1 #-----------------------------------------------------------
2 # Multicorn, multiprocessing inspired by gunicorn
3 # TODO rename class: Multicorn -> Arbiter ?
4 #-----------------------------------------------------------
18 import werkzeug.serving
20 from setproctitle import setproctitle
22 setproctitle = lambda x: None
25 import openerp.tools.config as config
27 _logger = logging.getLogger(__name__)
29 class Multicorn(object):
30 """ Multiprocessing inspired by (g)unicorn.
31 Multicorn currently uses accept(2) as dispatching method between workers
32 but we plan to replace it by a more intelligent dispatcher to will parse
33 the first HTTP request line.
35 def __init__(self, app):
37 self.address = (config['xmlrpc_interface'] or '0.0.0.0', config['xmlrpc_port'])
38 self.population = config['workers']
39 self.timeout = config['limit_time_real']
40 self.limit_request = config['limit_request']
44 self.pid = os.getpid()
46 self.workers_http = {}
47 self.workers_cron = {}
56 flags = fcntl.fcntl(fd, fcntl.F_GETFL) | os.O_NONBLOCK
57 fcntl.fcntl(fd, fcntl.F_SETFL, flags)
59 flags = fcntl.fcntl(fd, fcntl.F_GETFD) | fcntl.FD_CLOEXEC
60 fcntl.fcntl(fd, fcntl.F_SETFD, flags)
63 def pipe_ping(self, pipe):
65 os.write(pipe[1], '.')
67 if e.errno not in [errno.EAGAIN, errno.EINTR]:
70 def signal_handler(self, sig, frame):
71 if len(self.queue) < 5 or sig == signal.SIGCHLD:
72 self.queue.append(sig)
73 self.pipe_ping(self.pipe)
75 _logger.warn("Dropping signal: %s", sig)
77 def worker_spawn(self, klass, workers_registry):
83 self.workers[pid] = worker
84 workers_registry[pid] = worker
90 def worker_pop(self, pid):
91 if pid in self.workers:
92 _logger.debug("Worker (%s) unregistered",pid)
94 self.workers_http.pop(pid,None)
95 self.workers_cron.pop(pid,None)
96 u = self.workers.pop(pid)
101 def worker_kill(self, pid, sig):
105 if e.errno == errno.ESRCH:
108 def process_signals(self):
109 while len(self.queue):
110 sig = self.queue.pop(0)
111 if sig in [signal.SIGINT,signal.SIGTERM]:
112 raise KeyboardInterrupt
114 def process_zombie(self):
118 wpid, status = os.waitpid(-1, os.WNOHANG)
121 if (status >> 8) == 3:
122 msg = "Critial worker error (%s)"
123 _logger.critical(msg, wpid)
124 raise Exception(msg % wpid)
125 self.worker_pop(wpid)
127 if e.errno == errno.ECHILD:
131 def process_timeout(self):
133 for (pid, worker) in self.workers.items():
134 if now - worker.watchdog_time >= worker.watchdog_timeout:
135 _logger.error("Worker (%s) timeout", pid)
136 self.worker_kill(pid, signal.SIGKILL)
138 def process_spawn(self):
139 while len(self.workers_http) < self.population:
140 self.worker_spawn(WorkerHTTP, self.workers_http)
141 while len(self.workers_cron) < config['max_cron_threads']:
142 self.worker_spawn(WorkerCron, self.workers_cron)
146 # map of fd -> worker
147 fds = dict([(w.watchdog_pipe[0],w) for k,w in self.workers.items()])
148 fd_in = fds.keys() + [self.pipe[0]]
149 # check for ping or internal wakeups
150 ready = select.select(fd_in, [], [], self.beat)
151 # update worker watchdogs
154 fds[fd].watchdog_time = time.time()
157 while os.read(fd, 1):
160 if e.errno not in [errno.EAGAIN]:
162 except select.error, e:
163 if e[0] not in [errno.EINTR]:
167 # wakeup pipe, python doesnt throw EINTR when a syscall is interrupted
168 # by a signal simulating a pseudo SA_RESTART. We write to a pipe in the
169 # signal handler to overcome this behaviour
170 self.pipe = self.pipe_new()
172 signal.signal(signal.SIGINT, self.signal_handler)
173 signal.signal(signal.SIGTERM, self.signal_handler)
174 signal.signal(signal.SIGCHLD, self.signal_handler)
176 self.socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
177 self.socket.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
178 self.socket.setblocking(0)
179 self.socket.bind(self.address)
180 self.socket.listen(8*self.population)
182 def stop(self, graceful=True):
184 _logger.info("Stopping gracefully")
185 limit = time.time() + self.timeout
186 for pid in self.workers.keys():
187 self.worker_kill(pid, signal.SIGTERM)
188 while self.workers and time.time() < limit:
189 self.process_zombie()
192 _logger.info("Stopping forcefully")
193 for pid in self.workers.keys():
194 self.worker_kill(pid, signal.SIGTERM)
196 openerp.cli.server.quit_signals_received = 1
200 _logger.debug("Multiprocess starting")
203 #_logger.debug("Multiprocess beat (%s)",time.time())
204 self.process_signals()
205 self.process_zombie()
206 self.process_timeout()
209 except KeyboardInterrupt:
210 _logger.debug("Multiprocess clean stop")
218 class Worker(object):
220 def __init__(self, multi):
222 self.watchdog_time = time.time()
223 self.watchdog_pipe = multi.pipe_new()
224 self.watchdog_timeout = multi.timeout
225 self.ppid = os.getpid()
228 # should we rename into lifetime ?
229 self.request_max = multi.limit_request
230 self.request_count = 0
232 def setproctitle(self, title=""):
233 setproctitle('openerp: %s %s %s' % (self.__class__.__name__, self.pid, title))
236 os.close(self.watchdog_pipe[0])
237 os.close(self.watchdog_pipe[1])
239 def signal_handler(self, sig, frame):
244 ret = select.select([self.multi.socket], [], [], self.multi.beat)
245 except select.error, e:
246 if e[0] not in [errno.EINTR]:
249 def process_limit(self):
250 # If our parent changed sucide
251 if self.ppid != os.getppid():
252 _logger.info("Worker (%s) Parent changed", self.pid)
255 if self.request_count >= self.request_max:
256 _logger.info("Worker (%d) max request (%s) reached.", self.pid, self.request_count)
258 # Reset the worker if it consumes too much memory (e.g. caused by a memory leak).
259 rss, vms = psutil.Process(os.getpid()).get_memory_info()
260 if vms > config['limit_memory_soft']:
261 _logger.info('Worker (%d) virtual memory limit (%s) reached.', self.pid, vms)
262 self.alive = False # Commit suicide after the request.
264 # VMS and RLIMIT_AS are the same thing: virtual memory, a.k.a. address space
265 soft, hard = resource.getrlimit(resource.RLIMIT_AS)
266 resource.setrlimit(resource.RLIMIT_AS, (config['limit_memory_hard'], hard))
268 # SIGXCPU (exceeded CPU time) signal handler will raise an exception.
269 r = resource.getrusage(resource.RUSAGE_SELF)
270 cpu_time = r.ru_utime + r.ru_stime
271 def time_expired(n, stack):
272 _logger.info('Worker (%d) CPU time limit (%s) reached.', config['limit_time_cpu'])
273 # We dont suicide in such case
274 raise Exception('CPU time limit exceeded.')
275 signal.signal(signal.SIGXCPU, time_expired)
276 soft, hard = resource.getrlimit(resource.RLIMIT_CPU)
277 resource.setrlimit(resource.RLIMIT_CPU, (cpu_time + config['limit_time_cpu'], hard))
279 def process_work(self):
283 self.pid = os.getpid()
285 _logger.info("Worker %s (%s) alive", self.__class__.__name__, self.pid)
286 # Reseed the random number generator
288 # Prevent fd inherientence close_on_exec
289 flags = fcntl.fcntl(self.multi.socket, fcntl.F_GETFD) | fcntl.FD_CLOEXEC
290 fcntl.fcntl(self.multi.socket, fcntl.F_SETFD, flags)
291 # reset blocking status
292 self.multi.socket.setblocking(0)
293 signal.signal(signal.SIGINT, self.signal_handler)
294 signal.signal(signal.SIGTERM, signal.SIG_DFL)
295 signal.signal(signal.SIGCHLD, signal.SIG_DFL)
305 self.multi.pipe_ping(self.watchdog_pipe)
308 _logger.info("Worker (%s) exiting. request_count: %s.", self.pid, self.request_count)
311 _logger.exception("Worker (%s) Exception occured, exiting..." % self.pid)
312 # should we use 3 to abort everything ?
315 class WorkerHTTP(Worker):
316 """ HTTP Request workers """
317 def process_request(self, client, addr):
318 client.setblocking(1)
319 client.setsockopt(socket.IPPROTO_TCP, socket.TCP_NODELAY, 1)
320 # Prevent fd inherientence close_on_exec
321 flags = fcntl.fcntl(client, fcntl.F_GETFD) | fcntl.FD_CLOEXEC
322 fcntl.fcntl(client, fcntl.F_SETFD, flags)
323 # do request using WorkerBaseWSGIServer monkey patched with socket
324 self.server.socket = client
325 # tolerate broken pipe when the http client closes the socket before
326 # receiving the full reply
328 self.server.process_request(client,addr)
330 if e.errno != errno.EPIPE:
332 self.request_count += 1
334 def process_work(self):
336 client, addr = self.multi.socket.accept()
337 self.process_request(client, addr)
338 except socket.error, e:
339 if e[0] not in (errno.EAGAIN, errno.ECONNABORTED):
344 self.server = WorkerBaseWSGIServer(self.multi.app)
346 class WorkerBaseWSGIServer(werkzeug.serving.BaseWSGIServer):
347 """ werkzeug WSGI Server patched to allow using an external listen socket
349 def __init__(self, app):
350 werkzeug.serving.BaseWSGIServer.__init__(self, "1", "1", app)
351 def server_bind(self):
352 # we dont bind beause we use the listen socket of Multicorn#socket
353 # instead we close the socket
356 def server_activate(self):
357 # dont listen as we use Multicorn#socket
360 class WorkerCron(Worker):
363 def __init__(self, multi):
364 super(WorkerCron, self).__init__(multi)
365 # process_work() below process a single database per call.
366 # The variable db_index is keeping track of the next database to
371 # Really sleep once all the databases have been processed.
372 if self.db_index == 0:
373 interval = 60 + self.pid % 10 # chorus effect
377 if config['db_name']:
378 db_names = config['db_name'].split(',')
380 db_names = openerp.netsvc.ExportService._services['db'].exp_list(True)
383 def process_work(self):
384 rpc_request = logging.getLogger('openerp.netsvc.rpc.request')
385 rpc_request_flag = rpc_request.isEnabledFor(logging.DEBUG)
386 _logger.debug("WorkerCron (%s) polling for jobs", self.pid)
387 db_names = self._db_list()
389 self.db_index = (self.db_index + 1) % len(db_names)
390 db_name = db_names[self.db_index]
391 self.setproctitle(db_name)
393 start_time = time.time()
394 start_rss, start_vms = psutil.Process(os.getpid()).get_memory_info()
396 # acquired = openerp.addons.base.ir.ir_cron.ir_cron._acquire_job(db_name)
397 # TODO why isnt openerp.addons.base defined ?
399 acquired = base.ir.ir_cron.ir_cron._acquire_job(db_name)
401 openerp.modules.registry.RegistryManager.delete(db_name)
403 # dont keep cursors in multi database mode
404 if len(db_names) > 1:
405 openerp.sql_db.close_db(db_name)
407 end_time = time.time()
408 end_rss, end_vms = psutil.Process(os.getpid()).get_memory_info()
409 logline = '%s time:%.3fs mem: %sk -> %sk (diff: %sk)' % (db_name, end_time - start_time, start_vms / 1024, end_vms / 1024, (end_vms - start_vms)/1024)
410 _logger.debug("WorkerCron (%s) %s", self.pid, logline)
412 self.request_count += 1
413 if self.request_count >= self.request_max and self.request_max < len(db_names):
414 _logger.error("There are more dabatases to process than allowed "
415 "by the `limit_request` configuration variable: %s more.",
416 len(db_names) - self.request_max)
421 os.nice(10) # mommy always told me to be nice with others...
423 openerp.service.start_internal()
425 # chorus effect: make cron workers do not all start at first database
426 mct = config['max_cron_threads']
427 p = float(self.pid % mct) / mct
428 self.db_index = int(len(self._db_list()) * p)
431 # vim:expandtab:smartindent:tabstop=4:softtabstop=4:shiftwidth=4: