gevent.pywsgi – A pure-Python, gevent-friendly WSGI server

The server is provided in WSGIServer, but most of the actual WSGI work is handled by WSGIHandler — a new instance is created for each request. The server can be customized to use different subclasses of WSGIHandler.

class WSGIServer(listener, application=None, backlog=None, spawn='default', log='default', error_log='default', handler_class=None, environ=None, **ssl_args)

Bases: gevent.server.StreamServer

A WSGI server based on StreamServer that supports HTTPS.

Parameters:
  • log – If given, an object with a write method to which request (access) logs will be written. If not given, defaults to sys.stderr. You may pass None to disable request logging. You may use a wrapper, around e.g., logging, to support objects that don’t implement a write method. (If you pass a Logger instance, or in general something that provides a log method but not a write method, such a wrapper will automatically be created and it will be logged to at the INFO level.)
  • error_log – If given, a file-like object with write, writelines and flush methods to which error logs will be written. If not given, defaults to sys.stderr. You may pass None to disable error logging (not recommended). You may use a wrapper, around e.g., logging, to support objects that don’t implement the proper methods. This parameter will become the value for wsgi.errors in the WSGI environment (if not already set). (As with log, wrappers for Logger instances and the like will be created automatically and logged to at the ERROR level.)

See also

LoggingLogAdapter
See important warnings before attempting to use logging.

Changed in version 1.1a3: Added the error_log parameter, and set wsgi.errors in the WSGI environment to this value.

Changed in version 1.1a3: Add support for passing logging.Logger objects to the log and error_log arguments.

environ_class

The class of environ objects passed to the handlers. Must be a dict subclass. For compliance with PEP 3333 and libraries like WebOb, this is simply dict but this can be customized in a subclass or per-instance (probably to WSGISecureEnviron).

New in version 1.2a1.

alias of dict

error_log = None

The object to which error logs will be written. It must never be None. Initialized from the error_log constructor parameter.

handle(sock, address)

Create an instance of handler_class to handle the request.

This method blocks until the handler returns.

handler_class

A callable taking three arguments: (socket, address, server) and returning an object with a handle() method. The callable is called once for each incoming socket request, as is its handle method. The handle method should not return until all use of the socket is complete.

This class uses the WSGIHandler object as the default value. You may subclass this class and set a different default value, or you may pass a value to use in the handler_class keyword constructor argument.

alias of WSGIHandler

log = None

The object to which request logs will be written. It must never be None. Initialized from the log constructor parameter.

secure_environ_class

alias of WSGISecureEnviron

update_environ()

Called before the first request is handled to fill in WSGI environment values.

This includes getting the correct server name and port.

class WSGIHandler(sock, address, server, rfile=None)

Bases: object

Handles HTTP requests from a socket, creates the WSGI environment, and interacts with the WSGI application.

This is the default value of WSGIServer.handler_class. This class may be subclassed carefully, and that class set on a WSGIServer instance through a keyword argument at construction time.

Instances are constructed with the same arguments as passed to the server’s WSGIServer.handle() method followed by the server itself. The application and environment are obtained from the server.

get_environ()

Construct and return a new WSGI environment dictionary for a specific request.

This should begin with asking the server for the base environment using WSGIServer.get_environ(), and then proceed to add the request specific values.

By the time this method is invoked the request line and request shall have been parsed and self.headers shall be populated.

handle()

The main request handling method, called by the server.

This method runs a request handling loop, calling handle_one_request() until all requests on the connection have been handled (that is, it implements keep-alive).

handle_one_request()

Handles one HTTP request using self.socket and self.rfile.

Each invocation of this method will do several things, including (but not limited to):

  • Read the request line using read_requestline();
  • Read the rest of the request, including headers, with read_request();
  • Construct a new WSGI environment in self.environ using get_environ();
  • Store the application in self.application, retrieving it from the server;
  • Handle the remainder of the request, including invoking the application, with handle_one_response()

There are several possible return values to indicate the state of the client connection:

  • None
    The client connection is already closed or should be closed because the WSGI application or client set the Connection: close header. The request handling loop should terminate and perform cleanup steps.
  • (status, body)
    An HTTP status and body tuple. The request was in error, as detailed by the status and body. The request handling loop should terminate, close the connection, and perform cleanup steps. Note that the body is the complete contents to send to the client, including all headers and the initial status line.
  • True
    The literal True value. The request was successfully handled and the response sent to the client by handle_one_response(). The connection remains open to process more requests and the connection handling loop should call this method again. This is the typical return value.

See also

handle()

Changed in version 1.1b6: Funnel exceptions having to do with invalid HTTP requests through _handle_client_error() to allow subclasses to customize. Note that this is experimental and may change in the future.

handle_one_response()

Invoke the application to produce one response.

This is called by handle_one_request() after all the state for the request has been established. It is responsible for error handling.

ignored_socket_errors = (32, 104)

These errors are silently ignored by handle_one_response() to avoid producing excess log entries on normal operating conditions. They indicate a remote client has disconnected and there is little or nothing this process can be expected to do about it. You may change this value in a subclass.

The default value includes errno.EPIPE and errno.ECONNRESET. On Windows this also includes errno.WSAECONNABORTED.

This is a provisional API, subject to change. See pull request #377, pull request #999 and issue #136.

New in version 1.3.

read_request(raw_requestline)

Parse the incoming request.

Parses various headers into self.headers using MessageClass. Other attributes that are set upon a successful return of this method include self.content_length and self.close_connection.

Parameters:raw_requestline (str) – A native str representing the request line. A processed version of this will be stored into self.requestline.
Raises:ValueError – If the request is invalid. This error will not be logged as a traceback (because it’s a client issue, not a server problem).
Returns:A boolean value indicating whether the request was successfully parsed. This method should either return a true value or have raised a ValueError with details about the parsing error.

Changed in version 1.1b6: Raise the previously documented ValueError in more cases instead of returning a false value; this allows subclasses more opportunity to customize behaviour.

read_requestline()

Read and return the HTTP request line.

Under both Python 2 and 3, this should return the native str type; under Python 3, this probably means the bytes read from the network need to be decoded (using the ISO-8859-1 charset, aka latin-1).

start_response(status, headers, exc_info=None)

Changed in version 1.2a1: Avoid HTTP header injection by raising a ValueError if status or any header name or value contains a carriage return or newline.

Changed in version 1.1b5: Pro-actively handle checking the encoding of the status line and headers during this method. On Python 2, avoid some extra encodings.

class LoggingLogAdapter(logger, level=20)

Bases: object

An adapter for logging.Logger instances to let them be used with WSGIServer.

Warning

Unless the entire process is monkey-patched at a very early part of the lifecycle (before logging is configured), loggers are likely to not be gevent-cooperative. For example, the socket and syslog handlers use the socket module in a way that can block, and most handlers acquire threading locks.

Warning

It may be possible for the logging functions to be called in the gevent.Hub greenlet. Code running in the hub greenlet cannot use any gevent blocking functions without triggering a LoopExit.

New in version 1.1a3.

Changed in version 1.1b6: Attributes not present on this object are proxied to the underlying logger instance. This permits using custom Logger subclasses (or indeed, even duck-typed objects).

Changed in version 1.1: Strip trailing newline characters on the message passed to write() because log handlers will usually add one themselves.

Write information to the logger at the given level (default to INFO).

flush()

No-op; required to be a file-like object

class Environ

Bases: dict

A base class that can be used for WSGI environment objects.

Provisional API.

New in version 1.2a1.

class SecureEnviron

Bases: gevent.pywsgi.Environ

An environment that does not print its keys and values by default.

Provisional API.

This is intended to keep potentially sensitive information like HTTP authorization and cookies from being inadvertently printed or logged.

For debugging, each instance can have its secure_repr attribute set to False, which will cause it to print like a normal dict.

When secure_repr is True (the default), then the value of the whitelist_keys attribute is consulted; if this value is true-ish, it should be a container (something that responds to in) of key names (typically a list or set). Keys and values in this dictionary that are in whitelist_keys will then be printed, while all other values will be masked. These values may be customized on the class by setting the default_secure_repr and default_whitelist_keys, respectively:

>>> environ = SecureEnviron(key='value')
>>> environ 
<pywsgi.SecureEnviron dict (keys: 1) at ...

If we whitelist the key, it gets printed:

>>> environ.whitelist_keys = {'key'}
>>> environ
{'key': 'value'}

A non-whitelisted key (only, to avoid doctest issues) is masked:

>>> environ['secure'] = 'secret'; del environ['key']
>>> environ
{'secure': '<MASKED>'}

We can turn it off entirely for the instance:

>>> environ.secure_repr = False
>>> environ
{'secure': 'secret'}

We can also customize it at the class level (here we use a new class to be explicit and to avoid polluting the true default values; we would set this class to be the environ_class of the server):

>>> class MyEnviron(SecureEnviron):
...    default_whitelist_keys = ('key',)
...
>>> environ = MyEnviron({'key': 'value'})
>>> environ
{'key': 'value'}

New in version 1.2a1.

class WSGISecureEnviron

Bases: gevent.pywsgi.SecureEnviron

Specializes the default list of whitelisted keys to a few common WSGI variables.

Example:

>>> environ = WSGISecureEnviron(REMOTE_ADDR='::1', HTTP_AUTHORIZATION='secret')
>>> environ
{'REMOTE_ADDR': '::1', (hidden keys: 1)}
>>> import pprint
>>> pprint.pprint(environ)
{'REMOTE_ADDR': '::1', (hidden keys: 1)}
>>> print(pprint.pformat(environ))
{'REMOTE_ADDR': '::1', (hidden keys: 1)}

Next page: gevent.wsgi – Backwards compatibility alias for gevent.pywsgi