2 * Copyright (c) 2008, 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.
18 #include "dpif-provider.h"
28 #include "dynamic-string.h"
32 #include "odp-execute.h"
34 #include "ofp-errors.h"
35 #include "ofp-print.h"
38 #include "packet-dpif.h"
40 #include "poll-loop.h"
48 VLOG_DEFINE_THIS_MODULE(dpif);
50 COVERAGE_DEFINE(dpif_destroy);
51 COVERAGE_DEFINE(dpif_port_add);
52 COVERAGE_DEFINE(dpif_port_del);
53 COVERAGE_DEFINE(dpif_flow_flush);
54 COVERAGE_DEFINE(dpif_flow_get);
55 COVERAGE_DEFINE(dpif_flow_put);
56 COVERAGE_DEFINE(dpif_flow_del);
57 COVERAGE_DEFINE(dpif_execute);
58 COVERAGE_DEFINE(dpif_purge);
59 COVERAGE_DEFINE(dpif_execute_with_help);
61 static const struct dpif_class *base_dpif_classes[] = {
68 struct registered_dpif_class {
69 const struct dpif_class *dpif_class;
72 static struct shash dpif_classes = SHASH_INITIALIZER(&dpif_classes);
73 static struct sset dpif_blacklist = SSET_INITIALIZER(&dpif_blacklist);
75 /* Protects 'dpif_classes', including the refcount, and 'dpif_blacklist'. */
76 static struct ovs_mutex dpif_mutex = OVS_MUTEX_INITIALIZER;
78 /* Rate limit for individual messages going to or from the datapath, output at
79 * DBG level. This is very high because, if these are enabled, it is because
80 * we really need to see them. */
81 static struct vlog_rate_limit dpmsg_rl = VLOG_RATE_LIMIT_INIT(600, 600);
83 /* Not really much point in logging many dpif errors. */
84 static struct vlog_rate_limit error_rl = VLOG_RATE_LIMIT_INIT(60, 5);
86 static void log_flow_message(const struct dpif *dpif, int error,
87 const char *operation,
88 const struct nlattr *key, size_t key_len,
89 const struct nlattr *mask, size_t mask_len,
90 const struct dpif_flow_stats *stats,
91 const struct nlattr *actions, size_t actions_len);
92 static void log_operation(const struct dpif *, const char *operation,
94 static bool should_log_flow_message(int error);
95 static void log_flow_put_message(struct dpif *, const struct dpif_flow_put *,
97 static void log_flow_del_message(struct dpif *, const struct dpif_flow_del *,
99 static void log_execute_message(struct dpif *, const struct dpif_execute *,
100 bool subexecute, int error);
105 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
107 if (ovsthread_once_start(&once)) {
110 for (i = 0; i < ARRAY_SIZE(base_dpif_classes); i++) {
111 dp_register_provider(base_dpif_classes[i]);
113 dpctl_unixctl_register();
114 ovsthread_once_done(&once);
119 dp_register_provider__(const struct dpif_class *new_class)
121 struct registered_dpif_class *registered_class;
123 if (sset_contains(&dpif_blacklist, new_class->type)) {
124 VLOG_DBG("attempted to register blacklisted provider: %s",
129 if (shash_find(&dpif_classes, new_class->type)) {
130 VLOG_WARN("attempted to register duplicate datapath provider: %s",
135 registered_class = xmalloc(sizeof *registered_class);
136 registered_class->dpif_class = new_class;
137 registered_class->refcount = 0;
139 shash_add(&dpif_classes, new_class->type, registered_class);
144 /* Registers a new datapath provider. After successful registration, new
145 * datapaths of that type can be opened using dpif_open(). */
147 dp_register_provider(const struct dpif_class *new_class)
151 ovs_mutex_lock(&dpif_mutex);
152 error = dp_register_provider__(new_class);
153 ovs_mutex_unlock(&dpif_mutex);
158 /* Unregisters a datapath provider. 'type' must have been previously
159 * registered and not currently be in use by any dpifs. After unregistration
160 * new datapaths of that type cannot be opened using dpif_open(). */
162 dp_unregister_provider__(const char *type)
164 struct shash_node *node;
165 struct registered_dpif_class *registered_class;
167 node = shash_find(&dpif_classes, type);
169 VLOG_WARN("attempted to unregister a datapath provider that is not "
170 "registered: %s", type);
174 registered_class = node->data;
175 if (registered_class->refcount) {
176 VLOG_WARN("attempted to unregister in use datapath provider: %s", type);
180 shash_delete(&dpif_classes, node);
181 free(registered_class);
186 /* Unregisters a datapath provider. 'type' must have been previously
187 * registered and not currently be in use by any dpifs. After unregistration
188 * new datapaths of that type cannot be opened using dpif_open(). */
190 dp_unregister_provider(const char *type)
196 ovs_mutex_lock(&dpif_mutex);
197 error = dp_unregister_provider__(type);
198 ovs_mutex_unlock(&dpif_mutex);
203 /* Blacklists a provider. Causes future calls of dp_register_provider() with
204 * a dpif_class which implements 'type' to fail. */
206 dp_blacklist_provider(const char *type)
208 ovs_mutex_lock(&dpif_mutex);
209 sset_add(&dpif_blacklist, type);
210 ovs_mutex_unlock(&dpif_mutex);
213 /* Clears 'types' and enumerates the types of all currently registered datapath
214 * providers into it. The caller must first initialize the sset. */
216 dp_enumerate_types(struct sset *types)
218 struct shash_node *node;
223 ovs_mutex_lock(&dpif_mutex);
224 SHASH_FOR_EACH(node, &dpif_classes) {
225 const struct registered_dpif_class *registered_class = node->data;
226 sset_add(types, registered_class->dpif_class->type);
228 ovs_mutex_unlock(&dpif_mutex);
232 dp_class_unref(struct registered_dpif_class *rc)
234 ovs_mutex_lock(&dpif_mutex);
235 ovs_assert(rc->refcount);
237 ovs_mutex_unlock(&dpif_mutex);
240 static struct registered_dpif_class *
241 dp_class_lookup(const char *type)
243 struct registered_dpif_class *rc;
245 ovs_mutex_lock(&dpif_mutex);
246 rc = shash_find_data(&dpif_classes, type);
250 ovs_mutex_unlock(&dpif_mutex);
255 /* Clears 'names' and enumerates the names of all known created datapaths with
256 * the given 'type'. The caller must first initialize the sset. Returns 0 if
257 * successful, otherwise a positive errno value.
259 * Some kinds of datapaths might not be practically enumerable. This is not
260 * considered an error. */
262 dp_enumerate_names(const char *type, struct sset *names)
264 struct registered_dpif_class *registered_class;
265 const struct dpif_class *dpif_class;
271 registered_class = dp_class_lookup(type);
272 if (!registered_class) {
273 VLOG_WARN("could not enumerate unknown type: %s", type);
277 dpif_class = registered_class->dpif_class;
278 error = (dpif_class->enumerate
279 ? dpif_class->enumerate(names, dpif_class)
282 VLOG_WARN("failed to enumerate %s datapaths: %s", dpif_class->type,
283 ovs_strerror(error));
285 dp_class_unref(registered_class);
290 /* Parses 'datapath_name_', which is of the form [type@]name into its
291 * component pieces. 'name' and 'type' must be freed by the caller.
293 * The returned 'type' is normalized, as if by dpif_normalize_type(). */
295 dp_parse_name(const char *datapath_name_, char **name, char **type)
297 char *datapath_name = xstrdup(datapath_name_);
300 separator = strchr(datapath_name, '@');
303 *type = datapath_name;
304 *name = xstrdup(dpif_normalize_type(separator + 1));
306 *name = datapath_name;
307 *type = xstrdup(dpif_normalize_type(NULL));
312 do_open(const char *name, const char *type, bool create, struct dpif **dpifp)
314 struct dpif *dpif = NULL;
316 struct registered_dpif_class *registered_class;
320 type = dpif_normalize_type(type);
321 registered_class = dp_class_lookup(type);
322 if (!registered_class) {
323 VLOG_WARN("could not create datapath %s of unknown type %s", name,
325 error = EAFNOSUPPORT;
329 error = registered_class->dpif_class->open(registered_class->dpif_class,
330 name, create, &dpif);
332 ovs_assert(dpif->dpif_class == registered_class->dpif_class);
334 dp_class_unref(registered_class);
338 *dpifp = error ? NULL : dpif;
342 /* Tries to open an existing datapath named 'name' and type 'type'. Will fail
343 * if no datapath with 'name' and 'type' exists. 'type' may be either NULL or
344 * the empty string to specify the default system type. Returns 0 if
345 * successful, otherwise a positive errno value. On success stores a pointer
346 * to the datapath in '*dpifp', otherwise a null pointer. */
348 dpif_open(const char *name, const char *type, struct dpif **dpifp)
350 return do_open(name, type, false, dpifp);
353 /* Tries to create and open a new datapath with the given 'name' and 'type'.
354 * 'type' may be either NULL or the empty string to specify the default system
355 * type. Will fail if a datapath with 'name' and 'type' already exists.
356 * Returns 0 if successful, otherwise a positive errno value. On success
357 * stores a pointer to the datapath in '*dpifp', otherwise a null pointer. */
359 dpif_create(const char *name, const char *type, struct dpif **dpifp)
361 return do_open(name, type, true, dpifp);
364 /* Tries to open a datapath with the given 'name' and 'type', creating it if it
365 * does not exist. 'type' may be either NULL or the empty string to specify
366 * the default system type. Returns 0 if successful, otherwise a positive
367 * errno value. On success stores a pointer to the datapath in '*dpifp',
368 * otherwise a null pointer. */
370 dpif_create_and_open(const char *name, const char *type, struct dpif **dpifp)
374 error = dpif_create(name, type, dpifp);
375 if (error == EEXIST || error == EBUSY) {
376 error = dpif_open(name, type, dpifp);
378 VLOG_WARN("datapath %s already exists but cannot be opened: %s",
379 name, ovs_strerror(error));
382 VLOG_WARN("failed to create datapath %s: %s",
383 name, ovs_strerror(error));
388 /* Closes and frees the connection to 'dpif'. Does not destroy the datapath
389 * itself; call dpif_delete() first, instead, if that is desirable. */
391 dpif_close(struct dpif *dpif)
394 struct registered_dpif_class *rc;
396 rc = shash_find_data(&dpif_classes, dpif->dpif_class->type);
397 dpif_uninit(dpif, true);
402 /* Performs periodic work needed by 'dpif'. */
404 dpif_run(struct dpif *dpif)
406 if (dpif->dpif_class->run) {
407 dpif->dpif_class->run(dpif);
411 /* Arranges for poll_block() to wake up when dp_run() needs to be called for
414 dpif_wait(struct dpif *dpif)
416 if (dpif->dpif_class->wait) {
417 dpif->dpif_class->wait(dpif);
421 /* Returns the name of datapath 'dpif' prefixed with the type
422 * (for use in log messages). */
424 dpif_name(const struct dpif *dpif)
426 return dpif->full_name;
429 /* Returns the name of datapath 'dpif' without the type
430 * (for use in device names). */
432 dpif_base_name(const struct dpif *dpif)
434 return dpif->base_name;
437 /* Returns the type of datapath 'dpif'. */
439 dpif_type(const struct dpif *dpif)
441 return dpif->dpif_class->type;
444 /* Returns the fully spelled out name for the given datapath 'type'.
446 * Normalized type string can be compared with strcmp(). Unnormalized type
447 * string might be the same even if they have different spellings. */
449 dpif_normalize_type(const char *type)
451 return type && type[0] ? type : "system";
454 /* Destroys the datapath that 'dpif' is connected to, first removing all of its
455 * ports. After calling this function, it does not make sense to pass 'dpif'
456 * to any functions other than dpif_name() or dpif_close(). */
458 dpif_delete(struct dpif *dpif)
462 COVERAGE_INC(dpif_destroy);
464 error = dpif->dpif_class->destroy(dpif);
465 log_operation(dpif, "delete", error);
469 /* Retrieves statistics for 'dpif' into 'stats'. Returns 0 if successful,
470 * otherwise a positive errno value. */
472 dpif_get_dp_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
474 int error = dpif->dpif_class->get_stats(dpif, stats);
476 memset(stats, 0, sizeof *stats);
478 log_operation(dpif, "get_stats", error);
483 dpif_port_open_type(const char *datapath_type, const char *port_type)
485 struct registered_dpif_class *rc;
487 datapath_type = dpif_normalize_type(datapath_type);
489 ovs_mutex_lock(&dpif_mutex);
490 rc = shash_find_data(&dpif_classes, datapath_type);
491 if (rc && rc->dpif_class->port_open_type) {
492 port_type = rc->dpif_class->port_open_type(rc->dpif_class, port_type);
494 ovs_mutex_unlock(&dpif_mutex);
499 /* Attempts to add 'netdev' as a port on 'dpif'. If 'port_nop' is
500 * non-null and its value is not ODPP_NONE, then attempts to use the
501 * value as the port number.
503 * If successful, returns 0 and sets '*port_nop' to the new port's port
504 * number (if 'port_nop' is non-null). On failure, returns a positive
505 * errno value and sets '*port_nop' to ODPP_NONE (if 'port_nop' is
508 dpif_port_add(struct dpif *dpif, struct netdev *netdev, odp_port_t *port_nop)
510 const char *netdev_name = netdev_get_name(netdev);
511 odp_port_t port_no = ODPP_NONE;
514 COVERAGE_INC(dpif_port_add);
520 error = dpif->dpif_class->port_add(dpif, netdev, &port_no);
522 VLOG_DBG_RL(&dpmsg_rl, "%s: added %s as port %"PRIu32,
523 dpif_name(dpif), netdev_name, port_no);
525 VLOG_WARN_RL(&error_rl, "%s: failed to add %s as port: %s",
526 dpif_name(dpif), netdev_name, ovs_strerror(error));
535 /* Attempts to remove 'dpif''s port number 'port_no'. Returns 0 if successful,
536 * otherwise a positive errno value. */
538 dpif_port_del(struct dpif *dpif, odp_port_t port_no)
542 COVERAGE_INC(dpif_port_del);
544 error = dpif->dpif_class->port_del(dpif, port_no);
546 VLOG_DBG_RL(&dpmsg_rl, "%s: port_del(%"PRIu32")",
547 dpif_name(dpif), port_no);
549 log_operation(dpif, "port_del", error);
554 /* Makes a deep copy of 'src' into 'dst'. */
556 dpif_port_clone(struct dpif_port *dst, const struct dpif_port *src)
558 dst->name = xstrdup(src->name);
559 dst->type = xstrdup(src->type);
560 dst->port_no = src->port_no;
563 /* Frees memory allocated to members of 'dpif_port'.
565 * Do not call this function on a dpif_port obtained from
566 * dpif_port_dump_next(): that function retains ownership of the data in the
569 dpif_port_destroy(struct dpif_port *dpif_port)
571 free(dpif_port->name);
572 free(dpif_port->type);
575 /* Checks if port named 'devname' exists in 'dpif'. If so, returns
576 * true; otherwise, returns false. */
578 dpif_port_exists(const struct dpif *dpif, const char *devname)
580 int error = dpif->dpif_class->port_query_by_name(dpif, devname, NULL);
581 if (error != 0 && error != ENOENT && error != ENODEV) {
582 VLOG_WARN_RL(&error_rl, "%s: failed to query port %s: %s",
583 dpif_name(dpif), devname, ovs_strerror(error));
589 /* Looks up port number 'port_no' in 'dpif'. On success, returns 0 and
590 * initializes '*port' appropriately; on failure, returns a positive errno
593 * The caller owns the data in 'port' and must free it with
594 * dpif_port_destroy() when it is no longer needed. */
596 dpif_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
597 struct dpif_port *port)
599 int error = dpif->dpif_class->port_query_by_number(dpif, port_no, port);
601 VLOG_DBG_RL(&dpmsg_rl, "%s: port %"PRIu32" is device %s",
602 dpif_name(dpif), port_no, port->name);
604 memset(port, 0, sizeof *port);
605 VLOG_WARN_RL(&error_rl, "%s: failed to query port %"PRIu32": %s",
606 dpif_name(dpif), port_no, ovs_strerror(error));
611 /* Looks up port named 'devname' in 'dpif'. On success, returns 0 and
612 * initializes '*port' appropriately; on failure, returns a positive errno
615 * The caller owns the data in 'port' and must free it with
616 * dpif_port_destroy() when it is no longer needed. */
618 dpif_port_query_by_name(const struct dpif *dpif, const char *devname,
619 struct dpif_port *port)
621 int error = dpif->dpif_class->port_query_by_name(dpif, devname, port);
623 VLOG_DBG_RL(&dpmsg_rl, "%s: device %s is on port %"PRIu32,
624 dpif_name(dpif), devname, port->port_no);
626 memset(port, 0, sizeof *port);
628 /* For ENOENT or ENODEV we use DBG level because the caller is probably
629 * interested in whether 'dpif' actually has a port 'devname', so that
630 * it's not an issue worth logging if it doesn't. Other errors are
631 * uncommon and more likely to indicate a real problem. */
633 error == ENOENT || error == ENODEV ? VLL_DBG : VLL_WARN,
634 "%s: failed to query port %s: %s",
635 dpif_name(dpif), devname, ovs_strerror(error));
640 /* Returns the Netlink PID value to supply in OVS_ACTION_ATTR_USERSPACE
641 * actions as the OVS_USERSPACE_ATTR_PID attribute's value, for use in
642 * flows whose packets arrived on port 'port_no'. In the case where the
643 * provider allocates multiple Netlink PIDs to a single port, it may use
644 * 'hash' to spread load among them. The caller need not use a particular
645 * hash function; a 5-tuple hash is suitable.
647 * (The datapath implementation might use some different hash function for
648 * distributing packets received via flow misses among PIDs. This means
649 * that packets received via flow misses might be reordered relative to
650 * packets received via userspace actions. This is not ordinarily a
653 * A 'port_no' of ODPP_NONE is a special case: it returns a reserved PID, not
654 * allocated to any port, that the client may use for special purposes.
656 * The return value is only meaningful when DPIF_UC_ACTION has been enabled in
657 * the 'dpif''s listen mask. It is allowed to change when DPIF_UC_ACTION is
658 * disabled and then re-enabled, so a client that does that must be prepared to
659 * update all of the flows that it installed that contain
660 * OVS_ACTION_ATTR_USERSPACE actions. */
662 dpif_port_get_pid(const struct dpif *dpif, odp_port_t port_no, uint32_t hash)
664 return (dpif->dpif_class->port_get_pid
665 ? (dpif->dpif_class->port_get_pid)(dpif, port_no, hash)
669 /* Looks up port number 'port_no' in 'dpif'. On success, returns 0 and copies
670 * the port's name into the 'name_size' bytes in 'name', ensuring that the
671 * result is null-terminated. On failure, returns a positive errno value and
672 * makes 'name' the empty string. */
674 dpif_port_get_name(struct dpif *dpif, odp_port_t port_no,
675 char *name, size_t name_size)
677 struct dpif_port port;
680 ovs_assert(name_size > 0);
682 error = dpif_port_query_by_number(dpif, port_no, &port);
684 ovs_strlcpy(name, port.name, name_size);
685 dpif_port_destroy(&port);
692 /* Initializes 'dump' to begin dumping the ports in a dpif.
694 * This function provides no status indication. An error status for the entire
695 * dump operation is provided when it is completed by calling
696 * dpif_port_dump_done().
699 dpif_port_dump_start(struct dpif_port_dump *dump, const struct dpif *dpif)
702 dump->error = dpif->dpif_class->port_dump_start(dpif, &dump->state);
703 log_operation(dpif, "port_dump_start", dump->error);
706 /* Attempts to retrieve another port from 'dump', which must have been
707 * initialized with dpif_port_dump_start(). On success, stores a new dpif_port
708 * into 'port' and returns true. On failure, returns false.
710 * Failure might indicate an actual error or merely that the last port has been
711 * dumped. An error status for the entire dump operation is provided when it
712 * is completed by calling dpif_port_dump_done().
714 * The dpif owns the data stored in 'port'. It will remain valid until at
715 * least the next time 'dump' is passed to dpif_port_dump_next() or
716 * dpif_port_dump_done(). */
718 dpif_port_dump_next(struct dpif_port_dump *dump, struct dpif_port *port)
720 const struct dpif *dpif = dump->dpif;
726 dump->error = dpif->dpif_class->port_dump_next(dpif, dump->state, port);
727 if (dump->error == EOF) {
728 VLOG_DBG_RL(&dpmsg_rl, "%s: dumped all ports", dpif_name(dpif));
730 log_operation(dpif, "port_dump_next", dump->error);
734 dpif->dpif_class->port_dump_done(dpif, dump->state);
740 /* Completes port table dump operation 'dump', which must have been initialized
741 * with dpif_port_dump_start(). Returns 0 if the dump operation was
742 * error-free, otherwise a positive errno value describing the problem. */
744 dpif_port_dump_done(struct dpif_port_dump *dump)
746 const struct dpif *dpif = dump->dpif;
748 dump->error = dpif->dpif_class->port_dump_done(dpif, dump->state);
749 log_operation(dpif, "port_dump_done", dump->error);
751 return dump->error == EOF ? 0 : dump->error;
754 /* Polls for changes in the set of ports in 'dpif'. If the set of ports in
755 * 'dpif' has changed, this function does one of the following:
757 * - Stores the name of the device that was added to or deleted from 'dpif' in
758 * '*devnamep' and returns 0. The caller is responsible for freeing
759 * '*devnamep' (with free()) when it no longer needs it.
761 * - Returns ENOBUFS and sets '*devnamep' to NULL.
763 * This function may also return 'false positives', where it returns 0 and
764 * '*devnamep' names a device that was not actually added or deleted or it
765 * returns ENOBUFS without any change.
767 * Returns EAGAIN if the set of ports in 'dpif' has not changed. May also
768 * return other positive errno values to indicate that something has gone
771 dpif_port_poll(const struct dpif *dpif, char **devnamep)
773 int error = dpif->dpif_class->port_poll(dpif, devnamep);
780 /* Arranges for the poll loop to wake up when port_poll(dpif) will return a
781 * value other than EAGAIN. */
783 dpif_port_poll_wait(const struct dpif *dpif)
785 dpif->dpif_class->port_poll_wait(dpif);
788 /* Extracts the flow stats for a packet. The 'flow' and 'packet'
789 * arguments must have been initialized through a call to flow_extract().
790 * 'used' is stored into stats->used. */
792 dpif_flow_stats_extract(const struct flow *flow, const struct ofpbuf *packet,
793 long long int used, struct dpif_flow_stats *stats)
795 stats->tcp_flags = ntohs(flow->tcp_flags);
796 stats->n_bytes = ofpbuf_size(packet);
797 stats->n_packets = 1;
801 /* Appends a human-readable representation of 'stats' to 's'. */
803 dpif_flow_stats_format(const struct dpif_flow_stats *stats, struct ds *s)
805 ds_put_format(s, "packets:%"PRIu64", bytes:%"PRIu64", used:",
806 stats->n_packets, stats->n_bytes);
808 ds_put_format(s, "%.3fs", (time_msec() - stats->used) / 1000.0);
810 ds_put_format(s, "never");
812 if (stats->tcp_flags) {
813 ds_put_cstr(s, ", flags:");
814 packet_format_tcp_flags(s, stats->tcp_flags);
818 /* Deletes all flows from 'dpif'. Returns 0 if successful, otherwise a
819 * positive errno value. */
821 dpif_flow_flush(struct dpif *dpif)
825 COVERAGE_INC(dpif_flow_flush);
827 error = dpif->dpif_class->flow_flush(dpif);
828 log_operation(dpif, "flow_flush", error);
832 /* Queries 'dpif' for a flow entry. The flow is specified by the Netlink
833 * attributes with types OVS_KEY_ATTR_* in the 'key_len' bytes starting at
836 * Returns 0 if successful. If no flow matches, returns ENOENT. On other
837 * failure, returns a positive errno value.
839 * On success, '*bufp' will be set to an ofpbuf owned by the caller that
840 * contains the response for 'flow->mask' and 'flow->actions'. The caller must
841 * supply a valid pointer, and must free the ofpbuf (with ofpbuf_delete()) when
842 * it is no longer needed.
844 * On success, 'flow' will be populated with the mask, actions and stats for
845 * the datapath flow corresponding to 'key'. The mask and actions will point
848 * Implementations may opt to point 'flow->mask' and/or 'flow->actions' at
849 * RCU-protected data rather than making a copy of them. Therefore, callers
850 * that wish to hold these over quiescent periods must make a copy of these
851 * fields before quiescing. */
853 dpif_flow_get(const struct dpif *dpif,
854 const struct nlattr *key, size_t key_len,
855 struct ofpbuf **bufp, struct dpif_flow *flow)
858 struct nlattr *mask, *actions;
859 size_t mask_len, actions_len;
860 struct dpif_flow_stats stats;
862 COVERAGE_INC(dpif_flow_get);
865 error = dpif->dpif_class->flow_get(dpif, key, key_len, bufp,
867 &actions, &actions_len, &stats);
869 memset(flow, 0, sizeof *flow);
870 ofpbuf_delete(*bufp);
874 flow->mask_len = mask_len;
875 flow->actions = actions;
876 flow->actions_len = actions_len;
879 if (should_log_flow_message(error)) {
880 log_flow_message(dpif, error, "flow_get", key, key_len,
881 NULL, 0, &flow->stats,
882 flow->actions, flow->actions_len);
887 /* A dpif_operate() wrapper for performing a single DPIF_OP_FLOW_PUT. */
889 dpif_flow_put(struct dpif *dpif, enum dpif_flow_put_flags flags,
890 const struct nlattr *key, size_t key_len,
891 const struct nlattr *mask, size_t mask_len,
892 const struct nlattr *actions, size_t actions_len,
893 struct dpif_flow_stats *stats)
898 op.type = DPIF_OP_FLOW_PUT;
899 op.u.flow_put.flags = flags;
900 op.u.flow_put.key = key;
901 op.u.flow_put.key_len = key_len;
902 op.u.flow_put.mask = mask;
903 op.u.flow_put.mask_len = mask_len;
904 op.u.flow_put.actions = actions;
905 op.u.flow_put.actions_len = actions_len;
906 op.u.flow_put.stats = stats;
909 dpif_operate(dpif, &opp, 1);
914 /* A dpif_operate() wrapper for performing a single DPIF_OP_FLOW_DEL. */
916 dpif_flow_del(struct dpif *dpif,
917 const struct nlattr *key, size_t key_len,
918 struct dpif_flow_stats *stats)
923 op.type = DPIF_OP_FLOW_DEL;
924 op.u.flow_del.key = key;
925 op.u.flow_del.key_len = key_len;
926 op.u.flow_del.stats = stats;
929 dpif_operate(dpif, &opp, 1);
934 /* Creates and returns a new 'struct dpif_flow_dump' for iterating through the
937 * This function always successfully returns a dpif_flow_dump. Error
938 * reporting is deferred to dpif_flow_dump_destroy(). */
939 struct dpif_flow_dump *
940 dpif_flow_dump_create(const struct dpif *dpif)
942 return dpif->dpif_class->flow_dump_create(dpif);
945 /* Destroys 'dump', which must have been created with dpif_flow_dump_create().
946 * All dpif_flow_dump_thread structures previously created for 'dump' must
947 * previously have been destroyed.
949 * Returns 0 if the dump operation was error-free, otherwise a positive errno
950 * value describing the problem. */
952 dpif_flow_dump_destroy(struct dpif_flow_dump *dump)
954 const struct dpif *dpif = dump->dpif;
955 int error = dpif->dpif_class->flow_dump_destroy(dump);
956 log_operation(dpif, "flow_dump_destroy", error);
957 return error == EOF ? 0 : error;
960 /* Returns new thread-local state for use with dpif_flow_dump_next(). */
961 struct dpif_flow_dump_thread *
962 dpif_flow_dump_thread_create(struct dpif_flow_dump *dump)
964 return dump->dpif->dpif_class->flow_dump_thread_create(dump);
967 /* Releases 'thread'. */
969 dpif_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread)
971 thread->dpif->dpif_class->flow_dump_thread_destroy(thread);
974 /* Attempts to retrieve up to 'max_flows' more flows from 'thread'. Returns 0
975 * if and only if no flows remained to be retrieved, otherwise a positive
976 * number reflecting the number of elements in 'flows[]' that were updated.
977 * The number of flows returned might be less than 'max_flows' because
978 * fewer than 'max_flows' remained, because this particular datapath does not
979 * benefit from batching, or because an error occurred partway through
980 * retrieval. Thus, the caller should continue calling until a 0 return value,
981 * even if intermediate return values are less than 'max_flows'.
983 * No error status is immediately provided. An error status for the entire
984 * dump operation is provided when it is completed by calling
985 * dpif_flow_dump_destroy().
987 * All of the data stored into 'flows' is owned by the datapath, not by the
988 * caller, and the caller must not modify or free it. The datapath guarantees
989 * that it remains accessible and unchanged until the first of:
990 * - The next call to dpif_flow_dump_next() for 'thread', or
991 * - The next rcu quiescent period. */
993 dpif_flow_dump_next(struct dpif_flow_dump_thread *thread,
994 struct dpif_flow *flows, int max_flows)
996 struct dpif *dpif = thread->dpif;
999 ovs_assert(max_flows > 0);
1000 n = dpif->dpif_class->flow_dump_next(thread, flows, max_flows);
1002 struct dpif_flow *f;
1004 for (f = flows; f < &flows[n] && should_log_flow_message(0); f++) {
1005 log_flow_message(dpif, 0, "flow_dump",
1006 f->key, f->key_len, f->mask, f->mask_len,
1007 &f->stats, f->actions, f->actions_len);
1010 VLOG_DBG_RL(&dpmsg_rl, "%s: dumped all flows", dpif_name(dpif));
1015 struct dpif_execute_helper_aux {
1020 /* This is called for actions that need the context of the datapath to be
1023 dpif_execute_helper_cb(void *aux_, struct dpif_packet **packets, int cnt,
1024 struct pkt_metadata *md,
1025 const struct nlattr *action, bool may_steal OVS_UNUSED)
1027 struct dpif_execute_helper_aux *aux = aux_;
1028 int type = nl_attr_type(action);
1029 struct ofpbuf * packet = &packets[0]->ofpbuf;
1031 ovs_assert(cnt == 1);
1033 switch ((enum ovs_action_attr)type) {
1034 case OVS_ACTION_ATTR_OUTPUT:
1035 case OVS_ACTION_ATTR_USERSPACE:
1036 case OVS_ACTION_ATTR_RECIRC: {
1037 struct dpif_execute execute;
1038 struct ofpbuf execute_actions;
1039 uint64_t stub[256 / 8];
1041 if (md->tunnel.ip_dst) {
1042 /* The Linux kernel datapath throws away the tunnel information
1043 * that we supply as metadata. We have to use a "set" action to
1045 ofpbuf_use_stub(&execute_actions, stub, sizeof stub);
1046 odp_put_tunnel_action(&md->tunnel, &execute_actions);
1047 ofpbuf_put(&execute_actions, action, NLA_ALIGN(action->nla_len));
1049 execute.actions = ofpbuf_data(&execute_actions);
1050 execute.actions_len = ofpbuf_size(&execute_actions);
1052 execute.actions = action;
1053 execute.actions_len = NLA_ALIGN(action->nla_len);
1056 execute.packet = packet;
1058 execute.needs_help = false;
1059 aux->error = dpif_execute(aux->dpif, &execute);
1060 log_execute_message(aux->dpif, &execute, true, aux->error);
1062 if (md->tunnel.ip_dst) {
1063 ofpbuf_uninit(&execute_actions);
1068 case OVS_ACTION_ATTR_HASH:
1069 case OVS_ACTION_ATTR_PUSH_VLAN:
1070 case OVS_ACTION_ATTR_POP_VLAN:
1071 case OVS_ACTION_ATTR_PUSH_MPLS:
1072 case OVS_ACTION_ATTR_POP_MPLS:
1073 case OVS_ACTION_ATTR_SET:
1074 case OVS_ACTION_ATTR_SAMPLE:
1075 case OVS_ACTION_ATTR_UNSPEC:
1076 case __OVS_ACTION_ATTR_MAX:
1081 /* Executes 'execute' by performing most of the actions in userspace and
1082 * passing the fully constructed packets to 'dpif' for output and userspace
1085 * This helps with actions that a given 'dpif' doesn't implement directly. */
1087 dpif_execute_with_help(struct dpif *dpif, struct dpif_execute *execute)
1089 struct dpif_execute_helper_aux aux = {dpif, 0};
1090 struct dpif_packet packet, *pp;
1092 COVERAGE_INC(dpif_execute_with_help);
1094 packet.ofpbuf = *execute->packet;
1097 odp_execute_actions(&aux, &pp, 1, false, &execute->md, execute->actions,
1098 execute->actions_len, dpif_execute_helper_cb);
1100 /* Even though may_steal is set to false, some actions could modify or
1101 * reallocate the ofpbuf memory. We need to pass those changes to the
1103 *execute->packet = packet.ofpbuf;
1108 /* Returns true if the datapath needs help executing 'execute'. */
1110 dpif_execute_needs_help(const struct dpif_execute *execute)
1112 return execute->needs_help || nl_attr_oversized(execute->actions_len);
1115 /* A dpif_operate() wrapper for performing a single DPIF_OP_EXECUTE. */
1117 dpif_execute(struct dpif *dpif, struct dpif_execute *execute)
1119 if (execute->actions_len) {
1120 struct dpif_op *opp;
1123 op.type = DPIF_OP_EXECUTE;
1124 op.u.execute = *execute;
1127 dpif_operate(dpif, &opp, 1);
1135 /* Executes each of the 'n_ops' operations in 'ops' on 'dpif', in the order in
1136 * which they are specified. Places each operation's results in the "output"
1137 * members documented in comments, and 0 in the 'error' member on success or a
1138 * positive errno on failure. */
1140 dpif_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
1145 /* Count 'chunk', the number of ops that can be executed without
1146 * needing any help. Ops that need help should be rare, so we
1147 * expect this to ordinarily be 'n_ops', that is, all the ops. */
1148 for (chunk = 0; chunk < n_ops; chunk++) {
1149 struct dpif_op *op = ops[chunk];
1151 if (op->type == DPIF_OP_EXECUTE
1152 && dpif_execute_needs_help(&op->u.execute)) {
1158 /* Execute a chunk full of ops that the dpif provider can
1159 * handle itself, without help. */
1162 dpif->dpif_class->operate(dpif, ops, chunk);
1164 for (i = 0; i < chunk; i++) {
1165 struct dpif_op *op = ops[i];
1166 int error = op->error;
1169 case DPIF_OP_FLOW_PUT: {
1170 struct dpif_flow_put *put = &op->u.flow_put;
1172 COVERAGE_INC(dpif_flow_put);
1173 log_flow_put_message(dpif, put, error);
1174 if (error && put->stats) {
1175 memset(put->stats, 0, sizeof *put->stats);
1180 case DPIF_OP_FLOW_DEL: {
1181 struct dpif_flow_del *del = &op->u.flow_del;
1183 COVERAGE_INC(dpif_flow_del);
1184 log_flow_del_message(dpif, del, error);
1185 if (error && del->stats) {
1186 memset(del->stats, 0, sizeof *del->stats);
1191 case DPIF_OP_EXECUTE:
1192 COVERAGE_INC(dpif_execute);
1193 log_execute_message(dpif, &op->u.execute, false, error);
1201 /* Help the dpif provider to execute one op. */
1202 struct dpif_op *op = ops[0];
1204 COVERAGE_INC(dpif_execute);
1205 op->error = dpif_execute_with_help(dpif, &op->u.execute);
1212 /* Returns a string that represents 'type', for use in log messages. */
1214 dpif_upcall_type_to_string(enum dpif_upcall_type type)
1217 case DPIF_UC_MISS: return "miss";
1218 case DPIF_UC_ACTION: return "action";
1219 case DPIF_N_UC_TYPES: default: return "<unknown>";
1223 /* Enables or disables receiving packets with dpif_recv() on 'dpif'. Returns 0
1224 * if successful, otherwise a positive errno value.
1226 * Turning packet receive off and then back on may change the Netlink PID
1227 * assignments returned by dpif_port_get_pid(). If the client does this, it
1228 * must update all of the flows that have OVS_ACTION_ATTR_USERSPACE actions
1229 * using the new PID assignment. */
1231 dpif_recv_set(struct dpif *dpif, bool enable)
1235 if (dpif->dpif_class->recv_set) {
1236 error = dpif->dpif_class->recv_set(dpif, enable);
1237 log_operation(dpif, "recv_set", error);
1242 /* Refreshes the poll loops and Netlink sockets associated to each port,
1243 * when the number of upcall handlers (upcall receiving thread) is changed
1244 * to 'n_handlers' and receiving packets for 'dpif' is enabled by
1247 * Since multiple upcall handlers can read upcalls simultaneously from
1248 * 'dpif', each port can have multiple Netlink sockets, one per upcall
1249 * handler. So, handlers_set() is responsible for the following tasks:
1251 * When receiving upcall is enabled, extends or creates the
1252 * configuration to support:
1254 * - 'n_handlers' Netlink sockets for each port.
1256 * - 'n_handlers' poll loops, one for each upcall handler.
1258 * - registering the Netlink sockets for the same upcall handler to
1259 * the corresponding poll loop.
1261 * Returns 0 if successful, otherwise a positive errno value. */
1263 dpif_handlers_set(struct dpif *dpif, uint32_t n_handlers)
1267 if (dpif->dpif_class->handlers_set) {
1268 error = dpif->dpif_class->handlers_set(dpif, n_handlers);
1269 log_operation(dpif, "handlers_set", error);
1275 dpif_register_upcall_cb(struct dpif *dpif, exec_upcall_cb *cb)
1277 if (dpif->dpif_class->register_upcall_cb) {
1278 dpif->dpif_class->register_upcall_cb(dpif, cb);
1283 dpif_enable_upcall(struct dpif *dpif)
1285 if (dpif->dpif_class->enable_upcall) {
1286 dpif->dpif_class->enable_upcall(dpif);
1291 dpif_disable_upcall(struct dpif *dpif)
1293 if (dpif->dpif_class->disable_upcall) {
1294 dpif->dpif_class->disable_upcall(dpif);
1299 dpif_print_packet(struct dpif *dpif, struct dpif_upcall *upcall)
1301 if (!VLOG_DROP_DBG(&dpmsg_rl)) {
1305 packet = ofp_packet_to_string(ofpbuf_data(&upcall->packet),
1306 ofpbuf_size(&upcall->packet));
1309 odp_flow_key_format(upcall->key, upcall->key_len, &flow);
1311 VLOG_DBG("%s: %s upcall:\n%s\n%s",
1312 dpif_name(dpif), dpif_upcall_type_to_string(upcall->type),
1313 ds_cstr(&flow), packet);
1320 /* Polls for an upcall from 'dpif' for an upcall handler. Since there
1321 * there can be multiple poll loops, 'handler_id' is needed as index to
1322 * identify the corresponding poll loop. If successful, stores the upcall
1323 * into '*upcall', using 'buf' for storage. Should only be called if
1324 * 'recv_set' has been used to enable receiving packets from 'dpif'.
1326 * 'upcall->key' and 'upcall->userdata' point into data in the caller-provided
1327 * 'buf', so their memory cannot be freed separately from 'buf'.
1329 * The caller owns the data of 'upcall->packet' and may modify it. If
1330 * packet's headroom is exhausted as it is manipulated, 'upcall->packet'
1331 * will be reallocated. This requires the data of 'upcall->packet' to be
1332 * released with ofpbuf_uninit() before 'upcall' is destroyed. However,
1333 * when an error is returned, the 'upcall->packet' may be uninitialized
1334 * and should not be released.
1336 * Returns 0 if successful, otherwise a positive errno value. Returns EAGAIN
1337 * if no upcall is immediately available. */
1339 dpif_recv(struct dpif *dpif, uint32_t handler_id, struct dpif_upcall *upcall,
1344 if (dpif->dpif_class->recv) {
1345 error = dpif->dpif_class->recv(dpif, handler_id, upcall, buf);
1347 dpif_print_packet(dpif, upcall);
1348 } else if (error != EAGAIN) {
1349 log_operation(dpif, "recv", error);
1355 /* Discards all messages that would otherwise be received by dpif_recv() on
1358 dpif_recv_purge(struct dpif *dpif)
1360 COVERAGE_INC(dpif_purge);
1361 if (dpif->dpif_class->recv_purge) {
1362 dpif->dpif_class->recv_purge(dpif);
1366 /* Arranges for the poll loop for an upcall handler to wake up when 'dpif'
1367 * 'dpif' has a message queued to be received with the recv member
1368 * function. Since there can be multiple poll loops, 'handler_id' is
1369 * needed as index to identify the corresponding poll loop. */
1371 dpif_recv_wait(struct dpif *dpif, uint32_t handler_id)
1373 if (dpif->dpif_class->recv_wait) {
1374 dpif->dpif_class->recv_wait(dpif, handler_id);
1378 /* Obtains the NetFlow engine type and engine ID for 'dpif' into '*engine_type'
1379 * and '*engine_id', respectively. */
1381 dpif_get_netflow_ids(const struct dpif *dpif,
1382 uint8_t *engine_type, uint8_t *engine_id)
1384 *engine_type = dpif->netflow_engine_type;
1385 *engine_id = dpif->netflow_engine_id;
1388 /* Translates OpenFlow queue ID 'queue_id' (in host byte order) into a priority
1389 * value used for setting packet priority.
1390 * On success, returns 0 and stores the priority into '*priority'.
1391 * On failure, returns a positive errno value and stores 0 into '*priority'. */
1393 dpif_queue_to_priority(const struct dpif *dpif, uint32_t queue_id,
1396 int error = (dpif->dpif_class->queue_to_priority
1397 ? dpif->dpif_class->queue_to_priority(dpif, queue_id,
1403 log_operation(dpif, "queue_to_priority", error);
1408 dpif_init(struct dpif *dpif, const struct dpif_class *dpif_class,
1410 uint8_t netflow_engine_type, uint8_t netflow_engine_id)
1412 dpif->dpif_class = dpif_class;
1413 dpif->base_name = xstrdup(name);
1414 dpif->full_name = xasprintf("%s@%s", dpif_class->type, name);
1415 dpif->netflow_engine_type = netflow_engine_type;
1416 dpif->netflow_engine_id = netflow_engine_id;
1419 /* Undoes the results of initialization.
1421 * Normally this function only needs to be called from dpif_close().
1422 * However, it may be called by providers due to an error on opening
1423 * that occurs after initialization. It this case dpif_close() would
1424 * never be called. */
1426 dpif_uninit(struct dpif *dpif, bool close)
1428 char *base_name = dpif->base_name;
1429 char *full_name = dpif->full_name;
1432 dpif->dpif_class->close(dpif);
1440 log_operation(const struct dpif *dpif, const char *operation, int error)
1443 VLOG_DBG_RL(&dpmsg_rl, "%s: %s success", dpif_name(dpif), operation);
1444 } else if (ofperr_is_valid(error)) {
1445 VLOG_WARN_RL(&error_rl, "%s: %s failed (%s)",
1446 dpif_name(dpif), operation, ofperr_get_name(error));
1448 VLOG_WARN_RL(&error_rl, "%s: %s failed (%s)",
1449 dpif_name(dpif), operation, ovs_strerror(error));
1453 static enum vlog_level
1454 flow_message_log_level(int error)
1456 /* If flows arrive in a batch, userspace may push down multiple
1457 * unique flow definitions that overlap when wildcards are applied.
1458 * Kernels that support flow wildcarding will reject these flows as
1459 * duplicates (EEXIST), so lower the log level to debug for these
1460 * types of messages. */
1461 return (error && error != EEXIST) ? VLL_WARN : VLL_DBG;
1465 should_log_flow_message(int error)
1467 return !vlog_should_drop(THIS_MODULE, flow_message_log_level(error),
1468 error ? &error_rl : &dpmsg_rl);
1472 log_flow_message(const struct dpif *dpif, int error, const char *operation,
1473 const struct nlattr *key, size_t key_len,
1474 const struct nlattr *mask, size_t mask_len,
1475 const struct dpif_flow_stats *stats,
1476 const struct nlattr *actions, size_t actions_len)
1478 struct ds ds = DS_EMPTY_INITIALIZER;
1479 ds_put_format(&ds, "%s: ", dpif_name(dpif));
1481 ds_put_cstr(&ds, "failed to ");
1483 ds_put_format(&ds, "%s ", operation);
1485 ds_put_format(&ds, "(%s) ", ovs_strerror(error));
1487 odp_flow_format(key, key_len, mask, mask_len, NULL, &ds, true);
1489 ds_put_cstr(&ds, ", ");
1490 dpif_flow_stats_format(stats, &ds);
1492 if (actions || actions_len) {
1493 ds_put_cstr(&ds, ", actions:");
1494 format_odp_actions(&ds, actions, actions_len);
1496 vlog(THIS_MODULE, flow_message_log_level(error), "%s", ds_cstr(&ds));
1501 log_flow_put_message(struct dpif *dpif, const struct dpif_flow_put *put,
1504 if (should_log_flow_message(error)) {
1508 ds_put_cstr(&s, "put");
1509 if (put->flags & DPIF_FP_CREATE) {
1510 ds_put_cstr(&s, "[create]");
1512 if (put->flags & DPIF_FP_MODIFY) {
1513 ds_put_cstr(&s, "[modify]");
1515 if (put->flags & DPIF_FP_ZERO_STATS) {
1516 ds_put_cstr(&s, "[zero]");
1518 log_flow_message(dpif, error, ds_cstr(&s),
1519 put->key, put->key_len, put->mask, put->mask_len,
1520 put->stats, put->actions, put->actions_len);
1526 log_flow_del_message(struct dpif *dpif, const struct dpif_flow_del *del,
1529 if (should_log_flow_message(error)) {
1530 log_flow_message(dpif, error, "flow_del", del->key, del->key_len,
1531 NULL, 0, !error ? del->stats : NULL, NULL, 0);
1535 /* Logs that 'execute' was executed on 'dpif' and completed with errno 'error'
1536 * (0 for success). 'subexecute' should be true if the execution is a result
1537 * of breaking down a larger execution that needed help, false otherwise.
1540 * XXX In theory, the log message could be deceptive because this function is
1541 * called after the dpif_provider's '->execute' function, which is allowed to
1542 * modify execute->packet and execute->md. In practice, though:
1544 * - dpif-linux doesn't modify execute->packet or execute->md.
1546 * - dpif-netdev does modify them but it is less likely to have problems
1547 * because it is built into ovs-vswitchd and cannot have version skew,
1550 * It would still be better to avoid the potential problem. I don't know of a
1551 * good way to do that, though, that isn't expensive. */
1553 log_execute_message(struct dpif *dpif, const struct dpif_execute *execute,
1554 bool subexecute, int error)
1556 if (!(error ? VLOG_DROP_WARN(&error_rl) : VLOG_DROP_DBG(&dpmsg_rl))) {
1557 struct ds ds = DS_EMPTY_INITIALIZER;
1560 packet = ofp_packet_to_string(ofpbuf_data(execute->packet),
1561 ofpbuf_size(execute->packet));
1562 ds_put_format(&ds, "%s: %sexecute ",
1564 (subexecute ? "sub-"
1565 : dpif_execute_needs_help(execute) ? "super-"
1567 format_odp_actions(&ds, execute->actions, execute->actions_len);
1569 ds_put_format(&ds, " failed (%s)", ovs_strerror(error));
1571 ds_put_format(&ds, " on packet %s", packet);
1572 vlog(THIS_MODULE, error ? VLL_WARN : VLL_DBG, "%s", ds_cstr(&ds));