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
637 * actions as the OVS_USERSPACE_ATTR_PID attribute's value, for use in
638 * flows whose packets arrived on port 'port_no'. In the case where the
639 * provider allocates multiple Netlink PIDs to a single port, it may use
640 * 'hash' to spread load among them. The caller need not use a particular
641 * hash function; a 5-tuple hash is suitable.
643 * (The datapath implementation might use some different hash function for
644 * distributing packets received via flow misses among PIDs. This means
645 * that packets received via flow misses might be reordered relative to
646 * packets received via userspace actions. This is not ordinarily a
649 * A 'port_no' of ODPP_NONE is a special case: it returns a reserved PID, not
650 * allocated to any port, that the client may use for special purposes.
652 * The return value is only meaningful when DPIF_UC_ACTION has been enabled in
653 * the 'dpif''s listen mask. It is allowed to change when DPIF_UC_ACTION is
654 * disabled and then re-enabled, so a client that does that must be prepared to
655 * update all of the flows that it installed that contain
656 * OVS_ACTION_ATTR_USERSPACE actions. */
658 dpif_port_get_pid(const struct dpif *dpif, odp_port_t port_no, uint32_t hash)
660 return (dpif->dpif_class->port_get_pid
661 ? (dpif->dpif_class->port_get_pid)(dpif, port_no, hash)
665 /* Looks up port number 'port_no' in 'dpif'. On success, returns 0 and copies
666 * the port's name into the 'name_size' bytes in 'name', ensuring that the
667 * result is null-terminated. On failure, returns a positive errno value and
668 * makes 'name' the empty string. */
670 dpif_port_get_name(struct dpif *dpif, odp_port_t port_no,
671 char *name, size_t name_size)
673 struct dpif_port port;
676 ovs_assert(name_size > 0);
678 error = dpif_port_query_by_number(dpif, port_no, &port);
680 ovs_strlcpy(name, port.name, name_size);
681 dpif_port_destroy(&port);
688 /* Initializes 'dump' to begin dumping the ports in a dpif.
690 * This function provides no status indication. An error status for the entire
691 * dump operation is provided when it is completed by calling
692 * dpif_port_dump_done().
695 dpif_port_dump_start(struct dpif_port_dump *dump, const struct dpif *dpif)
698 dump->error = dpif->dpif_class->port_dump_start(dpif, &dump->state);
699 log_operation(dpif, "port_dump_start", dump->error);
702 /* Attempts to retrieve another port from 'dump', which must have been
703 * initialized with dpif_port_dump_start(). On success, stores a new dpif_port
704 * into 'port' and returns true. On failure, returns false.
706 * Failure might indicate an actual error or merely that the last port has been
707 * dumped. An error status for the entire dump operation is provided when it
708 * is completed by calling dpif_port_dump_done().
710 * The dpif owns the data stored in 'port'. It will remain valid until at
711 * least the next time 'dump' is passed to dpif_port_dump_next() or
712 * dpif_port_dump_done(). */
714 dpif_port_dump_next(struct dpif_port_dump *dump, struct dpif_port *port)
716 const struct dpif *dpif = dump->dpif;
722 dump->error = dpif->dpif_class->port_dump_next(dpif, dump->state, port);
723 if (dump->error == EOF) {
724 VLOG_DBG_RL(&dpmsg_rl, "%s: dumped all ports", dpif_name(dpif));
726 log_operation(dpif, "port_dump_next", dump->error);
730 dpif->dpif_class->port_dump_done(dpif, dump->state);
736 /* Completes port table dump operation 'dump', which must have been initialized
737 * with dpif_port_dump_start(). Returns 0 if the dump operation was
738 * error-free, otherwise a positive errno value describing the problem. */
740 dpif_port_dump_done(struct dpif_port_dump *dump)
742 const struct dpif *dpif = dump->dpif;
744 dump->error = dpif->dpif_class->port_dump_done(dpif, dump->state);
745 log_operation(dpif, "port_dump_done", dump->error);
747 return dump->error == EOF ? 0 : dump->error;
750 /* Polls for changes in the set of ports in 'dpif'. If the set of ports in
751 * 'dpif' has changed, this function does one of the following:
753 * - Stores the name of the device that was added to or deleted from 'dpif' in
754 * '*devnamep' and returns 0. The caller is responsible for freeing
755 * '*devnamep' (with free()) when it no longer needs it.
757 * - Returns ENOBUFS and sets '*devnamep' to NULL.
759 * This function may also return 'false positives', where it returns 0 and
760 * '*devnamep' names a device that was not actually added or deleted or it
761 * returns ENOBUFS without any change.
763 * Returns EAGAIN if the set of ports in 'dpif' has not changed. May also
764 * return other positive errno values to indicate that something has gone
767 dpif_port_poll(const struct dpif *dpif, char **devnamep)
769 int error = dpif->dpif_class->port_poll(dpif, devnamep);
776 /* Arranges for the poll loop to wake up when port_poll(dpif) will return a
777 * value other than EAGAIN. */
779 dpif_port_poll_wait(const struct dpif *dpif)
781 dpif->dpif_class->port_poll_wait(dpif);
784 /* Extracts the flow stats for a packet. The 'flow' and 'packet'
785 * arguments must have been initialized through a call to flow_extract().
786 * 'used' is stored into stats->used. */
788 dpif_flow_stats_extract(const struct flow *flow, const struct ofpbuf *packet,
789 long long int used, struct dpif_flow_stats *stats)
791 stats->tcp_flags = ntohs(flow->tcp_flags);
792 stats->n_bytes = ofpbuf_size(packet);
793 stats->n_packets = 1;
797 /* Appends a human-readable representation of 'stats' to 's'. */
799 dpif_flow_stats_format(const struct dpif_flow_stats *stats, struct ds *s)
801 ds_put_format(s, "packets:%"PRIu64", bytes:%"PRIu64", used:",
802 stats->n_packets, stats->n_bytes);
804 ds_put_format(s, "%.3fs", (time_msec() - stats->used) / 1000.0);
806 ds_put_format(s, "never");
808 if (stats->tcp_flags) {
809 ds_put_cstr(s, ", flags:");
810 packet_format_tcp_flags(s, stats->tcp_flags);
814 /* Deletes all flows from 'dpif'. Returns 0 if successful, otherwise a
815 * positive errno value. */
817 dpif_flow_flush(struct dpif *dpif)
821 COVERAGE_INC(dpif_flow_flush);
823 error = dpif->dpif_class->flow_flush(dpif);
824 log_operation(dpif, "flow_flush", error);
828 /* Queries 'dpif' for a flow entry. The flow is specified by the Netlink
829 * attributes with types OVS_KEY_ATTR_* in the 'key_len' bytes starting at
832 * Returns 0 if successful. If no flow matches, returns ENOENT. On other
833 * failure, returns a positive errno value.
835 * If 'actionsp' is nonnull, then on success '*actionsp' will be set to an
836 * ofpbuf owned by the caller that contains the Netlink attributes for the
837 * flow's actions. The caller must free the ofpbuf (with ofpbuf_delete()) when
838 * it is no longer needed.
840 * If 'stats' is nonnull, then on success it will be updated with the flow's
843 dpif_flow_get(const struct dpif *dpif,
844 const struct nlattr *key, size_t key_len,
845 struct ofpbuf **actionsp, struct dpif_flow_stats *stats)
849 COVERAGE_INC(dpif_flow_get);
851 error = dpif->dpif_class->flow_get(dpif, key, key_len, actionsp, stats);
857 memset(stats, 0, sizeof *stats);
860 if (should_log_flow_message(error)) {
861 const struct nlattr *actions;
864 if (!error && actionsp) {
865 actions = ofpbuf_data(*actionsp);
866 actions_len = ofpbuf_size(*actionsp);
871 log_flow_message(dpif, error, "flow_get", key, key_len,
872 NULL, 0, stats, actions, actions_len);
878 dpif_flow_put__(struct dpif *dpif, const struct dpif_flow_put *put)
882 COVERAGE_INC(dpif_flow_put);
883 ovs_assert(!(put->flags & ~(DPIF_FP_CREATE | DPIF_FP_MODIFY
884 | DPIF_FP_ZERO_STATS)));
886 error = dpif->dpif_class->flow_put(dpif, put);
887 if (error && put->stats) {
888 memset(put->stats, 0, sizeof *put->stats);
890 log_flow_put_message(dpif, put, error);
894 /* Adds or modifies a flow in 'dpif'. The flow is specified by the Netlink
895 * attribute OVS_FLOW_ATTR_KEY with types OVS_KEY_ATTR_* in the 'key_len' bytes
896 * starting at 'key', and OVS_FLOW_ATTR_MASK with types of OVS_KEY_ATTR_* in
897 * the 'mask_len' bytes starting at 'mask'. The associated actions are
898 * specified by the Netlink attributes with types OVS_ACTION_ATTR_* in the
899 * 'actions_len' bytes starting at 'actions'.
901 * - If the flow's key does not exist in 'dpif', then the flow will be added if
902 * 'flags' includes DPIF_FP_CREATE. Otherwise the operation will fail with
905 * The datapath may reject attempts to insert overlapping flows with EINVAL
906 * or EEXIST, but clients should not rely on this: avoiding overlapping flows
907 * is primarily the client's responsibility.
909 * If the operation succeeds, then 'stats', if nonnull, will be zeroed.
911 * - If the flow's key does exist in 'dpif', then the flow's actions will be
912 * updated if 'flags' includes DPIF_FP_MODIFY. Otherwise the operation will
913 * fail with EEXIST. If the flow's actions are updated, then its statistics
914 * will be zeroed if 'flags' includes DPIF_FP_ZERO_STATS, and left as-is
917 * If the operation succeeds, then 'stats', if nonnull, will be set to the
918 * flow's statistics before the update.
921 dpif_flow_put(struct dpif *dpif, enum dpif_flow_put_flags flags,
922 const struct nlattr *key, size_t key_len,
923 const struct nlattr *mask, size_t mask_len,
924 const struct nlattr *actions, size_t actions_len,
925 struct dpif_flow_stats *stats)
927 struct dpif_flow_put put;
931 put.key_len = key_len;
933 put.mask_len = mask_len;
934 put.actions = actions;
935 put.actions_len = actions_len;
937 return dpif_flow_put__(dpif, &put);
941 dpif_flow_del__(struct dpif *dpif, struct dpif_flow_del *del)
945 COVERAGE_INC(dpif_flow_del);
947 error = dpif->dpif_class->flow_del(dpif, del);
948 if (error && del->stats) {
949 memset(del->stats, 0, sizeof *del->stats);
951 log_flow_del_message(dpif, del, error);
955 /* Deletes a flow from 'dpif' and returns 0, or returns ENOENT if 'dpif' does
956 * not contain such a flow. The flow is specified by the Netlink attributes
957 * with types OVS_KEY_ATTR_* in the 'key_len' bytes starting at 'key'.
959 * If the operation succeeds, then 'stats', if nonnull, will be set to the
960 * flow's statistics before its deletion. */
962 dpif_flow_del(struct dpif *dpif,
963 const struct nlattr *key, size_t key_len,
964 struct dpif_flow_stats *stats)
966 struct dpif_flow_del del;
969 del.key_len = key_len;
971 return dpif_flow_del__(dpif, &del);
974 /* Allocates thread-local state for use with the 'flow_dump_next' function for
975 * 'dpif'. On return, initializes '*statep' with any private data needed for
978 dpif_flow_dump_state_init(const struct dpif *dpif, void **statep)
980 dpif->dpif_class->flow_dump_state_init(statep);
983 /* Releases 'state' which was initialized by a call to the
984 * 'flow_dump_state_init' function for 'dpif'. */
986 dpif_flow_dump_state_uninit(const struct dpif *dpif, void *state)
988 dpif->dpif_class->flow_dump_state_uninit(state);
991 /* Initializes 'dump' to begin dumping the flows in a dpif. On sucess,
992 * initializes 'dump' with any data needed for iteration and returns 0.
993 * Otherwise, returns a positive errno value describing the problem. */
995 dpif_flow_dump_start(struct dpif_flow_dump *dump, const struct dpif *dpif)
999 error = dpif->dpif_class->flow_dump_start(dpif, &dump->iter);
1000 log_operation(dpif, "flow_dump_start", error);
1004 /* Attempts to retrieve another flow from 'dump', using 'state' for
1005 * thread-local storage. 'dump' must have been initialized with a successful
1006 * call to dpif_flow_dump_start(), and 'state' must have been initialized with
1007 * dpif_flow_state_init().
1009 * On success, updates the output parameters as described below and returns
1010 * true. Otherwise, returns false. Failure might indicate an actual error or
1011 * merely the end of the flow table. An error status for the entire dump
1012 * operation is provided when it is completed by calling dpif_flow_dump_done().
1013 * Multiple threads may use the same 'dump' with this function, but all other
1014 * parameters must not be shared.
1016 * On success, if 'key' and 'key_len' are nonnull then '*key' and '*key_len'
1017 * will be set to Netlink attributes with types OVS_KEY_ATTR_* representing the
1018 * dumped flow's key. If 'actions' and 'actions_len' are nonnull then they are
1019 * set to Netlink attributes with types OVS_ACTION_ATTR_* representing the
1020 * dumped flow's actions. If 'stats' is nonnull then it will be set to the
1021 * dumped flow's statistics.
1023 * All of the returned data is owned by 'dpif', not by the caller, and the
1024 * caller must not modify or free it. 'dpif' guarantees that it remains
1025 * accessible and unchanging until at least the next call to 'flow_dump_next'
1026 * or 'flow_dump_done' for 'dump' and 'state'. */
1028 dpif_flow_dump_next(struct dpif_flow_dump *dump, void *state,
1029 const struct nlattr **key, size_t *key_len,
1030 const struct nlattr **mask, size_t *mask_len,
1031 const struct nlattr **actions, size_t *actions_len,
1032 const struct dpif_flow_stats **stats)
1034 const struct dpif *dpif = dump->dpif;
1037 error = dpif->dpif_class->flow_dump_next(dpif, dump->iter, state,
1038 key, key_len, mask, mask_len,
1039 actions, actions_len, stats);
1058 VLOG_DBG_RL(&dpmsg_rl, "%s: dumped all flows", dpif_name(dpif));
1059 } else if (should_log_flow_message(error)) {
1060 log_flow_message(dpif, error, "flow_dump",
1061 key ? *key : NULL, key ? *key_len : 0,
1062 mask ? *mask : NULL, mask ? *mask_len : 0,
1063 stats ? *stats : NULL, actions ? *actions : NULL,
1064 actions ? *actions_len : 0);
1069 /* Determines whether the next call to 'dpif_flow_dump_next' for 'dump' and
1070 * 'state' will modify or free the keys that it previously returned. 'state'
1071 * must have been initialized by a call to 'dpif_flow_dump_state_init' for
1074 * 'dpif' guarantees that data returned by flow_dump_next() will remain
1075 * accessible and unchanging until the next call. This function provides a way
1076 * for callers to determine whether that guarantee extends beyond the next
1079 * Returns true if the next call to flow_dump_next() is expected to be
1080 * destructive to previously returned keys for 'state', false otherwise. */
1082 dpif_flow_dump_next_may_destroy_keys(struct dpif_flow_dump *dump, void *state)
1084 const struct dpif *dpif = dump->dpif;
1085 return (dpif->dpif_class->flow_dump_next_may_destroy_keys
1086 ? dpif->dpif_class->flow_dump_next_may_destroy_keys(state)
1090 /* Completes flow table dump operation 'dump', which must have been initialized
1091 * with a successful call to dpif_flow_dump_start(). Returns 0 if the dump
1092 * operation was error-free, otherwise a positive errno value describing the
1095 dpif_flow_dump_done(struct dpif_flow_dump *dump)
1097 const struct dpif *dpif = dump->dpif;
1098 int error = dpif->dpif_class->flow_dump_done(dpif, dump->iter);
1099 log_operation(dpif, "flow_dump_done", error);
1100 return error == EOF ? 0 : error;
1103 struct dpif_execute_helper_aux {
1108 /* This is called for actions that need the context of the datapath to be
1111 dpif_execute_helper_cb(void *aux_, struct ofpbuf *packet,
1112 struct pkt_metadata *md,
1113 const struct nlattr *action, bool may_steal OVS_UNUSED)
1115 struct dpif_execute_helper_aux *aux = aux_;
1116 struct dpif_execute execute;
1117 int type = nl_attr_type(action);
1119 switch ((enum ovs_action_attr)type) {
1120 case OVS_ACTION_ATTR_OUTPUT:
1121 case OVS_ACTION_ATTR_USERSPACE:
1122 execute.actions = action;
1123 execute.actions_len = NLA_ALIGN(action->nla_len);
1124 execute.packet = packet;
1126 execute.needs_help = false;
1127 aux->error = aux->dpif->dpif_class->execute(aux->dpif, &execute);
1130 case OVS_ACTION_ATTR_PUSH_VLAN:
1131 case OVS_ACTION_ATTR_POP_VLAN:
1132 case OVS_ACTION_ATTR_PUSH_MPLS:
1133 case OVS_ACTION_ATTR_POP_MPLS:
1134 case OVS_ACTION_ATTR_SET:
1135 case OVS_ACTION_ATTR_SAMPLE:
1136 case OVS_ACTION_ATTR_UNSPEC:
1137 case OVS_ACTION_ATTR_RECIRC:
1138 case OVS_ACTION_ATTR_HASH:
1139 case __OVS_ACTION_ATTR_MAX:
1144 /* Executes 'execute' by performing most of the actions in userspace and
1145 * passing the fully constructed packets to 'dpif' for output and userspace
1148 * This helps with actions that a given 'dpif' doesn't implement directly. */
1150 dpif_execute_with_help(struct dpif *dpif, struct dpif_execute *execute)
1152 struct dpif_execute_helper_aux aux = {dpif, 0};
1154 COVERAGE_INC(dpif_execute_with_help);
1156 odp_execute_actions(&aux, execute->packet, false, &execute->md,
1157 execute->actions, execute->actions_len,
1158 dpif_execute_helper_cb);
1162 /* Causes 'dpif' to perform the 'execute->actions_len' bytes of actions in
1163 * 'execute->actions' on the Ethernet frame in 'execute->packet' and on packet
1164 * metadata in 'execute->md'. The implementation is allowed to modify both the
1165 * '*execute->packet' and 'execute->md'.
1167 * Some dpif providers do not implement every action. The Linux kernel
1168 * datapath, in particular, does not implement ARP field modification. If
1169 * 'needs_help' is true, the dpif layer executes in userspace all of the
1170 * actions that it can, and for OVS_ACTION_ATTR_OUTPUT and
1171 * OVS_ACTION_ATTR_USERSPACE actions it passes the packet through to the dpif
1174 * This works even if 'execute->actions_len' is too long for a Netlink
1177 * Returns 0 if successful, otherwise a positive errno value. */
1179 dpif_execute(struct dpif *dpif, struct dpif_execute *execute)
1183 COVERAGE_INC(dpif_execute);
1184 if (execute->actions_len > 0) {
1185 error = (execute->needs_help || nl_attr_oversized(execute->actions_len)
1186 ? dpif_execute_with_help(dpif, execute)
1187 : dpif->dpif_class->execute(dpif, execute));
1192 log_execute_message(dpif, execute, error);
1197 /* Executes each of the 'n_ops' operations in 'ops' on 'dpif', in the order in
1198 * which they are specified, placing each operation's results in the "output"
1199 * members documented in comments.
1201 * This function exists because some datapaths can perform batched operations
1202 * faster than individual operations. */
1204 dpif_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
1206 if (dpif->dpif_class->operate) {
1210 /* Count 'chunk', the number of ops that can be executed without
1211 * needing any help. Ops that need help should be rare, so we
1212 * expect this to ordinarily be 'n_ops', that is, all the ops. */
1213 for (chunk = 0; chunk < n_ops; chunk++) {
1214 struct dpif_op *op = ops[chunk];
1216 if (op->type == DPIF_OP_EXECUTE && op->u.execute.needs_help) {
1222 /* Execute a chunk full of ops that the dpif provider can
1223 * handle itself, without help. */
1226 dpif->dpif_class->operate(dpif, ops, chunk);
1228 for (i = 0; i < chunk; i++) {
1229 struct dpif_op *op = ops[i];
1232 case DPIF_OP_FLOW_PUT:
1233 log_flow_put_message(dpif, &op->u.flow_put, op->error);
1236 case DPIF_OP_FLOW_DEL:
1237 log_flow_del_message(dpif, &op->u.flow_del, op->error);
1240 case DPIF_OP_EXECUTE:
1241 log_execute_message(dpif, &op->u.execute, op->error);
1249 /* Help the dpif provider to execute one op. */
1250 struct dpif_op *op = ops[0];
1252 op->error = dpif_execute(dpif, &op->u.execute);
1260 for (i = 0; i < n_ops; i++) {
1261 struct dpif_op *op = ops[i];
1264 case DPIF_OP_FLOW_PUT:
1265 op->error = dpif_flow_put__(dpif, &op->u.flow_put);
1268 case DPIF_OP_FLOW_DEL:
1269 op->error = dpif_flow_del__(dpif, &op->u.flow_del);
1272 case DPIF_OP_EXECUTE:
1273 op->error = dpif_execute(dpif, &op->u.execute);
1283 /* Returns a string that represents 'type', for use in log messages. */
1285 dpif_upcall_type_to_string(enum dpif_upcall_type type)
1288 case DPIF_UC_MISS: return "miss";
1289 case DPIF_UC_ACTION: return "action";
1290 case DPIF_N_UC_TYPES: default: return "<unknown>";
1294 /* Enables or disables receiving packets with dpif_recv() on 'dpif'. Returns 0
1295 * if successful, otherwise a positive errno value.
1297 * Turning packet receive off and then back on may change the Netlink PID
1298 * assignments returned by dpif_port_get_pid(). If the client does this, it
1299 * must update all of the flows that have OVS_ACTION_ATTR_USERSPACE actions
1300 * using the new PID assignment. */
1302 dpif_recv_set(struct dpif *dpif, bool enable)
1304 int error = dpif->dpif_class->recv_set(dpif, enable);
1305 log_operation(dpif, "recv_set", error);
1309 /* Refreshes the poll loops and Netlink sockets associated to each port,
1310 * when the number of upcall handlers (upcall receiving thread) is changed
1311 * to 'n_handlers' and receiving packets for 'dpif' is enabled by
1314 * Since multiple upcall handlers can read upcalls simultaneously from
1315 * 'dpif', each port can have multiple Netlink sockets, one per upcall
1316 * handler. So, handlers_set() is responsible for the following tasks:
1318 * When receiving upcall is enabled, extends or creates the
1319 * configuration to support:
1321 * - 'n_handlers' Netlink sockets for each port.
1323 * - 'n_handlers' poll loops, one for each upcall handler.
1325 * - registering the Netlink sockets for the same upcall handler to
1326 * the corresponding poll loop.
1328 * Returns 0 if successful, otherwise a positive errno value. */
1330 dpif_handlers_set(struct dpif *dpif, uint32_t n_handlers)
1332 int error = dpif->dpif_class->handlers_set(dpif, n_handlers);
1333 log_operation(dpif, "handlers_set", error);
1337 /* Polls for an upcall from 'dpif' for an upcall handler. Since there
1338 * there can be multiple poll loops, 'handler_id' is needed as index to
1339 * identify the corresponding poll loop. If successful, stores the upcall
1340 * into '*upcall', using 'buf' for storage. Should only be called if
1341 * 'recv_set' has been used to enable receiving packets from 'dpif'.
1343 * 'upcall->key' and 'upcall->userdata' point into data in the caller-provided
1344 * 'buf', so their memory cannot be freed separately from 'buf'.
1346 * The caller owns the data of 'upcall->packet' and may modify it. If
1347 * packet's headroom is exhausted as it is manipulated, 'upcall->packet'
1348 * will be reallocated. This requires the data of 'upcall->packet' to be
1349 * released with ofpbuf_uninit() before 'upcall' is destroyed. However,
1350 * when an error is returned, the 'upcall->packet' may be uninitialized
1351 * and should not be released.
1353 * Returns 0 if successful, otherwise a positive errno value. Returns EAGAIN
1354 * if no upcall is immediately available. */
1356 dpif_recv(struct dpif *dpif, uint32_t handler_id, struct dpif_upcall *upcall,
1359 int error = dpif->dpif_class->recv(dpif, handler_id, upcall, buf);
1360 if (!error && !VLOG_DROP_DBG(&dpmsg_rl)) {
1364 packet = ofp_packet_to_string(ofpbuf_data(&upcall->packet),
1365 ofpbuf_size(&upcall->packet));
1368 odp_flow_key_format(upcall->key, upcall->key_len, &flow);
1370 VLOG_DBG("%s: %s upcall:\n%s\n%s",
1371 dpif_name(dpif), dpif_upcall_type_to_string(upcall->type),
1372 ds_cstr(&flow), packet);
1376 } else if (error && error != EAGAIN) {
1377 log_operation(dpif, "recv", error);
1382 /* Discards all messages that would otherwise be received by dpif_recv() on
1385 dpif_recv_purge(struct dpif *dpif)
1387 COVERAGE_INC(dpif_purge);
1388 if (dpif->dpif_class->recv_purge) {
1389 dpif->dpif_class->recv_purge(dpif);
1393 /* Arranges for the poll loop for an upcall handler to wake up when 'dpif'
1394 * 'dpif' has a message queued to be received with the recv member
1395 * function. Since there can be multiple poll loops, 'handler_id' is
1396 * needed as index to identify the corresponding poll loop. */
1398 dpif_recv_wait(struct dpif *dpif, uint32_t handler_id)
1400 dpif->dpif_class->recv_wait(dpif, handler_id);
1403 /* Obtains the NetFlow engine type and engine ID for 'dpif' into '*engine_type'
1404 * and '*engine_id', respectively. */
1406 dpif_get_netflow_ids(const struct dpif *dpif,
1407 uint8_t *engine_type, uint8_t *engine_id)
1409 *engine_type = dpif->netflow_engine_type;
1410 *engine_id = dpif->netflow_engine_id;
1413 /* Translates OpenFlow queue ID 'queue_id' (in host byte order) into a priority
1414 * value used for setting packet priority.
1415 * On success, returns 0 and stores the priority into '*priority'.
1416 * On failure, returns a positive errno value and stores 0 into '*priority'. */
1418 dpif_queue_to_priority(const struct dpif *dpif, uint32_t queue_id,
1421 int error = (dpif->dpif_class->queue_to_priority
1422 ? dpif->dpif_class->queue_to_priority(dpif, queue_id,
1428 log_operation(dpif, "queue_to_priority", error);
1433 dpif_init(struct dpif *dpif, const struct dpif_class *dpif_class,
1435 uint8_t netflow_engine_type, uint8_t netflow_engine_id)
1437 dpif->dpif_class = dpif_class;
1438 dpif->base_name = xstrdup(name);
1439 dpif->full_name = xasprintf("%s@%s", dpif_class->type, name);
1440 dpif->netflow_engine_type = netflow_engine_type;
1441 dpif->netflow_engine_id = netflow_engine_id;
1444 /* Undoes the results of initialization.
1446 * Normally this function only needs to be called from dpif_close().
1447 * However, it may be called by providers due to an error on opening
1448 * that occurs after initialization. It this case dpif_close() would
1449 * never be called. */
1451 dpif_uninit(struct dpif *dpif, bool close)
1453 char *base_name = dpif->base_name;
1454 char *full_name = dpif->full_name;
1457 dpif->dpif_class->close(dpif);
1465 log_operation(const struct dpif *dpif, const char *operation, int error)
1468 VLOG_DBG_RL(&dpmsg_rl, "%s: %s success", dpif_name(dpif), operation);
1469 } else if (ofperr_is_valid(error)) {
1470 VLOG_WARN_RL(&error_rl, "%s: %s failed (%s)",
1471 dpif_name(dpif), operation, ofperr_get_name(error));
1473 VLOG_WARN_RL(&error_rl, "%s: %s failed (%s)",
1474 dpif_name(dpif), operation, ovs_strerror(error));
1478 static enum vlog_level
1479 flow_message_log_level(int error)
1481 /* If flows arrive in a batch, userspace may push down multiple
1482 * unique flow definitions that overlap when wildcards are applied.
1483 * Kernels that support flow wildcarding will reject these flows as
1484 * duplicates (EEXIST), so lower the log level to debug for these
1485 * types of messages. */
1486 return (error && error != EEXIST) ? VLL_WARN : VLL_DBG;
1490 should_log_flow_message(int error)
1492 return !vlog_should_drop(THIS_MODULE, flow_message_log_level(error),
1493 error ? &error_rl : &dpmsg_rl);
1497 log_flow_message(const struct dpif *dpif, int error, const char *operation,
1498 const struct nlattr *key, size_t key_len,
1499 const struct nlattr *mask, size_t mask_len,
1500 const struct dpif_flow_stats *stats,
1501 const struct nlattr *actions, size_t actions_len)
1503 struct ds ds = DS_EMPTY_INITIALIZER;
1504 ds_put_format(&ds, "%s: ", dpif_name(dpif));
1506 ds_put_cstr(&ds, "failed to ");
1508 ds_put_format(&ds, "%s ", operation);
1510 ds_put_format(&ds, "(%s) ", ovs_strerror(error));
1512 odp_flow_format(key, key_len, mask, mask_len, NULL, &ds, true);
1514 ds_put_cstr(&ds, ", ");
1515 dpif_flow_stats_format(stats, &ds);
1517 if (actions || actions_len) {
1518 ds_put_cstr(&ds, ", actions:");
1519 format_odp_actions(&ds, actions, actions_len);
1521 vlog(THIS_MODULE, flow_message_log_level(error), "%s", ds_cstr(&ds));
1526 log_flow_put_message(struct dpif *dpif, const struct dpif_flow_put *put,
1529 if (should_log_flow_message(error)) {
1533 ds_put_cstr(&s, "put");
1534 if (put->flags & DPIF_FP_CREATE) {
1535 ds_put_cstr(&s, "[create]");
1537 if (put->flags & DPIF_FP_MODIFY) {
1538 ds_put_cstr(&s, "[modify]");
1540 if (put->flags & DPIF_FP_ZERO_STATS) {
1541 ds_put_cstr(&s, "[zero]");
1543 log_flow_message(dpif, error, ds_cstr(&s),
1544 put->key, put->key_len, put->mask, put->mask_len,
1545 put->stats, put->actions, put->actions_len);
1551 log_flow_del_message(struct dpif *dpif, const struct dpif_flow_del *del,
1554 if (should_log_flow_message(error)) {
1555 log_flow_message(dpif, error, "flow_del", del->key, del->key_len,
1556 NULL, 0, !error ? del->stats : NULL, NULL, 0);
1561 log_execute_message(struct dpif *dpif, const struct dpif_execute *execute,
1564 if (!(error ? VLOG_DROP_WARN(&error_rl) : VLOG_DROP_DBG(&dpmsg_rl))) {
1565 struct ds ds = DS_EMPTY_INITIALIZER;
1568 packet = ofp_packet_to_string(ofpbuf_data(execute->packet),
1569 ofpbuf_size(execute->packet));
1570 ds_put_format(&ds, "%s: execute ", dpif_name(dpif));
1571 format_odp_actions(&ds, execute->actions, execute->actions_len);
1573 ds_put_format(&ds, " failed (%s)", ovs_strerror(error));
1575 ds_put_format(&ds, " on packet %s", packet);
1576 vlog(THIS_MODULE, error ? VLL_WARN : VLL_DBG, "%s", ds_cstr(&ds));