1Description of the "concap" encapsulation protocol interface
2============================================================
3
4The "concap" interface is intended to be used by network device
5drivers that need to process an encapsulation protocol.
6It is assumed that the protocol interacts with a linux network device by
7- data transmission
8- connection control (establish, release)
9Thus, the mnemonic: "CONnection CONtrolling eNCAPsulation Protocol".
10
11This is currently only used inside the isdn subsystem. But it might
12also be useful to other kinds of network devices. Thus, if you want
13to suggest changes that improve usability or performance of the
14interface, please let me know. I'm willing to include them in future
15releases (even if I needed to adapt the current isdn code to the
16changed interface).
17
18
19Why is this useful?
20===================
21
22The encapsulation protocol used on top of WAN connections or permanent
23point-to-point links are frequently chosen upon bilateral agreement.
24Thus, a device driver for a certain type of hardware must support
25several different encapsulation protocols at once.
26
27The isdn device driver did already support several different
28encapsulation protocols. The encapsulation protocol is configured by a
29user space utility (isdnctrl). The isdn network interface code then
30uses several case statements which select appropriate actions
31depending on the currently configured encapsulation protocol.
32
33In contrast, LAN network interfaces always used a single encapsulation
34protocol which is unique to the hardware type of the interface. The LAN
35encapsulation is usually done by just sticking a header on the data. Thus,
36traditional linux network device drivers used to process the
37encapsulation protocol directly (usually by just providing a hard_header()
38method in the device structure) using some hardware type specific support
39functions. This is simple, direct and efficient. But it doesn't fit all
40the requirements for complex WAN encapsulations.
41
42
43   The configurability of the encapsulation protocol to be used
44   makes isdn network interfaces more flexible, but also much more
45   complex than traditional lan network interfaces.
46
47
48Many Encapsulation protocols used on top of WAN connections will not just
49stick a header on the data. They also might need to set up or release
50the WAN connection. They also might want to send other data for their
51private purpose over the wire, e.g. ppp does a lot of link level
52negotiation before the first piece of user data can be transmitted.
53Such encapsulation protocols for WAN devices are typically more complex
54than encapsulation protocols for lan devices. Thus, network interface
55code for typical WAN devices also tends to be more complex.
56
57
58In order to support Linux' x25 PLP implementation on top of
59isdn network interfaces I could have introduced yet another branch to
60the various case statements inside drivers/isdn/isdn_net.c.
61This eventually made isdn_net.c even more complex. In addition, it made
62isdn_net.c harder to maintain. Thus, by identifying an abstract
63interface between the network interface code and the encapsulation
64protocol, complexity could be reduced and maintainability could be
65increased.
66
67
68Likewise, a similar encapsulation protocol will frequently be needed by
69several different interfaces of even different hardware type, e.g. the
70synchronous ppp implementation used by the isdn driver and the
71asynchronous ppp implementation used by the ppp driver have a lot of
72similar code in them. By cleanly separating the encapsulation protocol
73from the hardware specific interface stuff such code could be shared
74better in future.
75
76
77When operating over dial-up-connections (e.g. telephone lines via modem,
78non-permanent virtual circuits of wide area networks, ISDN) many
79encapsulation protocols will need to control the connection. Therefore,
80some basic connection control primitives are supported. The type and
81semantics of the connection (i.e the ISO layer where connection service
82is provided) is outside our scope and might be different depending on
83the encapsulation protocol used, e.g. for a ppp module using our service
84on top of a modem connection a connect_request will result in dialing
85a (somewhere else configured) remote phone number. For an X25-interface
86module (LAPB semantics, as defined in Documentation/networking/x25-iface.txt)
87a connect_request will ask for establishing a reliable lapb
88datalink connection.
89
90
91The encapsulation protocol currently provides the following
92service primitives to the network device.
93
94- create a new encapsulation protocol instance
95- delete encapsulation protocol instance and free all its resources
96- initialize (open) the encapsulation protocol instance for use.
97- deactivate (close) an encapsulation protocol instance.
98- process (xmit) data handed down by upper protocol layer
99- receive data from lower (hardware) layer
100- process connect indication from lower (hardware) layer
101- process disconnect indication from lower (hardware) layer
102
103
104The network interface driver accesses those primitives via callbacks
105provided by the encapsulation protocol instance within a
106struct concap_proto_ops.
107
108struct concap_proto_ops{
109
110	/* create a new encapsulation protocol instance of same type */
111	struct concap_proto *  (*proto_new) (void);
112
113	/* delete encapsulation protocol instance and free all its resources.
114	   cprot may no loger be referenced after calling this */
115	void (*proto_del)(struct concap_proto *cprot);
116
117	/* initialize the protocol's data. To be called at interface startup
118	   or when the device driver resets the interface. All services of the
119	   encapsulation protocol may be used after this*/
120	int (*restart)(struct concap_proto *cprot,
121		       struct net_device *ndev,
122		       struct concap_device_ops *dops);
123
124	/* deactivate an encapsulation protocol instance. The encapsulation
125	   protocol may not call any *dops methods after this. */
126	int (*close)(struct concap_proto *cprot);
127
128	/* process a frame handed down to us by upper layer */
129	int (*encap_and_xmit)(struct concap_proto *cprot, struct sk_buff *skb);
130
131	/* to be called for each data entity received from lower layer*/
132	int (*data_ind)(struct concap_proto *cprot, struct sk_buff *skb);
133
134	/* to be called when a connection was set up/down.
135	   Protocols that don't process these primitives might fill in
136	   dummy methods here */
137	int (*connect_ind)(struct concap_proto *cprot);
138	int (*disconn_ind)(struct concap_proto *cprot);
139};
140
141
142The data structures are defined in the header file include/linux/concap.h.
143
144
145A Network interface using encapsulation protocols must also provide
146some service primitives to the encapsulation protocol:
147
148- request data being submitted by lower layer (device hardware)
149- request a connection being set up by lower layer
150- request a connection being released by lower layer
151
152The encapsulation protocol accesses those primitives via callbacks
153provided by the network interface within a struct concap_device_ops.
154
155struct concap_device_ops{
156
157	/* to request data be submitted by device */
158	int (*data_req)(struct concap_proto *, struct sk_buff *);
159
160	/* Control methods must be set to NULL by devices which do not
161	   support connection control. */
162	/* to request a connection be set up */
163	int (*connect_req)(struct concap_proto *);
164
165	/* to request a connection be released */
166	int (*disconn_req)(struct concap_proto *);
167};
168
169The network interface does not explicitly provide a receive service
170because the encapsulation protocol directly calls netif_rx().
171
172
173
174
175An encapsulation protocol itself is actually the
176struct concap_proto{
177	struct net_device *net_dev;		/* net device using our service  */
178	struct concap_device_ops *dops; /* callbacks provided by device */
179 	struct concap_proto_ops  *pops; /* callbacks provided by us */
180	int flags;
181	void *proto_data;               /* protocol specific private data, to
182					   be accessed via *pops methods only*/
183	/*
184	  :
185	  whatever
186	  :
187	  */
188};
189
190Most of this is filled in when the device requests the protocol to
191be reset (opend). The network interface must provide the net_dev and
192dops pointers. Other concap_proto members should be considered private
193data that are only accessed by the pops callback functions. Likewise,
194a concap proto should access the network device's private data
195only by means of the callbacks referred to by the dops pointer.
196
197
198A possible extended device structure which uses the connection controlling
199encapsulation services could look like this:
200
201struct concap_device{
202	struct net_device net_dev;
203	struct my_priv  /* device->local stuff */
204			/* the my_priv struct might contain a
205			   struct concap_device_ops *dops;
206	                   to provide the device specific callbacks
207			*/
208	struct concap_proto *cprot;        /* callbacks provided by protocol */
209};
210
211
212
213Misc Thoughts
214=============
215
216The concept of the concap proto might help to reuse protocol code and
217reduce the complexity of certain network interface implementations.
218The trade off is that it introduces yet another procedure call layer
219when processing the protocol. This has of course some impact on
220performance. However, typically the concap interface will be used by
221devices attached to slow lines (like telephone, isdn, leased synchronous
222lines). For such slow lines, the overhead is probably negligible.
223This might no longer hold for certain high speed WAN links (like
224ATM).
225
226
227If general linux network interfaces explicitly supported concap
228protocols (e.g. by a member struct concap_proto* in struct net_device)
229then the interface of the service function could be changed
230by passing a pointer of type (struct net_device*) instead of
231type (struct concap_proto*). Doing so would make many of the service
232functions compatible to network device support functions.
233
234e.g. instead of the concap protocol's service function
235
236  int (*encap_and_xmit)(struct concap_proto *cprot, struct sk_buff *skb);
237
238we could have
239
240  int (*encap_and_xmit)(struct net_device *ndev, struct sk_buff *skb);
241
242As this is compatible to the dev->hard_start_xmit() method, the device
243driver could directly register the concap protocol's encap_and_xmit()
244function as its hard_start_xmit() method. This would eliminate one
245procedure call layer.
246
247
248The device's data request function could also be defined as
249
250  int (*data_req)(struct net_device *ndev, struct sk_buff *skb);
251
252This might even allow for some protocol stacking. And the network
253interface might even register the same data_req() function directly
254as its hard_start_xmit() method when a zero layer encapsulation
255protocol is configured. Thus, eliminating the performance penalty
256of the concap interface when a trivial concap protocol is used.
257Nevertheless, the device remains able to support encapsulation
258protocol configuration.
259
260