 # and do a speculative accept_nonblock on ready listeners
656 # before we sleep again in select().
657 redo unless nr == 0 # (nr < 0) => reopen logs
659 ppid == Process.ppid or return
660 alive.chmod(m = 0 == m ? 1 : 0)
662 # timeout used so we can detect parent death:
663 ret = IO.select(LISTENERS, nil, SELF_PIPE, timeout) or redo
672 logger.error "Unhandled listen loop exception #{e.inspect}."
673 logger.error e.backtrace.join("\n")
678 # delivers a signal to a worker and fails gracefully if the worker
679 # is no longer running.
680 def kill_worker(signal, wpid)
682 Process.kill(signal, wpid)
684 worker = WORKERS.delete(wpid) and worker.tmp.close rescue nil
688 # delivers a signal to each worker
689 def kill_each_worker(signal)
690 WORKERS.keys.each { |wpid| kill_worker(signal, wpid) }
693 # unlinks a PID file at given +path+ if it contains the current PID
694 # still potentially racy without locking the directory (which is
695 # non-portable and may interact badly with other programs), but the
696 # window for hitting the race condition is small
697 def unlink_pid_safe(path)
698 (File.read(path).to_i == $$ and File.unlink(path)) rescue nil
701 # returns a PID if a given path contains a non-stale PID file,
704 wpid = File.read(path).to_i
705 wpid <= 0 and return nil
707 Process.kill(0, wpid)
710 # don't unlink stale pid files, racy without non-portable locking...
717 logger.info "reloading config_file=#{config.config_file}"
718 config[:listeners].replace(init_listeners)
721 kill_each_worker(:QUIT)
722 Unicorn::Util.reopen_logs
724 build_app! if preload_app
725 logger.info "done reloading config_file=#{config.config_file}"
727 logger.error "error reloading config_file=#{config.config_file}: " \
728 "#{e.class} #{e.message}"
732 # returns an array of string names for the given listener array
733 def listener_names(listeners = LISTENERS)
734 listeners.map { |io| sock_name(io) }
738 if app.respond_to?(:arity) && app.arity == 0
739 # exploit COW in case of preload_app. Also avoids race
740 # conditions in Rainbows! since load/require are not thread-safe
741 Unicorn.constants.each { |x| Unicorn.const_get(x) }
743 if defined?(Gem) && Gem.respond_to?(:refresh)
744 logger.info "Refreshing Gem list"
752 $0 = ([ File.basename(START_CTX[0]), tag
753 ]).concat(START_CTX[:argv]).join(' ')
756 def redirect_io(io, path)
757 File.open(path, 'ab') { |fp| io.reopen(fp) } if path
762 SELF_PIPE.each { |io| io.close rescue nil }
763 SELF_PIPE.replace(IO.pipe)
764 SELF_PIPE.each { |io| io.fcntl(Fcntl::F_SETFD, Fcntl::FD_CLOEXEC) }