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"
27 #include "dynamic-string.h"
31 #include "odp-execute.h"
33 #include "ofp-errors.h"
34 #include "ofp-print.h"
38 #include "poll-loop.h"
46 VLOG_DEFINE_THIS_MODULE(dpif);
48 COVERAGE_DEFINE(dpif_destroy);
49 COVERAGE_DEFINE(dpif_port_add);
50 COVERAGE_DEFINE(dpif_port_del);
51 COVERAGE_DEFINE(dpif_flow_flush);
52 COVERAGE_DEFINE(dpif_flow_get);
53 COVERAGE_DEFINE(dpif_flow_put);
54 COVERAGE_DEFINE(dpif_flow_del);
55 COVERAGE_DEFINE(dpif_execute);
56 COVERAGE_DEFINE(dpif_purge);
57 COVERAGE_DEFINE(dpif_execute_with_help);
59 static const struct dpif_class *base_dpif_classes[] = {
64 &dpif_planetlab_class,
67 struct registered_dpif_class {
68 const struct dpif_class *dpif_class;
71 static struct shash dpif_classes = SHASH_INITIALIZER(&dpif_classes);
72 static struct sset dpif_blacklist = SSET_INITIALIZER(&dpif_blacklist);
74 /* Protects 'dpif_classes', including the refcount, and 'dpif_blacklist'. */
75 static struct ovs_mutex dpif_mutex = OVS_MUTEX_INITIALIZER;
77 /* Rate limit for individual messages going to or from the datapath, output at
78 * DBG level. This is very high because, if these are enabled, it is because
79 * we really need to see them. */
80 static struct vlog_rate_limit dpmsg_rl = VLOG_RATE_LIMIT_INIT(600, 600);
82 /* Not really much point in logging many dpif errors. */
83 static struct vlog_rate_limit error_rl = VLOG_RATE_LIMIT_INIT(60, 5);
85 static void log_flow_message(const struct dpif *dpif, int error,
86 const char *operation,
87 const struct nlattr *key, size_t key_len,
88 const struct nlattr *mask, size_t mask_len,
89 const struct dpif_flow_stats *stats,
90 const struct nlattr *actions, size_t actions_len);
91 static void log_operation(const struct dpif *, const char *operation,
93 static bool should_log_flow_message(int error);
94 static void log_flow_put_message(struct dpif *, const struct dpif_flow_put *,
96 static void log_flow_del_message(struct dpif *, const struct dpif_flow_del *,
98 static void log_execute_message(struct dpif *, const struct dpif_execute *,
104 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
106 if (ovsthread_once_start(&once)) {
109 for (i = 0; i < ARRAY_SIZE(base_dpif_classes); i++) {
110 dp_register_provider(base_dpif_classes[i]);
112 ovsthread_once_done(&once);
117 dp_register_provider__(const struct dpif_class *new_class)
119 struct registered_dpif_class *registered_class;
121 if (sset_contains(&dpif_blacklist, new_class->type)) {
122 VLOG_DBG("attempted to register blacklisted provider: %s",
127 if (shash_find(&dpif_classes, new_class->type)) {
128 VLOG_WARN("attempted to register duplicate datapath provider: %s",
133 registered_class = xmalloc(sizeof *registered_class);
134 registered_class->dpif_class = new_class;
135 registered_class->refcount = 0;
137 shash_add(&dpif_classes, new_class->type, registered_class);
142 /* Registers a new datapath provider. After successful registration, new
143 * datapaths of that type can be opened using dpif_open(). */
145 dp_register_provider(const struct dpif_class *new_class)
149 ovs_mutex_lock(&dpif_mutex);
150 error = dp_register_provider__(new_class);
151 ovs_mutex_unlock(&dpif_mutex);
156 /* Unregisters a datapath provider. 'type' must have been previously
157 * registered and not currently be in use by any dpifs. After unregistration
158 * new datapaths of that type cannot be opened using dpif_open(). */
160 dp_unregister_provider__(const char *type)
162 struct shash_node *node;
163 struct registered_dpif_class *registered_class;
165 node = shash_find(&dpif_classes, type);
167 VLOG_WARN("attempted to unregister a datapath provider that is not "
168 "registered: %s", type);
172 registered_class = node->data;
173 if (registered_class->refcount) {
174 VLOG_WARN("attempted to unregister in use datapath provider: %s", type);
178 shash_delete(&dpif_classes, node);
179 free(registered_class);
184 /* Unregisters a datapath provider. 'type' must have been previously
185 * registered and not currently be in use by any dpifs. After unregistration
186 * new datapaths of that type cannot be opened using dpif_open(). */
188 dp_unregister_provider(const char *type)
194 ovs_mutex_lock(&dpif_mutex);
195 error = dp_unregister_provider__(type);
196 ovs_mutex_unlock(&dpif_mutex);
201 /* Blacklists a provider. Causes future calls of dp_register_provider() with
202 * a dpif_class which implements 'type' to fail. */
204 dp_blacklist_provider(const char *type)
206 ovs_mutex_lock(&dpif_mutex);
207 sset_add(&dpif_blacklist, type);
208 ovs_mutex_unlock(&dpif_mutex);
211 /* Clears 'types' and enumerates the types of all currently registered datapath
212 * providers into it. The caller must first initialize the sset. */
214 dp_enumerate_types(struct sset *types)
216 struct shash_node *node;
221 ovs_mutex_lock(&dpif_mutex);
222 SHASH_FOR_EACH(node, &dpif_classes) {
223 const struct registered_dpif_class *registered_class = node->data;
224 sset_add(types, registered_class->dpif_class->type);
226 ovs_mutex_unlock(&dpif_mutex);
230 dp_class_unref(struct registered_dpif_class *rc)
232 ovs_mutex_lock(&dpif_mutex);
233 ovs_assert(rc->refcount);
235 ovs_mutex_unlock(&dpif_mutex);
238 static struct registered_dpif_class *
239 dp_class_lookup(const char *type)
241 struct registered_dpif_class *rc;
243 ovs_mutex_lock(&dpif_mutex);
244 rc = shash_find_data(&dpif_classes, type);
248 ovs_mutex_unlock(&dpif_mutex);
253 /* Clears 'names' and enumerates the names of all known created datapaths with
254 * the given 'type'. The caller must first initialize the sset. Returns 0 if
255 * successful, otherwise a positive errno value.
257 * Some kinds of datapaths might not be practically enumerable. This is not
258 * considered an error. */
260 dp_enumerate_names(const char *type, struct sset *names)
262 struct registered_dpif_class *registered_class;
263 const struct dpif_class *dpif_class;
269 registered_class = dp_class_lookup(type);
270 if (!registered_class) {
271 VLOG_WARN("could not enumerate unknown type: %s", type);
275 dpif_class = registered_class->dpif_class;
276 error = dpif_class->enumerate ? dpif_class->enumerate(names) : 0;
278 VLOG_WARN("failed to enumerate %s datapaths: %s", dpif_class->type,
279 ovs_strerror(error));
281 dp_class_unref(registered_class);
286 /* Parses 'datapath_name_', which is of the form [type@]name into its
287 * component pieces. 'name' and 'type' must be freed by the caller.
289 * The returned 'type' is normalized, as if by dpif_normalize_type(). */
291 dp_parse_name(const char *datapath_name_, char **name, char **type)
293 char *datapath_name = xstrdup(datapath_name_);
296 separator = strchr(datapath_name, '@');
299 *type = datapath_name;
300 *name = xstrdup(dpif_normalize_type(separator + 1));
302 *name = datapath_name;
303 *type = xstrdup(dpif_normalize_type(NULL));
308 do_open(const char *name, const char *type, bool create, struct dpif **dpifp)
310 struct dpif *dpif = NULL;
312 struct registered_dpif_class *registered_class;
316 type = dpif_normalize_type(type);
317 registered_class = dp_class_lookup(type);
318 if (!registered_class) {
319 VLOG_WARN("could not create datapath %s of unknown type %s", name,
321 error = EAFNOSUPPORT;
325 error = registered_class->dpif_class->open(registered_class->dpif_class,
326 name, create, &dpif);
328 ovs_assert(dpif->dpif_class == registered_class->dpif_class);
330 dp_class_unref(registered_class);
334 *dpifp = error ? NULL : dpif;
338 /* Tries to open an existing datapath named 'name' and type 'type'. Will fail
339 * if no datapath with 'name' and 'type' exists. 'type' may be either NULL or
340 * the empty string to specify the default system type. Returns 0 if
341 * successful, otherwise a positive errno value. On success stores a pointer
342 * to the datapath in '*dpifp', otherwise a null pointer. */
344 dpif_open(const char *name, const char *type, struct dpif **dpifp)
346 return do_open(name, type, false, dpifp);
349 /* Tries to create and open a new datapath with the given 'name' and 'type'.
350 * 'type' may be either NULL or the empty string to specify the default system
351 * type. Will fail if a datapath with 'name' and 'type' already exists.
352 * Returns 0 if successful, otherwise a positive errno value. On success
353 * stores a pointer to the datapath in '*dpifp', otherwise a null pointer. */
355 dpif_create(const char *name, const char *type, struct dpif **dpifp)
357 return do_open(name, type, true, dpifp);
360 /* Tries to open a datapath with the given 'name' and 'type', creating it if it
361 * does not exist. 'type' may be either NULL or the empty string to specify
362 * the default system type. Returns 0 if successful, otherwise a positive
363 * errno value. On success stores a pointer to the datapath in '*dpifp',
364 * otherwise a null pointer. */
366 dpif_create_and_open(const char *name, const char *type, struct dpif **dpifp)
370 error = dpif_create(name, type, dpifp);
371 if (error == EEXIST || error == EBUSY) {
372 error = dpif_open(name, type, dpifp);
374 VLOG_WARN("datapath %s already exists but cannot be opened: %s",
375 name, ovs_strerror(error));
378 VLOG_WARN("failed to create datapath %s: %s",
379 name, ovs_strerror(error));
384 /* Closes and frees the connection to 'dpif'. Does not destroy the datapath
385 * itself; call dpif_delete() first, instead, if that is desirable. */
387 dpif_close(struct dpif *dpif)
390 struct registered_dpif_class *rc;
392 rc = shash_find_data(&dpif_classes, dpif->dpif_class->type);
393 dpif_uninit(dpif, true);
398 /* Performs periodic work needed by 'dpif'. */
400 dpif_run(struct dpif *dpif)
402 if (dpif->dpif_class->run) {
403 dpif->dpif_class->run(dpif);
407 /* Arranges for poll_block() to wake up when dp_run() needs to be called for
410 dpif_wait(struct dpif *dpif)
412 if (dpif->dpif_class->wait) {
413 dpif->dpif_class->wait(dpif);
417 /* Returns the name of datapath 'dpif' prefixed with the type
418 * (for use in log messages). */
420 dpif_name(const struct dpif *dpif)
422 return dpif->full_name;
425 /* Returns the name of datapath 'dpif' without the type
426 * (for use in device names). */
428 dpif_base_name(const struct dpif *dpif)
430 return dpif->base_name;
433 /* Returns the type of datapath 'dpif'. */
435 dpif_type(const struct dpif *dpif)
437 return dpif->dpif_class->type;
440 /* Returns the fully spelled out name for the given datapath 'type'.
442 * Normalized type string can be compared with strcmp(). Unnormalized type
443 * string might be the same even if they have different spellings. */
445 dpif_normalize_type(const char *type)
447 return type && type[0] ? type : "system";
450 /* Destroys the datapath that 'dpif' is connected to, first removing all of its
451 * ports. After calling this function, it does not make sense to pass 'dpif'
452 * to any functions other than dpif_name() or dpif_close(). */
454 dpif_delete(struct dpif *dpif)
458 COVERAGE_INC(dpif_destroy);
460 error = dpif->dpif_class->destroy(dpif);
461 log_operation(dpif, "delete", error);
465 /* Retrieves statistics for 'dpif' into 'stats'. Returns 0 if successful,
466 * otherwise a positive errno value. */
468 dpif_get_dp_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
470 int error = dpif->dpif_class->get_stats(dpif, stats);
472 memset(stats, 0, sizeof *stats);
474 log_operation(dpif, "get_stats", error);
479 dpif_port_open_type(const char *datapath_type, const char *port_type)
481 struct registered_dpif_class *rc;
483 datapath_type = dpif_normalize_type(datapath_type);
485 ovs_mutex_lock(&dpif_mutex);
486 rc = shash_find_data(&dpif_classes, datapath_type);
487 if (rc && rc->dpif_class->port_open_type) {
488 port_type = rc->dpif_class->port_open_type(rc->dpif_class, port_type);
490 ovs_mutex_unlock(&dpif_mutex);
495 /* Attempts to add 'netdev' as a port on 'dpif'. If 'port_nop' is
496 * non-null and its value is not ODPP_NONE, then attempts to use the
497 * value as the port number.
499 * If successful, returns 0 and sets '*port_nop' to the new port's port
500 * number (if 'port_nop' is non-null). On failure, returns a positive
501 * errno value and sets '*port_nop' to ODPP_NONE (if 'port_nop' is
504 dpif_port_add(struct dpif *dpif, struct netdev *netdev, odp_port_t *port_nop)
506 const char *netdev_name = netdev_get_name(netdev);
507 odp_port_t port_no = ODPP_NONE;
510 COVERAGE_INC(dpif_port_add);
516 error = dpif->dpif_class->port_add(dpif, netdev, &port_no);
518 VLOG_DBG_RL(&dpmsg_rl, "%s: added %s as port %"PRIu32,
519 dpif_name(dpif), netdev_name, port_no);
521 VLOG_WARN_RL(&error_rl, "%s: failed to add %s as port: %s",
522 dpif_name(dpif), netdev_name, ovs_strerror(error));
531 /* Attempts to remove 'dpif''s port number 'port_no'. Returns 0 if successful,
532 * otherwise a positive errno value. */
534 dpif_port_del(struct dpif *dpif, odp_port_t port_no)
538 COVERAGE_INC(dpif_port_del);
540 error = dpif->dpif_class->port_del(dpif, port_no);
542 VLOG_DBG_RL(&dpmsg_rl, "%s: port_del(%"PRIu32")",
543 dpif_name(dpif), port_no);
545 log_operation(dpif, "port_del", error);
550 /* Makes a deep copy of 'src' into 'dst'. */
552 dpif_port_clone(struct dpif_port *dst, const struct dpif_port *src)
554 dst->name = xstrdup(src->name);
555 dst->type = xstrdup(src->type);
556 dst->port_no = src->port_no;
559 /* Frees memory allocated to members of 'dpif_port'.
561 * Do not call this function on a dpif_port obtained from
562 * dpif_port_dump_next(): that function retains ownership of the data in the
565 dpif_port_destroy(struct dpif_port *dpif_port)
567 free(dpif_port->name);
568 free(dpif_port->type);
571 /* Checks if port named 'devname' exists in 'dpif'. If so, returns
572 * true; otherwise, returns false. */
574 dpif_port_exists(const struct dpif *dpif, const char *devname)
576 int error = dpif->dpif_class->port_query_by_name(dpif, devname, NULL);
577 if (error != 0 && error != ENOENT && error != ENODEV) {
578 VLOG_WARN_RL(&error_rl, "%s: failed to query port %s: %s",
579 dpif_name(dpif), devname, ovs_strerror(error));
585 /* Looks up port number 'port_no' in 'dpif'. On success, returns 0 and
586 * initializes '*port' appropriately; on failure, returns a positive errno
589 * The caller owns the data in 'port' and must free it with
590 * dpif_port_destroy() when it is no longer needed. */
592 dpif_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
593 struct dpif_port *port)
595 int error = dpif->dpif_class->port_query_by_number(dpif, port_no, port);
597 VLOG_DBG_RL(&dpmsg_rl, "%s: port %"PRIu32" is device %s",
598 dpif_name(dpif), port_no, port->name);
600 memset(port, 0, sizeof *port);
601 VLOG_WARN_RL(&error_rl, "%s: failed to query port %"PRIu32": %s",
602 dpif_name(dpif), port_no, ovs_strerror(error));
607 /* Looks up port named 'devname' in 'dpif'. On success, returns 0 and
608 * initializes '*port' appropriately; on failure, returns a positive errno
611 * The caller owns the data in 'port' and must free it with
612 * dpif_port_destroy() when it is no longer needed. */
614 dpif_port_query_by_name(const struct dpif *dpif, const char *devname,
615 struct dpif_port *port)
617 int error = dpif->dpif_class->port_query_by_name(dpif, devname, port);
619 VLOG_DBG_RL(&dpmsg_rl, "%s: device %s is on port %"PRIu32,
620 dpif_name(dpif), devname, port->port_no);
622 memset(port, 0, sizeof *port);
624 /* For ENOENT or ENODEV we use DBG level because the caller is probably
625 * interested in whether 'dpif' actually has a port 'devname', so that
626 * it's not an issue worth logging if it doesn't. Other errors are
627 * uncommon and more likely to indicate a real problem. */
629 error == ENOENT || error == ENODEV ? VLL_DBG : VLL_WARN,
630 "%s: failed to query port %s: %s",
631 dpif_name(dpif), devname, ovs_strerror(error));
636 /* Returns the Netlink PID value to supply in OVS_ACTION_ATTR_USERSPACE actions
637 * as the OVS_USERSPACE_ATTR_PID attribute's value, for use in flows whose
638 * packets arrived on port 'port_no'.
640 * A 'port_no' of ODPP_NONE is a special case: it returns a reserved PID, not
641 * allocated to any port, that the client may use for special purposes.
643 * The return value is only meaningful when DPIF_UC_ACTION has been enabled in
644 * the 'dpif''s listen mask. It is allowed to change when DPIF_UC_ACTION is
645 * disabled and then re-enabled, so a client that does that must be prepared to
646 * update all of the flows that it installed that contain
647 * OVS_ACTION_ATTR_USERSPACE actions. */
649 dpif_port_get_pid(const struct dpif *dpif, odp_port_t port_no)
651 return (dpif->dpif_class->port_get_pid
652 ? (dpif->dpif_class->port_get_pid)(dpif, port_no)
656 /* Looks up port number 'port_no' in 'dpif'. On success, returns 0 and copies
657 * the port's name into the 'name_size' bytes in 'name', ensuring that the
658 * result is null-terminated. On failure, returns a positive errno value and
659 * makes 'name' the empty string. */
661 dpif_port_get_name(struct dpif *dpif, odp_port_t port_no,
662 char *name, size_t name_size)
664 struct dpif_port port;
667 ovs_assert(name_size > 0);
669 error = dpif_port_query_by_number(dpif, port_no, &port);
671 ovs_strlcpy(name, port.name, name_size);
672 dpif_port_destroy(&port);
679 /* Initializes 'dump' to begin dumping the ports in a dpif.
681 * This function provides no status indication. An error status for the entire
682 * dump operation is provided when it is completed by calling
683 * dpif_port_dump_done().
686 dpif_port_dump_start(struct dpif_port_dump *dump, const struct dpif *dpif)
689 dump->error = dpif->dpif_class->port_dump_start(dpif, &dump->state);
690 log_operation(dpif, "port_dump_start", dump->error);
693 /* Attempts to retrieve another port from 'dump', which must have been
694 * initialized with dpif_port_dump_start(). On success, stores a new dpif_port
695 * into 'port' and returns true. On failure, returns false.
697 * Failure might indicate an actual error or merely that the last port has been
698 * dumped. An error status for the entire dump operation is provided when it
699 * is completed by calling dpif_port_dump_done().
701 * The dpif owns the data stored in 'port'. It will remain valid until at
702 * least the next time 'dump' is passed to dpif_port_dump_next() or
703 * dpif_port_dump_done(). */
705 dpif_port_dump_next(struct dpif_port_dump *dump, struct dpif_port *port)
707 const struct dpif *dpif = dump->dpif;
713 dump->error = dpif->dpif_class->port_dump_next(dpif, dump->state, port);
714 if (dump->error == EOF) {
715 VLOG_DBG_RL(&dpmsg_rl, "%s: dumped all ports", dpif_name(dpif));
717 log_operation(dpif, "port_dump_next", dump->error);
721 dpif->dpif_class->port_dump_done(dpif, dump->state);
727 /* Completes port table dump operation 'dump', which must have been initialized
728 * with dpif_port_dump_start(). Returns 0 if the dump operation was
729 * error-free, otherwise a positive errno value describing the problem. */
731 dpif_port_dump_done(struct dpif_port_dump *dump)
733 const struct dpif *dpif = dump->dpif;
735 dump->error = dpif->dpif_class->port_dump_done(dpif, dump->state);
736 log_operation(dpif, "port_dump_done", dump->error);
738 return dump->error == EOF ? 0 : dump->error;
741 /* Polls for changes in the set of ports in 'dpif'. If the set of ports in
742 * 'dpif' has changed, this function does one of the following:
744 * - Stores the name of the device that was added to or deleted from 'dpif' in
745 * '*devnamep' and returns 0. The caller is responsible for freeing
746 * '*devnamep' (with free()) when it no longer needs it.
748 * - Returns ENOBUFS and sets '*devnamep' to NULL.
750 * This function may also return 'false positives', where it returns 0 and
751 * '*devnamep' names a device that was not actually added or deleted or it
752 * returns ENOBUFS without any change.
754 * Returns EAGAIN if the set of ports in 'dpif' has not changed. May also
755 * return other positive errno values to indicate that something has gone
758 dpif_port_poll(const struct dpif *dpif, char **devnamep)
760 int error = dpif->dpif_class->port_poll(dpif, devnamep);
767 /* Arranges for the poll loop to wake up when port_poll(dpif) will return a
768 * value other than EAGAIN. */
770 dpif_port_poll_wait(const struct dpif *dpif)
772 dpif->dpif_class->port_poll_wait(dpif);
775 /* Extracts the flow stats for a packet. The 'flow' and 'packet'
776 * arguments must have been initialized through a call to flow_extract().
777 * 'used' is stored into stats->used. */
779 dpif_flow_stats_extract(const struct flow *flow, const struct ofpbuf *packet,
780 long long int used, struct dpif_flow_stats *stats)
782 stats->tcp_flags = packet_get_tcp_flags(packet, flow);
783 stats->n_bytes = packet->size;
784 stats->n_packets = 1;
788 /* Appends a human-readable representation of 'stats' to 's'. */
790 dpif_flow_stats_format(const struct dpif_flow_stats *stats, struct ds *s)
792 ds_put_format(s, "packets:%"PRIu64", bytes:%"PRIu64", used:",
793 stats->n_packets, stats->n_bytes);
795 ds_put_format(s, "%.3fs", (time_msec() - stats->used) / 1000.0);
797 ds_put_format(s, "never");
799 if (stats->tcp_flags) {
800 ds_put_cstr(s, ", flags:");
801 packet_format_tcp_flags(s, stats->tcp_flags);
805 /* Deletes all flows from 'dpif'. Returns 0 if successful, otherwise a
806 * positive errno value. */
808 dpif_flow_flush(struct dpif *dpif)
812 COVERAGE_INC(dpif_flow_flush);
814 error = dpif->dpif_class->flow_flush(dpif);
815 log_operation(dpif, "flow_flush", error);
819 /* Queries 'dpif' for a flow entry. The flow is specified by the Netlink
820 * attributes with types OVS_KEY_ATTR_* in the 'key_len' bytes starting at
823 * Returns 0 if successful. If no flow matches, returns ENOENT. On other
824 * failure, returns a positive errno value.
826 * If 'actionsp' is nonnull, then on success '*actionsp' will be set to an
827 * ofpbuf owned by the caller that contains the Netlink attributes for the
828 * flow's actions. The caller must free the ofpbuf (with ofpbuf_delete()) when
829 * it is no longer needed.
831 * If 'stats' is nonnull, then on success it will be updated with the flow's
834 dpif_flow_get(const struct dpif *dpif,
835 const struct nlattr *key, size_t key_len,
836 struct ofpbuf **actionsp, struct dpif_flow_stats *stats)
840 COVERAGE_INC(dpif_flow_get);
842 error = dpif->dpif_class->flow_get(dpif, key, key_len, actionsp, stats);
848 memset(stats, 0, sizeof *stats);
851 if (should_log_flow_message(error)) {
852 const struct nlattr *actions;
855 if (!error && actionsp) {
856 actions = (*actionsp)->data;
857 actions_len = (*actionsp)->size;
862 log_flow_message(dpif, error, "flow_get", key, key_len,
863 NULL, 0, stats, actions, actions_len);
869 dpif_flow_put__(struct dpif *dpif, const struct dpif_flow_put *put)
873 COVERAGE_INC(dpif_flow_put);
874 ovs_assert(!(put->flags & ~(DPIF_FP_CREATE | DPIF_FP_MODIFY
875 | DPIF_FP_ZERO_STATS)));
877 error = dpif->dpif_class->flow_put(dpif, put);
878 if (error && put->stats) {
879 memset(put->stats, 0, sizeof *put->stats);
881 log_flow_put_message(dpif, put, error);
885 /* Adds or modifies a flow in 'dpif'. The flow is specified by the Netlink
886 * attribute OVS_FLOW_ATTR_KEY with types OVS_KEY_ATTR_* in the 'key_len' bytes
887 * starting at 'key', and OVS_FLOW_ATTR_MASK with types of OVS_KEY_ATTR_* in
888 * the 'mask_len' bytes starting at 'mask'. The associated actions are
889 * specified by the Netlink attributes with types OVS_ACTION_ATTR_* in the
890 * 'actions_len' bytes starting at 'actions'.
892 * - If the flow's key does not exist in 'dpif', then the flow will be added if
893 * 'flags' includes DPIF_FP_CREATE. Otherwise the operation will fail with
896 * The datapath may reject attempts to insert overlapping flows with EINVAL
897 * or EEXIST, but clients should not rely on this: avoiding overlapping flows
898 * is primarily the client's responsibility.
900 * If the operation succeeds, then 'stats', if nonnull, will be zeroed.
902 * - If the flow's key does exist in 'dpif', then the flow's actions will be
903 * updated if 'flags' includes DPIF_FP_MODIFY. Otherwise the operation will
904 * fail with EEXIST. If the flow's actions are updated, then its statistics
905 * will be zeroed if 'flags' includes DPIF_FP_ZERO_STATS, and left as-is
908 * If the operation succeeds, then 'stats', if nonnull, will be set to the
909 * flow's statistics before the update.
912 dpif_flow_put(struct dpif *dpif, enum dpif_flow_put_flags flags,
913 const struct nlattr *key, size_t key_len,
914 const struct nlattr *mask, size_t mask_len,
915 const struct nlattr *actions, size_t actions_len,
916 struct dpif_flow_stats *stats)
918 struct dpif_flow_put put;
922 put.key_len = key_len;
924 put.mask_len = mask_len;
925 put.actions = actions;
926 put.actions_len = actions_len;
928 return dpif_flow_put__(dpif, &put);
932 dpif_flow_del__(struct dpif *dpif, struct dpif_flow_del *del)
936 COVERAGE_INC(dpif_flow_del);
938 error = dpif->dpif_class->flow_del(dpif, del);
939 if (error && del->stats) {
940 memset(del->stats, 0, sizeof *del->stats);
942 log_flow_del_message(dpif, del, error);
946 /* Deletes a flow from 'dpif' and returns 0, or returns ENOENT if 'dpif' does
947 * not contain such a flow. The flow is specified by the Netlink attributes
948 * with types OVS_KEY_ATTR_* in the 'key_len' bytes starting at 'key'.
950 * If the operation succeeds, then 'stats', if nonnull, will be set to the
951 * flow's statistics before its deletion. */
953 dpif_flow_del(struct dpif *dpif,
954 const struct nlattr *key, size_t key_len,
955 struct dpif_flow_stats *stats)
957 struct dpif_flow_del del;
960 del.key_len = key_len;
962 return dpif_flow_del__(dpif, &del);
965 /* Allocates thread-local state for use with the 'flow_dump_next' function for
966 * 'dpif'. On return, initializes '*statep' with any private data needed for
969 dpif_flow_dump_state_init(const struct dpif *dpif, void **statep)
971 dpif->dpif_class->flow_dump_state_init(statep);
974 /* Releases 'state' which was initialized by a call to the
975 * 'flow_dump_state_init' function for 'dpif'. */
977 dpif_flow_dump_state_uninit(const struct dpif *dpif, void *state)
979 dpif->dpif_class->flow_dump_state_uninit(state);
982 /* Initializes 'dump' to begin dumping the flows in a dpif. On sucess,
983 * initializes 'dump' with any data needed for iteration and returns 0.
984 * Otherwise, returns a positive errno value describing the problem. */
986 dpif_flow_dump_start(struct dpif_flow_dump *dump, const struct dpif *dpif)
990 error = dpif->dpif_class->flow_dump_start(dpif, &dump->iter);
991 log_operation(dpif, "flow_dump_start", error);
995 /* Attempts to retrieve another flow from 'dump', using 'state' for
996 * thread-local storage. 'dump' must have been initialized with a successful
997 * call to dpif_flow_dump_start(), and 'state' must have been initialized with
998 * dpif_flow_state_init().
1000 * On success, updates the output parameters as described below and returns
1001 * true. Otherwise, returns false. Failure might indicate an actual error or
1002 * merely the end of the flow table. An error status for the entire dump
1003 * operation is provided when it is completed by calling dpif_flow_dump_done().
1004 * Multiple threads may use the same 'dump' with this function, but all other
1005 * parameters must not be shared.
1007 * On success, if 'key' and 'key_len' are nonnull then '*key' and '*key_len'
1008 * will be set to Netlink attributes with types OVS_KEY_ATTR_* representing the
1009 * dumped flow's key. If 'actions' and 'actions_len' are nonnull then they are
1010 * set to Netlink attributes with types OVS_ACTION_ATTR_* representing the
1011 * dumped flow's actions. If 'stats' is nonnull then it will be set to the
1012 * dumped flow's statistics.
1014 * All of the returned data is owned by 'dpif', not by the caller, and the
1015 * caller must not modify or free it. 'dpif' guarantees that it remains
1016 * accessible and unchanging until at least the next call to 'flow_dump_next'
1017 * or 'flow_dump_done' for 'dump' and 'state'. */
1019 dpif_flow_dump_next(struct dpif_flow_dump *dump, void *state,
1020 const struct nlattr **key, size_t *key_len,
1021 const struct nlattr **mask, size_t *mask_len,
1022 const struct nlattr **actions, size_t *actions_len,
1023 const struct dpif_flow_stats **stats)
1025 const struct dpif *dpif = dump->dpif;
1028 error = dpif->dpif_class->flow_dump_next(dpif, dump->iter, state,
1029 key, key_len, mask, mask_len,
1030 actions, actions_len, stats);
1049 VLOG_DBG_RL(&dpmsg_rl, "%s: dumped all flows", dpif_name(dpif));
1050 } else if (should_log_flow_message(error)) {
1051 log_flow_message(dpif, error, "flow_dump",
1052 key ? *key : NULL, key ? *key_len : 0,
1053 mask ? *mask : NULL, mask ? *mask_len : 0,
1054 stats ? *stats : NULL, actions ? *actions : NULL,
1055 actions ? *actions_len : 0);
1060 /* Determines whether the next call to 'dpif_flow_dump_next' for 'dump' and
1061 * 'state' will modify or free the keys that it previously returned. 'state'
1062 * must have been initialized by a call to 'dpif_flow_dump_state_init' for
1065 * 'dpif' guarantees that data returned by flow_dump_next() will remain
1066 * accessible and unchanging until the next call. This function provides a way
1067 * for callers to determine whether that guarantee extends beyond the next
1070 * Returns true if the next call to flow_dump_next() is expected to be
1071 * destructive to previously returned keys for 'state', false otherwise. */
1073 dpif_flow_dump_next_may_destroy_keys(struct dpif_flow_dump *dump, void *state)
1075 const struct dpif *dpif = dump->dpif;
1076 return (dpif->dpif_class->flow_dump_next_may_destroy_keys
1077 ? dpif->dpif_class->flow_dump_next_may_destroy_keys(state)
1081 /* Completes flow table dump operation 'dump', which must have been initialized
1082 * with a successful call to dpif_flow_dump_start(). Returns 0 if the dump
1083 * operation was error-free, otherwise a positive errno value describing the
1086 dpif_flow_dump_done(struct dpif_flow_dump *dump)
1088 const struct dpif *dpif = dump->dpif;
1089 int error = dpif->dpif_class->flow_dump_done(dpif, dump->iter);
1090 log_operation(dpif, "flow_dump_done", error);
1091 return error == EOF ? 0 : error;
1094 struct dpif_execute_helper_aux {
1099 /* This is called for actions that need the context of the datapath to be
1102 dpif_execute_helper_cb(void *aux_, struct ofpbuf *packet,
1103 const struct pkt_metadata *md,
1104 const struct nlattr *action, bool may_steal OVS_UNUSED)
1106 struct dpif_execute_helper_aux *aux = aux_;
1107 struct dpif_execute execute;
1108 int type = nl_attr_type(action);
1110 switch ((enum ovs_action_attr)type) {
1111 case OVS_ACTION_ATTR_OUTPUT:
1112 case OVS_ACTION_ATTR_USERSPACE:
1113 execute.actions = action;
1114 execute.actions_len = NLA_ALIGN(action->nla_len);
1115 execute.packet = packet;
1117 execute.needs_help = false;
1118 aux->error = aux->dpif->dpif_class->execute(aux->dpif, &execute);
1121 case OVS_ACTION_ATTR_PUSH_VLAN:
1122 case OVS_ACTION_ATTR_POP_VLAN:
1123 case OVS_ACTION_ATTR_PUSH_MPLS:
1124 case OVS_ACTION_ATTR_POP_MPLS:
1125 case OVS_ACTION_ATTR_SET:
1126 case OVS_ACTION_ATTR_SAMPLE:
1127 case OVS_ACTION_ATTR_UNSPEC:
1128 case __OVS_ACTION_ATTR_MAX:
1133 /* Executes 'execute' by performing most of the actions in userspace and
1134 * passing the fully constructed packets to 'dpif' for output and userspace
1137 * This helps with actions that a given 'dpif' doesn't implement directly. */
1139 dpif_execute_with_help(struct dpif *dpif, struct dpif_execute *execute)
1141 struct dpif_execute_helper_aux aux = {dpif, 0};
1143 COVERAGE_INC(dpif_execute_with_help);
1145 odp_execute_actions(&aux, execute->packet, &execute->md,
1146 execute->actions, execute->actions_len,
1147 dpif_execute_helper_cb);
1151 /* Causes 'dpif' to perform the 'execute->actions_len' bytes of actions in
1152 * 'execute->actions' on the Ethernet frame in 'execute->packet' and on packet
1153 * metadata in 'execute->md'. The implementation is allowed to modify both the
1154 * '*execute->packet' and 'execute->md'.
1156 * Some dpif providers do not implement every action. The Linux kernel
1157 * datapath, in particular, does not implement ARP field modification. If
1158 * 'needs_help' is true, the dpif layer executes in userspace all of the
1159 * actions that it can, and for OVS_ACTION_ATTR_OUTPUT and
1160 * OVS_ACTION_ATTR_USERSPACE actions it passes the packet through to the dpif
1163 * This works even if 'execute->actions_len' is too long for a Netlink
1166 * Returns 0 if successful, otherwise a positive errno value. */
1168 dpif_execute(struct dpif *dpif, struct dpif_execute *execute)
1172 COVERAGE_INC(dpif_execute);
1173 if (execute->actions_len > 0) {
1174 error = (execute->needs_help || nl_attr_oversized(execute->actions_len)
1175 ? dpif_execute_with_help(dpif, execute)
1176 : dpif->dpif_class->execute(dpif, execute));
1181 log_execute_message(dpif, execute, error);
1186 /* Executes each of the 'n_ops' operations in 'ops' on 'dpif', in the order in
1187 * which they are specified, placing each operation's results in the "output"
1188 * members documented in comments.
1190 * This function exists because some datapaths can perform batched operations
1191 * faster than individual operations. */
1193 dpif_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
1195 if (dpif->dpif_class->operate) {
1199 /* Count 'chunk', the number of ops that can be executed without
1200 * needing any help. Ops that need help should be rare, so we
1201 * expect this to ordinarily be 'n_ops', that is, all the ops. */
1202 for (chunk = 0; chunk < n_ops; chunk++) {
1203 struct dpif_op *op = ops[chunk];
1205 if (op->type == DPIF_OP_EXECUTE && op->u.execute.needs_help) {
1211 /* Execute a chunk full of ops that the dpif provider can
1212 * handle itself, without help. */
1215 dpif->dpif_class->operate(dpif, ops, chunk);
1217 for (i = 0; i < chunk; i++) {
1218 struct dpif_op *op = ops[i];
1221 case DPIF_OP_FLOW_PUT:
1222 log_flow_put_message(dpif, &op->u.flow_put, op->error);
1225 case DPIF_OP_FLOW_DEL:
1226 log_flow_del_message(dpif, &op->u.flow_del, op->error);
1229 case DPIF_OP_EXECUTE:
1230 log_execute_message(dpif, &op->u.execute, op->error);
1238 /* Help the dpif provider to execute one op. */
1239 struct dpif_op *op = ops[0];
1241 op->error = dpif_execute(dpif, &op->u.execute);
1249 for (i = 0; i < n_ops; i++) {
1250 struct dpif_op *op = ops[i];
1253 case DPIF_OP_FLOW_PUT:
1254 op->error = dpif_flow_put__(dpif, &op->u.flow_put);
1257 case DPIF_OP_FLOW_DEL:
1258 op->error = dpif_flow_del__(dpif, &op->u.flow_del);
1261 case DPIF_OP_EXECUTE:
1262 op->error = dpif_execute(dpif, &op->u.execute);
1272 /* Returns a string that represents 'type', for use in log messages. */
1274 dpif_upcall_type_to_string(enum dpif_upcall_type type)
1277 case DPIF_UC_MISS: return "miss";
1278 case DPIF_UC_ACTION: return "action";
1279 case DPIF_N_UC_TYPES: default: return "<unknown>";
1283 /* Enables or disables receiving packets with dpif_recv() on 'dpif'. Returns 0
1284 * if successful, otherwise a positive errno value.
1286 * Turning packet receive off and then back on may change the Netlink PID
1287 * assignments returned by dpif_port_get_pid(). If the client does this, it
1288 * must update all of the flows that have OVS_ACTION_ATTR_USERSPACE actions
1289 * using the new PID assignment. */
1291 dpif_recv_set(struct dpif *dpif, bool enable)
1293 int error = dpif->dpif_class->recv_set(dpif, enable);
1294 log_operation(dpif, "recv_set", error);
1298 /* Polls for an upcall from 'dpif'. If successful, stores the upcall into
1299 * '*upcall', using 'buf' for storage. Should only be called if
1300 * dpif_recv_set() has been used to enable receiving packets on 'dpif'.
1302 * 'upcall->key' and 'upcall->userdata' point into data in the caller-provided
1303 * 'buf', so their memory cannot be freed separately from 'buf'.
1305 * The caller owns the data of 'upcall->packet' and may modify it. If
1306 * packet's headroom is exhausted as it is manipulated, 'upcall->packet'
1307 * will be reallocated. This requires the data of 'upcall->packet' to be
1308 * released with ofpbuf_uninit() before 'upcall' is destroyed. However,
1309 * when an error is returned, the 'upcall->packet' may be uninitialized
1310 * and should not be released.
1312 * Returns 0 if successful, otherwise a positive errno value. Returns EAGAIN
1313 * if no upcall is immediately available. */
1315 dpif_recv(struct dpif *dpif, struct dpif_upcall *upcall, struct ofpbuf *buf)
1317 int error = dpif->dpif_class->recv(dpif, upcall, buf);
1318 if (!error && !VLOG_DROP_DBG(&dpmsg_rl)) {
1322 packet = ofp_packet_to_string(upcall->packet.data,
1323 upcall->packet.size);
1326 odp_flow_key_format(upcall->key, upcall->key_len, &flow);
1328 VLOG_DBG("%s: %s upcall:\n%s\n%s",
1329 dpif_name(dpif), dpif_upcall_type_to_string(upcall->type),
1330 ds_cstr(&flow), packet);
1334 } else if (error && error != EAGAIN) {
1335 log_operation(dpif, "recv", error);
1340 /* Discards all messages that would otherwise be received by dpif_recv() on
1343 dpif_recv_purge(struct dpif *dpif)
1345 COVERAGE_INC(dpif_purge);
1346 if (dpif->dpif_class->recv_purge) {
1347 dpif->dpif_class->recv_purge(dpif);
1351 /* Arranges for the poll loop to wake up when 'dpif' has a message queued to be
1352 * received with dpif_recv(). */
1354 dpif_recv_wait(struct dpif *dpif)
1356 dpif->dpif_class->recv_wait(dpif);
1359 /* Obtains the NetFlow engine type and engine ID for 'dpif' into '*engine_type'
1360 * and '*engine_id', respectively. */
1362 dpif_get_netflow_ids(const struct dpif *dpif,
1363 uint8_t *engine_type, uint8_t *engine_id)
1365 *engine_type = dpif->netflow_engine_type;
1366 *engine_id = dpif->netflow_engine_id;
1369 /* Translates OpenFlow queue ID 'queue_id' (in host byte order) into a priority
1370 * value used for setting packet priority.
1371 * On success, returns 0 and stores the priority into '*priority'.
1372 * On failure, returns a positive errno value and stores 0 into '*priority'. */
1374 dpif_queue_to_priority(const struct dpif *dpif, uint32_t queue_id,
1377 int error = (dpif->dpif_class->queue_to_priority
1378 ? dpif->dpif_class->queue_to_priority(dpif, queue_id,
1384 log_operation(dpif, "queue_to_priority", error);
1389 dpif_init(struct dpif *dpif, const struct dpif_class *dpif_class,
1391 uint8_t netflow_engine_type, uint8_t netflow_engine_id)
1393 dpif->dpif_class = dpif_class;
1394 dpif->base_name = xstrdup(name);
1395 dpif->full_name = xasprintf("%s@%s", dpif_class->type, name);
1396 dpif->netflow_engine_type = netflow_engine_type;
1397 dpif->netflow_engine_id = netflow_engine_id;
1400 /* Undoes the results of initialization.
1402 * Normally this function only needs to be called from dpif_close().
1403 * However, it may be called by providers due to an error on opening
1404 * that occurs after initialization. It this case dpif_close() would
1405 * never be called. */
1407 dpif_uninit(struct dpif *dpif, bool close)
1409 char *base_name = dpif->base_name;
1410 char *full_name = dpif->full_name;
1413 dpif->dpif_class->close(dpif);
1421 log_operation(const struct dpif *dpif, const char *operation, int error)
1424 VLOG_DBG_RL(&dpmsg_rl, "%s: %s success", dpif_name(dpif), operation);
1425 } else if (ofperr_is_valid(error)) {
1426 VLOG_WARN_RL(&error_rl, "%s: %s failed (%s)",
1427 dpif_name(dpif), operation, ofperr_get_name(error));
1429 VLOG_WARN_RL(&error_rl, "%s: %s failed (%s)",
1430 dpif_name(dpif), operation, ovs_strerror(error));
1434 static enum vlog_level
1435 flow_message_log_level(int error)
1437 /* If flows arrive in a batch, userspace may push down multiple
1438 * unique flow definitions that overlap when wildcards are applied.
1439 * Kernels that support flow wildcarding will reject these flows as
1440 * duplicates (EEXIST), so lower the log level to debug for these
1441 * types of messages. */
1442 return (error && error != EEXIST) ? VLL_WARN : VLL_DBG;
1446 should_log_flow_message(int error)
1448 return !vlog_should_drop(THIS_MODULE, flow_message_log_level(error),
1449 error ? &error_rl : &dpmsg_rl);
1453 log_flow_message(const struct dpif *dpif, int error, const char *operation,
1454 const struct nlattr *key, size_t key_len,
1455 const struct nlattr *mask, size_t mask_len,
1456 const struct dpif_flow_stats *stats,
1457 const struct nlattr *actions, size_t actions_len)
1459 struct ds ds = DS_EMPTY_INITIALIZER;
1460 ds_put_format(&ds, "%s: ", dpif_name(dpif));
1462 ds_put_cstr(&ds, "failed to ");
1464 ds_put_format(&ds, "%s ", operation);
1466 ds_put_format(&ds, "(%s) ", ovs_strerror(error));
1468 odp_flow_format(key, key_len, mask, mask_len, NULL, &ds, true);
1470 ds_put_cstr(&ds, ", ");
1471 dpif_flow_stats_format(stats, &ds);
1473 if (actions || actions_len) {
1474 ds_put_cstr(&ds, ", actions:");
1475 format_odp_actions(&ds, actions, actions_len);
1477 vlog(THIS_MODULE, flow_message_log_level(error), "%s", ds_cstr(&ds));
1482 log_flow_put_message(struct dpif *dpif, const struct dpif_flow_put *put,
1485 if (should_log_flow_message(error)) {
1489 ds_put_cstr(&s, "put");
1490 if (put->flags & DPIF_FP_CREATE) {
1491 ds_put_cstr(&s, "[create]");
1493 if (put->flags & DPIF_FP_MODIFY) {
1494 ds_put_cstr(&s, "[modify]");
1496 if (put->flags & DPIF_FP_ZERO_STATS) {
1497 ds_put_cstr(&s, "[zero]");
1499 log_flow_message(dpif, error, ds_cstr(&s),
1500 put->key, put->key_len, put->mask, put->mask_len,
1501 put->stats, put->actions, put->actions_len);
1507 log_flow_del_message(struct dpif *dpif, const struct dpif_flow_del *del,
1510 if (should_log_flow_message(error)) {
1511 log_flow_message(dpif, error, "flow_del", del->key, del->key_len,
1512 NULL, 0, !error ? del->stats : NULL, NULL, 0);
1517 log_execute_message(struct dpif *dpif, const struct dpif_execute *execute,
1520 if (!(error ? VLOG_DROP_WARN(&error_rl) : VLOG_DROP_DBG(&dpmsg_rl))) {
1521 struct ds ds = DS_EMPTY_INITIALIZER;
1524 packet = ofp_packet_to_string(execute->packet->data,
1525 execute->packet->size);
1526 ds_put_format(&ds, "%s: execute ", dpif_name(dpif));
1527 format_odp_actions(&ds, execute->actions, execute->actions_len);
1529 ds_put_format(&ds, " failed (%s)", ovs_strerror(error));
1531 ds_put_format(&ds, " on packet %s", packet);
1532 vlog(THIS_MODULE, error ? VLL_WARN : VLL_DBG, "%s", ds_cstr(&ds));