Glance源码架构探秘(二)
本帖最后由 坎蒂丝_Swan 于 2014-12-6 19:36 编辑问题导读
问题1:WSGI的作用流程是怎样的?问题2:OpenStack的服务为什么没有使用市面上的web server?而又使用了什么库来实现自己的server功能?
static/image/hrline/4.gif
Glance源码架构探秘(一)
Glance源码架构探秘(二)
Glance源码架构探秘(三)
Glance源码架构探秘(四)
上一章的最后,为大家介绍了Glance服务的对外启动接口/bin/glance-api,其中最用要的部分就是通过server = eventlet.wsgi.Server()生成了一个http server,并通过server.start()启动了一个WSGI程序。
首先为大家说说这个WSGI程序的作用。WSGI(web server gateway interface)web服务器网关接口。简单来说,WSGI的作用就是将client发给web server的请求转发给要实际处理这个请求的程序。
WSGI在Python中有官方的参考实现wsgiref,http://docs.python.org/2/library/wsgiref.html#module-wsgiref。
from wsgiref.simple_server import make_server
# Every WSGI application must have an application object - a callable
# object that accepts two arguments. For that purpose, we're going to
# use a function (note that you're not limited to a function, you can
# use a class for example). The first argument passed to the function
# is a dictionary containing CGI-style envrironment variables and the
# second variable is the callable object (see PEP 333).
def hello_world_app(environ, start_response):
status = '200 OK' # HTTP Status
headers = [('Content-type', 'text/plain')] # HTTP Headers
start_response(status, headers)
# The returned object is going to be printed
return ["Hello World"]
httpd = make_server('', 8000, hello_world_app)
print "Serving on port 8000..."
# Serve until process is killed
httpd.serve_forever()
参考流程大概就是server中定义一个start_response()函数,返回值为一个write()函数,用来返回给client的响应。application函数要实现两个接口参数,environ和start_response(),前者就是服务器server传递过来的request请求,application控制后者将程序的返回值发回给web server。
这并不是本文所要讨论的重点,如有兴趣可自行做引申阅读。下面我们分析openstack中wsgi接口的实现,/glance/common/wsgi.py
class Server(object):
"""Server class to manage multiple WSGI sockets and applications."""
def __init__(self, threads=1000):
self.threads = threads
self.children = []
self.running = True
def start(self, application, default_port):
"""
Run a WSGI server with the given application.
:param application: The application to be run in the WSGI server
:param default_port: Port to bind to if none is specified in conf
"""
def kill_children(*args):
"""Kills the entire process group."""
self.logger.info(_('SIGTERM or SIGINT received'))
signal.signal(signal.SIGTERM, signal.SIG_IGN)
signal.signal(signal.SIGINT, signal.SIG_IGN)
self.running = False
os.killpg(0, signal.SIGTERM)
def hup(*args):
"""
Shuts down the server, but allows running requests to complete
"""
self.logger.info(_('SIGHUP received'))
signal.signal(signal.SIGHUP, signal.SIG_IGN)
self.running = False
self.application = application
self.sock = get_socket(default_port)
os.umask(027)# ensure files are created with the correct privileges
self.logger = os_logging.getLogger('eventlet.wsgi.server')
if CONF.workers == 0:
# Useful for profiling, test, debug etc.
self.pool = self.create_pool()
self.pool.spawn_n(self._single_run, self.application, self.sock)
return
else:
self.logger.info(_("Starting %d workers") % CONF.workers)
signal.signal(signal.SIGTERM, kill_children)
signal.signal(signal.SIGINT, kill_children)
signal.signal(signal.SIGHUP, hup)
while len(self.children) < CONF.workers:
self.run_child()
def create_pool(self):
eventlet.patcher.monkey_patch(all=False, socket=True)
return eventlet.GreenPool(size=self.threads)
def wait_on_children(self):
while self.running:
try:
pid, status = os.wait()
if os.WIFEXITED(status) or os.WIFSIGNALED(status):
self.logger.info(_('Removing dead child %s') % pid)
self.children.remove(pid)
if os.WIFEXITED(status) and os.WEXITSTATUS(status) != 0:
self.logger.error(_('Not respawning child %d, cannot '
'recover from termination') % pid)
if not self.children:
self.logger.info(
_('All workers have terminated. Exiting'))
self.running = False
else:
self.run_child()
except OSError, err:
if err.errno not in (errno.EINTR, errno.ECHILD):
raise
except KeyboardInterrupt:
self.logger.info(_('Caught keyboard interrupt. Exiting.'))
break
eventlet.greenio.shutdown_safe(self.sock)
self.sock.close()
self.logger.debug(_('Exited'))
def wait(self):
"""Wait until all servers have completed running."""
try:
if self.children:
self.wait_on_children()
else:
self.pool.waitall()
except KeyboardInterrupt:
pass
def run_child(self):
pid = os.fork()
if pid == 0:
signal.signal(signal.SIGHUP, signal.SIG_DFL)
signal.signal(signal.SIGTERM, signal.SIG_DFL)
# ignore the interrupt signal to avoid a race whereby
# a child worker receives the signal before the parent
# and is respawned unneccessarily as a result
signal.signal(signal.SIGINT, signal.SIG_IGN)
self.run_server()
self.logger.info(_('Child %d exiting normally') % os.getpid())
# self.pool.waitall() has been called by run_server, so
# its safe to exit here
sys.exit(0)
else:
self.logger.info(_('Started child %s') % pid)
self.children.append(pid)
def run_server(self):
"""Run a WSGI server."""
if cfg.CONF.pydev_worker_debug_host:
utils.setup_remote_pydev_debug(cfg.CONF.pydev_worker_debug_host,
cfg.CONF.pydev_worker_debug_port)
eventlet.wsgi.HttpProtocol.default_request_version = "HTTP/1.0"
try:
eventlet.hubs.use_hub('poll')
except Exception:
msg = _("eventlet 'poll' hub is not available on this platform")
raise exception.WorkerCreationFailure(reason=msg)
self.pool = self.create_pool()
try:
eventlet.wsgi.server(self.sock,
self.application,
log=WritableLogger(self.logger),
custom_pool=self.pool)
except socket.error, err:
if err != errno.EINVAL:
raise
self.pool.waitall()
大家可能看了代码之后又觉得头痛,其实没关系,前面有许多行无非是设置log写入程序,打开conf读取ip,port等信息,最重要的内容就是代码最后几行中的
self.pool = self.create_pool()
eventlet.wsgi.server(self.sock,self.application,log=WritableLogger(self.logger),custom_pool=self.pool)
刚才提到了server和application之间的通讯接口WSGI,现在我们要讲讲server。OpenStack并没有使用Python标准库中的BaseHTTPServer,而是使用了在网络并发等领域处理效率非常优异的eventlet库http://eventlet.net/。
eventlet提供了一套API以实现“协程”(coroutines)。所谓的“协程”可以简单的看做是“假线程”,他可以实现线程的非阻塞异步IO调用的功能,但是协程没有独立的堆栈,这和线程有自己独立的堆栈是有区别的。eventlet会维护一个协程“池”,用来存放所有创建的协程。但是不同于线程,协程同时只能有一个实例在运行,其他的协程要运行,必须等待当前协程显式的被挂起。不同于线程的执行顺序随机,协程的执行时按调用顺序的。
OpenStack的服务,没用使用市面上常见的web server的原因大概就是其处理并发无非就是使用多线程或IO复用等。然而,当多客户端并发访问时,OpenStack内部的一些共享资源,并不能十分安全的利用互斥锁等方法进行线程共享资源的互斥。为了防止并发出现资源死锁,简化架构设计流程,采用“协程”是个非常不错的选择。并且,线程间的切换需要大量的时间和空间的开销,而协程可以有效的避免这个问题。
import eventlet
def handle(client):
while True:
c = client.recv(1)
if not c: break
client.sendall(c)
server = eventlet.listen(('0.0.0.0', 6000))
pool = eventlet.GreenPool(10000)
while True:
new_sock, address = server.accept()
pool.spawn_n(handle, new_sock)
上面是eventlet一个简单服务器端的示例,首先用eventlet.GreenPool(1000)生成一个最大容量为1000的“协程”缓冲池,server.accept()等待,服务器server端收到一个客户端的连接请求,就用pool.spawn_n()启动一个“协程”进行处理响应。
回到glance中,OpenStack将Python原版的CGIHTTPServer进行“绿化”,提供了eventlet.wsgi.server进行http的响应,其内部实现结构和作用和上面的代码相似,同样都是来一个http请求,就会启动一个协程server进行响应。参数custom_pool就是我们上面刚刚申请的GreenPool协程池。参数self.application为WSGI程序入口。这样我们就成功运行了一个WSGI服务程序。
本章结束,我们已经成功运行启动了Glance的WSGI服务,下一章将会开始具体介绍WSGI程序所使用的请求分发组件Routes和request与response的包装类webob的相关内容。
本章提到了些高并发访问处理方面的相关内容,实际上eventlet协程的设置上也使用了eventlet.hubs.use_hub('poll'),欢迎大家继续展开阅读与讨论Select I/O poll epoll等相关内容。
页:
[1]