2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
17 #ifndef DPIF_PROVIDER_H
18 #define DPIF_PROVIDER_H 1
20 /* Provider interface to dpifs, which provide an interface to an Open vSwitch
21 * datapath. A datapath is a collection of physical or virtual ports that are
22 * exposed over OpenFlow as a single switch. Datapaths and the collections of
23 * ports that they contain may be fixed or dynamic. */
25 #include "openflow/openflow.h"
33 /* Open vSwitch datapath interface.
35 * This structure should be treated as opaque by dpif implementations. */
37 const struct dpif_class *dpif_class;
40 uint8_t netflow_engine_type;
41 uint8_t netflow_engine_id;
44 void dpif_init(struct dpif *, const struct dpif_class *, const char *name,
45 uint8_t netflow_engine_type, uint8_t netflow_engine_id);
46 void dpif_uninit(struct dpif *dpif, bool close);
48 static inline void dpif_assert_class(const struct dpif *dpif,
49 const struct dpif_class *dpif_class)
51 ovs_assert(dpif->dpif_class == dpif_class);
54 struct dpif_flow_dump {
59 dpif_flow_dump_init(struct dpif_flow_dump *dump, const struct dpif *dpif)
61 dump->dpif = CONST_CAST(struct dpif *, dpif);
64 struct dpif_flow_dump_thread {
69 dpif_flow_dump_thread_init(struct dpif_flow_dump_thread *thread,
70 struct dpif_flow_dump *dump)
72 thread->dpif = dump->dpif;
75 /* Datapath interface class structure, to be defined by each implementation of
76 * a datapath interface.
78 * These functions return 0 if successful or a positive errno value on failure,
79 * except where otherwise noted.
81 * These functions are expected to execute synchronously, that is, to block as
82 * necessary to obtain a result. Thus, they may not return EAGAIN or
83 * EWOULDBLOCK or EINPROGRESS. We may relax this requirement in the future if
84 * and when we encounter performance problems. */
86 /* Type of dpif in this class, e.g. "system", "netdev", etc.
88 * One of the providers should supply a "system" type, since this is
89 * the type assumed if no type is specified when opening a dpif. */
92 /* Enumerates the names of all known created datapaths (of class
93 * 'dpif_class'), if possible, into 'all_dps'. The caller has already
94 * initialized 'all_dps' and other dpif classes might already have added
97 * This is used by the vswitch at startup, so that it can delete any
98 * datapaths that are not configured.
100 * Some kinds of datapaths might not be practically enumerable, in which
101 * case this function may be a null pointer. */
102 int (*enumerate)(struct sset *all_dps, const struct dpif_class *dpif_class);
104 /* Returns the type to pass to netdev_open() when a dpif of class
105 * 'dpif_class' has a port of type 'type', for a few special cases
106 * when a netdev type differs from a port type. For example, when
107 * using the userspace datapath, a port of type "internal" needs to
108 * be opened as "tap".
110 * Returns either 'type' itself or a string literal, which must not
112 const char *(*port_open_type)(const struct dpif_class *dpif_class,
115 /* Attempts to open an existing dpif called 'name', if 'create' is false,
116 * or to open an existing dpif or create a new one, if 'create' is true.
118 * 'dpif_class' is the class of dpif to open.
120 * If successful, stores a pointer to the new dpif in '*dpifp', which must
121 * have class 'dpif_class'. On failure there are no requirements on what
122 * is stored in '*dpifp'. */
123 int (*open)(const struct dpif_class *dpif_class,
124 const char *name, bool create, struct dpif **dpifp);
126 /* Closes 'dpif' and frees associated memory. */
127 void (*close)(struct dpif *dpif);
129 /* Attempts to destroy the dpif underlying 'dpif'.
131 * If successful, 'dpif' will not be used again except as an argument for
132 * the 'close' member function. */
133 int (*destroy)(struct dpif *dpif);
135 /* Performs periodic work needed by 'dpif', if any is necessary. */
136 void (*run)(struct dpif *dpif);
138 /* Arranges for poll_block() to wake up if the "run" member function needs
139 * to be called for 'dpif'. */
140 void (*wait)(struct dpif *dpif);
142 /* Retrieves statistics for 'dpif' into 'stats'. */
143 int (*get_stats)(const struct dpif *dpif, struct dpif_dp_stats *stats);
145 /* Adds 'netdev' as a new port in 'dpif'. If '*port_no' is not
146 * UINT32_MAX, attempts to use that as the port's port number.
148 * If port is successfully added, sets '*port_no' to the new port's
149 * port number. Returns EBUSY if caller attempted to choose a port
150 * number, and it was in use. */
151 int (*port_add)(struct dpif *dpif, struct netdev *netdev,
152 odp_port_t *port_no);
154 /* Removes port numbered 'port_no' from 'dpif'. */
155 int (*port_del)(struct dpif *dpif, odp_port_t port_no);
157 /* Queries 'dpif' for a port with the given 'port_no' or 'devname'.
158 * If 'port' is not null, stores information about the port into
159 * '*port' if successful.
161 * If 'port' is not null, the caller takes ownership of data in
162 * 'port' and must free it with dpif_port_destroy() when it is no
164 int (*port_query_by_number)(const struct dpif *dpif, odp_port_t port_no,
165 struct dpif_port *port);
166 int (*port_query_by_name)(const struct dpif *dpif, const char *devname,
167 struct dpif_port *port);
169 /* Returns the Netlink PID value to supply in OVS_ACTION_ATTR_USERSPACE
170 * actions as the OVS_USERSPACE_ATTR_PID attribute's value, for use in
171 * flows whose packets arrived on port 'port_no'. In the case where the
172 * provider allocates multiple Netlink PIDs to a single port, it may use
173 * 'hash' to spread load among them. The caller need not use a particular
174 * hash function; a 5-tuple hash is suitable.
176 * (The datapath implementation might use some different hash function for
177 * distributing packets received via flow misses among PIDs. This means
178 * that packets received via flow misses might be reordered relative to
179 * packets received via userspace actions. This is not ordinarily a
182 * A 'port_no' of UINT32_MAX should be treated as a special case. The
183 * implementation should return a reserved PID, not allocated to any port,
184 * that the client may use for special purposes.
186 * The return value only needs to be meaningful when DPIF_UC_ACTION has
187 * been enabled in the 'dpif''s listen mask, and it is allowed to change
188 * when DPIF_UC_ACTION is disabled and then re-enabled.
190 * A dpif provider that doesn't have meaningful Netlink PIDs can use NULL
191 * for this function. This is equivalent to always returning 0. */
192 uint32_t (*port_get_pid)(const struct dpif *dpif, odp_port_t port_no,
195 /* Attempts to begin dumping the ports in a dpif. On success, returns 0
196 * and initializes '*statep' with any data needed for iteration. On
197 * failure, returns a positive errno value. */
198 int (*port_dump_start)(const struct dpif *dpif, void **statep);
200 /* Attempts to retrieve another port from 'dpif' for 'state', which was
201 * initialized by a successful call to the 'port_dump_start' function for
202 * 'dpif'. On success, stores a new dpif_port into 'port' and returns 0.
203 * Returns EOF if the end of the port table has been reached, or a positive
204 * errno value on error. This function will not be called again once it
205 * returns nonzero once for a given iteration (but the 'port_dump_done'
206 * function will be called afterward).
208 * The dpif provider retains ownership of the data stored in 'port'. It
209 * must remain valid until at least the next call to 'port_dump_next' or
210 * 'port_dump_done' for 'state'. */
211 int (*port_dump_next)(const struct dpif *dpif, void *state,
212 struct dpif_port *port);
214 /* Releases resources from 'dpif' for 'state', which was initialized by a
215 * successful call to the 'port_dump_start' function for 'dpif'. */
216 int (*port_dump_done)(const struct dpif *dpif, void *state);
218 /* Polls for changes in the set of ports in 'dpif'. If the set of ports in
219 * 'dpif' has changed, then this function should do one of the
222 * - Preferably: store the name of the device that was added to or deleted
223 * from 'dpif' in '*devnamep' and return 0. The caller is responsible
224 * for freeing '*devnamep' (with free()) when it no longer needs it.
226 * - Alternatively: return ENOBUFS, without indicating the device that was
229 * Occasional 'false positives', in which the function returns 0 while
230 * indicating a device that was not actually added or deleted or returns
231 * ENOBUFS without any change, are acceptable.
233 * If the set of ports in 'dpif' has not changed, returns EAGAIN. May also
234 * return other positive errno values to indicate that something has gone
236 int (*port_poll)(const struct dpif *dpif, char **devnamep);
238 /* Arranges for the poll loop to wake up when 'port_poll' will return a
239 * value other than EAGAIN. */
240 void (*port_poll_wait)(const struct dpif *dpif);
242 /* Queries 'dpif' for a flow entry. The flow is specified by the Netlink
243 * attributes with types OVS_KEY_ATTR_* in the 'key_len' bytes starting at
246 * Returns 0 if successful. If no flow matches, returns ENOENT. On other
247 * failure, returns a positive errno value.
249 * On success, '*bufp' will be set to an ofpbuf owned by the caller that
250 * contains the response for 'maskp' and/or 'actionsp'. The caller must
251 * supply a valid pointer, and must free the ofpbuf (with ofpbuf_delete())
252 * when it is no longer needed.
254 * If 'maskp' is nonnull, then on success '*maskp' will point to the
255 * Netlink attributes for the flow's mask. '*mask_len' will be set to the
256 * length of the mask attributes. Implementations may opt to point 'maskp'
257 * at RCU-protected data rather than making a copy in '*bufp'.
259 * If 'actionsp' is nonnull, then on success '*actionsp' will point to the
260 * Netlink attributes for the flow's actions. '*actions_len' will be set to
261 * the length of the actions attributes. Implementations may opt to point
262 * 'actionsp' at RCU-protected data rather than making a copy in '*bufp'.
264 * If 'stats' is nonnull, then on success it must be updated with the
265 * flow's statistics. */
266 int (*flow_get)(const struct dpif *dpif,
267 const struct nlattr *key, size_t key_len,
268 struct ofpbuf **bufp,
269 struct nlattr **maskp, size_t *mask_len,
270 struct nlattr **actionsp, size_t *acts_len,
271 struct dpif_flow_stats *stats);
273 /* Adds or modifies a flow in 'dpif'. The flow is specified by the Netlink
274 * attributes with types OVS_KEY_ATTR_* in the 'put->key_len' bytes
275 * starting at 'put->key'. The associated actions are specified by the
276 * Netlink attributes with types OVS_ACTION_ATTR_* in the
277 * 'put->actions_len' bytes starting at 'put->actions'.
279 * - If the flow's key does not exist in 'dpif', then the flow will be
280 * added if 'put->flags' includes DPIF_FP_CREATE. Otherwise the
281 * operation will fail with ENOENT.
283 * If the operation succeeds, then 'put->stats', if nonnull, must be
286 * - If the flow's key does exist in 'dpif', then the flow's actions will
287 * be updated if 'put->flags' includes DPIF_FP_MODIFY. Otherwise the
288 * operation will fail with EEXIST. If the flow's actions are updated,
289 * then its statistics will be zeroed if 'put->flags' includes
290 * DPIF_FP_ZERO_STATS, and left as-is otherwise.
292 * If the operation succeeds, then 'put->stats', if nonnull, must be set
293 * to the flow's statistics before the update.
295 int (*flow_put)(struct dpif *dpif, const struct dpif_flow_put *put);
297 /* Deletes a flow from 'dpif' and returns 0, or returns ENOENT if 'dpif'
298 * does not contain such a flow. The flow is specified by the Netlink
299 * attributes with types OVS_KEY_ATTR_* in the 'del->key_len' bytes
300 * starting at 'del->key'.
302 * If the operation succeeds, then 'del->stats', if nonnull, must be set to
303 * the flow's statistics before its deletion. */
304 int (*flow_del)(struct dpif *dpif, const struct dpif_flow_del *del);
306 /* Deletes all flows from 'dpif' and clears all of its queues of received
308 int (*flow_flush)(struct dpif *dpif);
310 /* Flow dumping interface.
312 * This is the back-end for the flow dumping interface described in
313 * dpif.h. Please read the comments there first, because this code
314 * closely follows it.
316 * 'flow_dump_create' and 'flow_dump_thread_create' must always return an
317 * initialized and usable data structure and defer error return until
318 * flow_dump_destroy(). This hasn't been a problem for the dpifs that
321 * 'flow_dump_create' and 'flow_dump_thread_create' must initialize the
322 * structures that they return with dpif_flow_dump_init() and
323 * dpif_flow_dump_thread_init(), respectively. */
324 struct dpif_flow_dump *(*flow_dump_create)(const struct dpif *dpif);
325 int (*flow_dump_destroy)(struct dpif_flow_dump *dump);
327 struct dpif_flow_dump_thread *(*flow_dump_thread_create)(
328 struct dpif_flow_dump *dump);
329 void (*flow_dump_thread_destroy)(struct dpif_flow_dump_thread *thread);
331 int (*flow_dump_next)(struct dpif_flow_dump_thread *thread,
332 struct dpif_flow *flows, int max_flows);
334 /* Performs the 'execute->actions_len' bytes of actions in
335 * 'execute->actions' on the Ethernet frame in 'execute->packet'
336 * and on the packet metadata in 'execute->md'.
337 * May modify both packet and metadata. */
338 int (*execute)(struct dpif *dpif, struct dpif_execute *execute);
340 /* Executes each of the 'n_ops' operations in 'ops' on 'dpif', in the order
341 * in which they are specified, placing each operation's results in the
342 * "output" members documented in comments.
344 * This function is optional. It is only worthwhile to implement it if
345 * 'dpif' can perform operations in batch faster than individually. */
346 void (*operate)(struct dpif *dpif, struct dpif_op **ops, size_t n_ops);
348 /* Enables or disables receiving packets with dpif_recv() for 'dpif'.
349 * Turning packet receive off and then back on is allowed to change Netlink
350 * PID assignments (see ->port_get_pid()). The client is responsible for
351 * updating flows as necessary if it does this. */
352 int (*recv_set)(struct dpif *dpif, bool enable);
354 /* Refreshes the poll loops and Netlink sockets associated to each port,
355 * when the number of upcall handlers (upcall receiving thread) is changed
356 * to 'n_handlers' and receiving packets for 'dpif' is enabled by
359 * Since multiple upcall handlers can read upcalls simultaneously from
360 * 'dpif', each port can have multiple Netlink sockets, one per upcall
361 * handler. So, handlers_set() is responsible for the following tasks:
363 * When receiving upcall is enabled, extends or creates the
364 * configuration to support:
366 * - 'n_handlers' Netlink sockets for each port.
368 * - 'n_handlers' poll loops, one for each upcall handler.
370 * - registering the Netlink sockets for the same upcall handler to
371 * the corresponding poll loop.
373 int (*handlers_set)(struct dpif *dpif, uint32_t n_handlers);
375 /* Translates OpenFlow queue ID 'queue_id' (in host byte order) into a
376 * priority value used for setting packet priority. */
377 int (*queue_to_priority)(const struct dpif *dpif, uint32_t queue_id,
380 /* Polls for an upcall from 'dpif' for an upcall handler. Since there
381 * can be multiple poll loops (see ->handlers_set()), 'handler_id' is
382 * needed as index to identify the corresponding poll loop. If
383 * successful, stores the upcall into '*upcall', using 'buf' for
384 * storage. Should only be called if 'recv_set' has been used to enable
385 * receiving packets from 'dpif'.
387 * The implementation should point 'upcall->key' and 'upcall->userdata'
388 * (if any) into data in the caller-provided 'buf'. The implementation may
389 * also use 'buf' for storing the data of 'upcall->packet'. If necessary
390 * to make room, the implementation may reallocate the data in 'buf'.
392 * The caller owns the data of 'upcall->packet' and may modify it. If
393 * packet's headroom is exhausted as it is manipulated, 'upcall->packet'
394 * will be reallocated. This requires the data of 'upcall->packet' to be
395 * released with ofpbuf_uninit() before 'upcall' is destroyed. However,
396 * when an error is returned, the 'upcall->packet' may be uninitialized
397 * and should not be released.
399 * This function must not block. If no upcall is pending when it is
400 * called, it should return EAGAIN without blocking. */
401 int (*recv)(struct dpif *dpif, uint32_t handler_id,
402 struct dpif_upcall *upcall, struct ofpbuf *buf);
404 /* Arranges for the poll loop for an upcall handler to wake up when 'dpif'
405 * has a message queued to be received with the recv member functions.
406 * Since there can be multiple poll loops (see ->handlers_set()),
407 * 'handler_id' is needed as index to identify the corresponding poll loop.
409 void (*recv_wait)(struct dpif *dpif, uint32_t handler_id);
411 /* Throws away any queued upcalls that 'dpif' currently has ready to
413 void (*recv_purge)(struct dpif *dpif);
416 extern const struct dpif_class dpif_linux_class;
417 extern const struct dpif_class dpif_netdev_class;
423 #endif /* dpif-provider.h */