From: Mathias Sundman Subject: Re: Interface with GUI agent, was: Re: [Openvpn-devel] [Patch] revoke scripts were broken In-Reply-To: <200407031729.30190.jim@...41...> Message-ID: References: <40E18559.8020407@...244...> <200407031601.24438.jim@...41...> <200407031729.30190.jim@...41...> MIME-Version: 1.0 Sender: openvpn-devel-admin@lists.sourceforge.net Errors-To: openvpn-devel-admin@lists.sourceforge.net List-Unsubscribe: , List-Id: List-Post: List-Help: List-Subscribe: , List-Archive: Date: Sat Jul 3 14:44:00 2004 Content-Type: TEXT/PLAIN; charset=US-ASCII; format=flowed Content-Transfer-Encoding: 8bit To: James Yonan Cc: Brandon Knitter , openvpn-devel@lists.sourceforge.net, Jan Kiszka On Sat, 3 Jul 2004, James Yonan wrote: >>> There are no hooks as of yet to control a running OpenVPN instance (other >>> than sending signals). >>> >>> While I initially thought that named pipes might be the appropriate >>> channel for this, I'm wondering if if might make more sense (for >>> portability) to use a local TCP/IP socket as the control channel, such as >>> 127.0.0.0:[some port]. >>> >>> Then for testing purposes, you could just telnet to the socket. >>> >>> I am thinking the interface should consist of a standard ascii command >>> response loop, such as is used by SMTP. >>> >>> Possible client -> server commands: >>> >>> (1) send auth credentials >>> (2) get status (as in SIGUSR2) >>> (3) send signal (SIGHUP, SIGUSR1, or SIGUSR2, SIGTERM) >>> >>> And server -> client commands: >>> >>> (1) need auth credentials -- GUI should query the user then return them >>> to the daemon via a "send auth credentials" command >>> >>> The basic mechanism of operation to enable the management interface would >>> be to pass an option like this in the config file: >>> >>> management 127.0.0.1 20001 >>> >>> This will cause OpenVPN to listen on 127.0.0.1:20001 as its management >>> interface port. >>> >> One little wonder though, is it good to have a bi-directional protocol >> over a single socket. Ofcource replys to command can go over the same >> channel, but if both the server and the client is allowed to "start the >> conversation", how do we handle the case when both the server and client >> sends a command at the same time. > > TCP connections are always full duplex so both sides can send and receive at > the same time without collision issues. > > Having said that, the application needs to support bidirectionality, by always > having a listen queued on the port for possible incoming data. > > On Windows, the best way to do this is to use the WSAAsyncSelect or > WSAEventSelect functions. WSAAsyncSelect is particularly useful for Windows > GUI-based programs because it gives you notification of incoming data on a > TCP socket through the standard Windows message notification mechanism. > >> How will we know if what we recieve is an answer to our command, or a >> command from the server? > > The unsolicited notification from the server would have a specific prefix that > indicates that it is an asynchronous notification from the server, as opposed > to a response to a client->server command. > > For example, the server would say something like this: > > NOTIFY:NEED_AUTH,USER,PASS > > The "ASYNC" indicates that the message is an asynchronous notification. > > NEED_AUTH is the command, and USER/PASS indicates that the GUI should prompt > the user for and return both username and password. > > Then the GUI would reply back: > > AUTH_USER [username] > AUTH_PASS [password] Okay. The NOTIFY prefix solves my wonder. I was aware that TCP connections are fullduplex, it was just the overlying protocol I was worried about. >> In addition to the information given by SIGUSR2 I'd would like to be able >> to get current state of the the link. We should be able to define several >> STATEs during the establishing of the connection. Like: >> >> STATE_WAIT_FIRST_RESPONE - When we have tried to connect to the remote >> peer but still not recieved any answer. >> >> STATE_WAIT_AUTH - We have recieved some answer from the remove peer, but >> not yet fully authenticated. (This state will never happend when using >> static keys.) >> >> STATE_WAIT_PUSH - We wait for options to be pushed from the server. >> >> STATE_WAIT_SETIP - We are waiting for the system to set the ip address. >> >> STATE_WAIT_APPLY_SETTINGS - We are waiting for the system to set other >> pushed parameters like routes. >> >> STATE_CONNECTED - We are fully connected! > > We already have something like this in ssl.h (see "key negotiation states"), > but your idea expands beyond this. > > I think the states are a good idea. The state info, like the request for auth > credentials, will be an asynchronous server -> client notification. > > For example: > > NOTIFY:STATE,STATE_CONNECTED I think it would be good both if the server sent NOTIFY messages when the state is change, and if there was a query command the client can issue to receive the current state. One use of the latter is that it would allow the gui to hook into an already running process (started by the service wrapper perhaps), and check the current state. -- _____________________________________________________________ Mathias Sundman (^) ASCII Ribbon Campaign NILINGS AB X NO HTML/RTF in e-mail Tel: +46-(0)8-666 32 28 / \ NO Word docs in e-mail