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[] = {
62 #if defined(__linux__) || defined(_WIN32)
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);
101 static void log_flow_get_message(const struct dpif *,
102 const struct dpif_flow_get *, int error);
107 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
109 if (ovsthread_once_start(&once)) {
112 for (i = 0; i < ARRAY_SIZE(base_dpif_classes); i++) {
113 dp_register_provider(base_dpif_classes[i]);
115 dpctl_unixctl_register();
116 ovsthread_once_done(&once);
121 dp_register_provider__(const struct dpif_class *new_class)
123 struct registered_dpif_class *registered_class;
125 if (sset_contains(&dpif_blacklist, new_class->type)) {
126 VLOG_DBG("attempted to register blacklisted provider: %s",
131 if (shash_find(&dpif_classes, new_class->type)) {
132 VLOG_WARN("attempted to register duplicate datapath provider: %s",
137 registered_class = xmalloc(sizeof *registered_class);
138 registered_class->dpif_class = new_class;
139 registered_class->refcount = 0;
141 shash_add(&dpif_classes, new_class->type, registered_class);
146 /* Registers a new datapath provider. After successful registration, new
147 * datapaths of that type can be opened using dpif_open(). */
149 dp_register_provider(const struct dpif_class *new_class)
153 ovs_mutex_lock(&dpif_mutex);
154 error = dp_register_provider__(new_class);
155 ovs_mutex_unlock(&dpif_mutex);
160 /* Unregisters a datapath provider. 'type' must have been previously
161 * registered and not currently be in use by any dpifs. After unregistration
162 * new datapaths of that type cannot be opened using dpif_open(). */
164 dp_unregister_provider__(const char *type)
166 struct shash_node *node;
167 struct registered_dpif_class *registered_class;
169 node = shash_find(&dpif_classes, type);
171 VLOG_WARN("attempted to unregister a datapath provider that is not "
172 "registered: %s", type);
176 registered_class = node->data;
177 if (registered_class->refcount) {
178 VLOG_WARN("attempted to unregister in use datapath provider: %s", type);
182 shash_delete(&dpif_classes, node);
183 free(registered_class);
188 /* Unregisters a datapath provider. 'type' must have been previously
189 * registered and not currently be in use by any dpifs. After unregistration
190 * new datapaths of that type cannot be opened using dpif_open(). */
192 dp_unregister_provider(const char *type)
198 ovs_mutex_lock(&dpif_mutex);
199 error = dp_unregister_provider__(type);
200 ovs_mutex_unlock(&dpif_mutex);
205 /* Blacklists a provider. Causes future calls of dp_register_provider() with
206 * a dpif_class which implements 'type' to fail. */
208 dp_blacklist_provider(const char *type)
210 ovs_mutex_lock(&dpif_mutex);
211 sset_add(&dpif_blacklist, type);
212 ovs_mutex_unlock(&dpif_mutex);
215 /* Clears 'types' and enumerates the types of all currently registered datapath
216 * providers into it. The caller must first initialize the sset. */
218 dp_enumerate_types(struct sset *types)
220 struct shash_node *node;
225 ovs_mutex_lock(&dpif_mutex);
226 SHASH_FOR_EACH(node, &dpif_classes) {
227 const struct registered_dpif_class *registered_class = node->data;
228 sset_add(types, registered_class->dpif_class->type);
230 ovs_mutex_unlock(&dpif_mutex);
234 dp_class_unref(struct registered_dpif_class *rc)
236 ovs_mutex_lock(&dpif_mutex);
237 ovs_assert(rc->refcount);
239 ovs_mutex_unlock(&dpif_mutex);
242 static struct registered_dpif_class *
243 dp_class_lookup(const char *type)
245 struct registered_dpif_class *rc;
247 ovs_mutex_lock(&dpif_mutex);
248 rc = shash_find_data(&dpif_classes, type);
252 ovs_mutex_unlock(&dpif_mutex);
257 /* Clears 'names' and enumerates the names of all known created datapaths with
258 * the given 'type'. The caller must first initialize the sset. Returns 0 if
259 * successful, otherwise a positive errno value.
261 * Some kinds of datapaths might not be practically enumerable. This is not
262 * considered an error. */
264 dp_enumerate_names(const char *type, struct sset *names)
266 struct registered_dpif_class *registered_class;
267 const struct dpif_class *dpif_class;
273 registered_class = dp_class_lookup(type);
274 if (!registered_class) {
275 VLOG_WARN("could not enumerate unknown type: %s", type);
279 dpif_class = registered_class->dpif_class;
280 error = (dpif_class->enumerate
281 ? dpif_class->enumerate(names, dpif_class)
284 VLOG_WARN("failed to enumerate %s datapaths: %s", dpif_class->type,
285 ovs_strerror(error));
287 dp_class_unref(registered_class);
292 /* Parses 'datapath_name_', which is of the form [type@]name into its
293 * component pieces. 'name' and 'type' must be freed by the caller.
295 * The returned 'type' is normalized, as if by dpif_normalize_type(). */
297 dp_parse_name(const char *datapath_name_, char **name, char **type)
299 char *datapath_name = xstrdup(datapath_name_);
302 separator = strchr(datapath_name, '@');
305 *type = datapath_name;
306 *name = xstrdup(dpif_normalize_type(separator + 1));
308 *name = datapath_name;
309 *type = xstrdup(dpif_normalize_type(NULL));
314 do_open(const char *name, const char *type, bool create, struct dpif **dpifp)
316 struct dpif *dpif = NULL;
318 struct registered_dpif_class *registered_class;
322 type = dpif_normalize_type(type);
323 registered_class = dp_class_lookup(type);
324 if (!registered_class) {
325 VLOG_WARN("could not create datapath %s of unknown type %s", name,
327 error = EAFNOSUPPORT;
331 error = registered_class->dpif_class->open(registered_class->dpif_class,
332 name, create, &dpif);
334 ovs_assert(dpif->dpif_class == registered_class->dpif_class);
336 dp_class_unref(registered_class);
340 *dpifp = error ? NULL : dpif;
344 /* Tries to open an existing datapath named 'name' and type 'type'. Will fail
345 * if no datapath with 'name' and 'type' exists. 'type' may be either NULL or
346 * the empty string to specify the default system type. Returns 0 if
347 * successful, otherwise a positive errno value. On success stores a pointer
348 * to the datapath in '*dpifp', otherwise a null pointer. */
350 dpif_open(const char *name, const char *type, struct dpif **dpifp)
352 return do_open(name, type, false, dpifp);
355 /* Tries to create and open a new datapath with the given 'name' and 'type'.
356 * 'type' may be either NULL or the empty string to specify the default system
357 * type. Will fail if a datapath with 'name' and 'type' already exists.
358 * Returns 0 if successful, otherwise a positive errno value. On success
359 * stores a pointer to the datapath in '*dpifp', otherwise a null pointer. */
361 dpif_create(const char *name, const char *type, struct dpif **dpifp)
363 return do_open(name, type, true, dpifp);
366 /* Tries to open a datapath with the given 'name' and 'type', creating it if it
367 * does not exist. 'type' may be either NULL or the empty string to specify
368 * the default system type. Returns 0 if successful, otherwise a positive
369 * errno value. On success stores a pointer to the datapath in '*dpifp',
370 * otherwise a null pointer. */
372 dpif_create_and_open(const char *name, const char *type, struct dpif **dpifp)
376 error = dpif_create(name, type, dpifp);
377 if (error == EEXIST || error == EBUSY) {
378 error = dpif_open(name, type, dpifp);
380 VLOG_WARN("datapath %s already exists but cannot be opened: %s",
381 name, ovs_strerror(error));
384 VLOG_WARN("failed to create datapath %s: %s",
385 name, ovs_strerror(error));
390 /* Closes and frees the connection to 'dpif'. Does not destroy the datapath
391 * itself; call dpif_delete() first, instead, if that is desirable. */
393 dpif_close(struct dpif *dpif)
396 struct registered_dpif_class *rc;
398 rc = shash_find_data(&dpif_classes, dpif->dpif_class->type);
399 dpif_uninit(dpif, true);
404 /* Performs periodic work needed by 'dpif'. */
406 dpif_run(struct dpif *dpif)
408 if (dpif->dpif_class->run) {
409 dpif->dpif_class->run(dpif);
413 /* Arranges for poll_block() to wake up when dp_run() needs to be called for
416 dpif_wait(struct dpif *dpif)
418 if (dpif->dpif_class->wait) {
419 dpif->dpif_class->wait(dpif);
423 /* Returns the name of datapath 'dpif' prefixed with the type
424 * (for use in log messages). */
426 dpif_name(const struct dpif *dpif)
428 return dpif->full_name;
431 /* Returns the name of datapath 'dpif' without the type
432 * (for use in device names). */
434 dpif_base_name(const struct dpif *dpif)
436 return dpif->base_name;
439 /* Returns the type of datapath 'dpif'. */
441 dpif_type(const struct dpif *dpif)
443 return dpif->dpif_class->type;
446 /* Returns the fully spelled out name for the given datapath 'type'.
448 * Normalized type string can be compared with strcmp(). Unnormalized type
449 * string might be the same even if they have different spellings. */
451 dpif_normalize_type(const char *type)
453 return type && type[0] ? type : "system";
456 /* Destroys the datapath that 'dpif' is connected to, first removing all of its
457 * ports. After calling this function, it does not make sense to pass 'dpif'
458 * to any functions other than dpif_name() or dpif_close(). */
460 dpif_delete(struct dpif *dpif)
464 COVERAGE_INC(dpif_destroy);
466 error = dpif->dpif_class->destroy(dpif);
467 log_operation(dpif, "delete", error);
471 /* Retrieves statistics for 'dpif' into 'stats'. Returns 0 if successful,
472 * otherwise a positive errno value. */
474 dpif_get_dp_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
476 int error = dpif->dpif_class->get_stats(dpif, stats);
478 memset(stats, 0, sizeof *stats);
480 log_operation(dpif, "get_stats", error);
485 dpif_port_open_type(const char *datapath_type, const char *port_type)
487 struct registered_dpif_class *rc;
489 datapath_type = dpif_normalize_type(datapath_type);
491 ovs_mutex_lock(&dpif_mutex);
492 rc = shash_find_data(&dpif_classes, datapath_type);
493 if (rc && rc->dpif_class->port_open_type) {
494 port_type = rc->dpif_class->port_open_type(rc->dpif_class, port_type);
496 ovs_mutex_unlock(&dpif_mutex);
501 /* Attempts to add 'netdev' as a port on 'dpif'. If 'port_nop' is
502 * non-null and its value is not ODPP_NONE, then attempts to use the
503 * value as the port number.
505 * If successful, returns 0 and sets '*port_nop' to the new port's port
506 * number (if 'port_nop' is non-null). On failure, returns a positive
507 * errno value and sets '*port_nop' to ODPP_NONE (if 'port_nop' is
510 dpif_port_add(struct dpif *dpif, struct netdev *netdev, odp_port_t *port_nop)
512 const char *netdev_name = netdev_get_name(netdev);
513 odp_port_t port_no = ODPP_NONE;
516 COVERAGE_INC(dpif_port_add);
522 error = dpif->dpif_class->port_add(dpif, netdev, &port_no);
524 VLOG_DBG_RL(&dpmsg_rl, "%s: added %s as port %"PRIu32,
525 dpif_name(dpif), netdev_name, port_no);
527 VLOG_WARN_RL(&error_rl, "%s: failed to add %s as port: %s",
528 dpif_name(dpif), netdev_name, ovs_strerror(error));
537 /* Attempts to remove 'dpif''s port number 'port_no'. Returns 0 if successful,
538 * otherwise a positive errno value. */
540 dpif_port_del(struct dpif *dpif, odp_port_t port_no)
544 COVERAGE_INC(dpif_port_del);
546 error = dpif->dpif_class->port_del(dpif, port_no);
548 VLOG_DBG_RL(&dpmsg_rl, "%s: port_del(%"PRIu32")",
549 dpif_name(dpif), port_no);
551 log_operation(dpif, "port_del", error);
556 /* Makes a deep copy of 'src' into 'dst'. */
558 dpif_port_clone(struct dpif_port *dst, const struct dpif_port *src)
560 dst->name = xstrdup(src->name);
561 dst->type = xstrdup(src->type);
562 dst->port_no = src->port_no;
565 /* Frees memory allocated to members of 'dpif_port'.
567 * Do not call this function on a dpif_port obtained from
568 * dpif_port_dump_next(): that function retains ownership of the data in the
571 dpif_port_destroy(struct dpif_port *dpif_port)
573 free(dpif_port->name);
574 free(dpif_port->type);
577 /* Checks if port named 'devname' exists in 'dpif'. If so, returns
578 * true; otherwise, returns false. */
580 dpif_port_exists(const struct dpif *dpif, const char *devname)
582 int error = dpif->dpif_class->port_query_by_name(dpif, devname, NULL);
583 if (error != 0 && error != ENOENT && error != ENODEV) {
584 VLOG_WARN_RL(&error_rl, "%s: failed to query port %s: %s",
585 dpif_name(dpif), devname, ovs_strerror(error));
591 /* Looks up port number 'port_no' in 'dpif'. On success, returns 0 and
592 * initializes '*port' appropriately; on failure, returns a positive errno
595 * The caller owns the data in 'port' and must free it with
596 * dpif_port_destroy() when it is no longer needed. */
598 dpif_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
599 struct dpif_port *port)
601 int error = dpif->dpif_class->port_query_by_number(dpif, port_no, port);
603 VLOG_DBG_RL(&dpmsg_rl, "%s: port %"PRIu32" is device %s",
604 dpif_name(dpif), port_no, port->name);
606 memset(port, 0, sizeof *port);
607 VLOG_WARN_RL(&error_rl, "%s: failed to query port %"PRIu32": %s",
608 dpif_name(dpif), port_no, ovs_strerror(error));
613 /* Looks up port named 'devname' in 'dpif'. On success, returns 0 and
614 * initializes '*port' appropriately; on failure, returns a positive errno
617 * The caller owns the data in 'port' and must free it with
618 * dpif_port_destroy() when it is no longer needed. */
620 dpif_port_query_by_name(const struct dpif *dpif, const char *devname,
621 struct dpif_port *port)
623 int error = dpif->dpif_class->port_query_by_name(dpif, devname, port);
625 VLOG_DBG_RL(&dpmsg_rl, "%s: device %s is on port %"PRIu32,
626 dpif_name(dpif), devname, port->port_no);
628 memset(port, 0, sizeof *port);
630 /* For ENOENT or ENODEV we use DBG level because the caller is probably
631 * interested in whether 'dpif' actually has a port 'devname', so that
632 * it's not an issue worth logging if it doesn't. Other errors are
633 * uncommon and more likely to indicate a real problem. */
635 error == ENOENT || error == ENODEV ? VLL_DBG : VLL_WARN,
636 "%s: failed to query port %s: %s",
637 dpif_name(dpif), devname, ovs_strerror(error));
642 /* Returns the Netlink PID value to supply in OVS_ACTION_ATTR_USERSPACE
643 * actions as the OVS_USERSPACE_ATTR_PID attribute's value, for use in
644 * flows whose packets arrived on port 'port_no'. In the case where the
645 * provider allocates multiple Netlink PIDs to a single port, it may use
646 * 'hash' to spread load among them. The caller need not use a particular
647 * hash function; a 5-tuple hash is suitable.
649 * (The datapath implementation might use some different hash function for
650 * distributing packets received via flow misses among PIDs. This means
651 * that packets received via flow misses might be reordered relative to
652 * packets received via userspace actions. This is not ordinarily a
655 * A 'port_no' of ODPP_NONE is a special case: it returns a reserved PID, not
656 * allocated to any port, that the client may use for special purposes.
658 * The return value is only meaningful when DPIF_UC_ACTION has been enabled in
659 * the 'dpif''s listen mask. It is allowed to change when DPIF_UC_ACTION is
660 * disabled and then re-enabled, so a client that does that must be prepared to
661 * update all of the flows that it installed that contain
662 * OVS_ACTION_ATTR_USERSPACE actions. */
664 dpif_port_get_pid(const struct dpif *dpif, odp_port_t port_no, uint32_t hash)
666 return (dpif->dpif_class->port_get_pid
667 ? (dpif->dpif_class->port_get_pid)(dpif, port_no, hash)
671 /* Looks up port number 'port_no' in 'dpif'. On success, returns 0 and copies
672 * the port's name into the 'name_size' bytes in 'name', ensuring that the
673 * result is null-terminated. On failure, returns a positive errno value and
674 * makes 'name' the empty string. */
676 dpif_port_get_name(struct dpif *dpif, odp_port_t port_no,
677 char *name, size_t name_size)
679 struct dpif_port port;
682 ovs_assert(name_size > 0);
684 error = dpif_port_query_by_number(dpif, port_no, &port);
686 ovs_strlcpy(name, port.name, name_size);
687 dpif_port_destroy(&port);
694 /* Initializes 'dump' to begin dumping the ports in a dpif.
696 * This function provides no status indication. An error status for the entire
697 * dump operation is provided when it is completed by calling
698 * dpif_port_dump_done().
701 dpif_port_dump_start(struct dpif_port_dump *dump, const struct dpif *dpif)
704 dump->error = dpif->dpif_class->port_dump_start(dpif, &dump->state);
705 log_operation(dpif, "port_dump_start", dump->error);
708 /* Attempts to retrieve another port from 'dump', which must have been
709 * initialized with dpif_port_dump_start(). On success, stores a new dpif_port
710 * into 'port' and returns true. On failure, returns false.
712 * Failure might indicate an actual error or merely that the last port has been
713 * dumped. An error status for the entire dump operation is provided when it
714 * is completed by calling dpif_port_dump_done().
716 * The dpif owns the data stored in 'port'. It will remain valid until at
717 * least the next time 'dump' is passed to dpif_port_dump_next() or
718 * dpif_port_dump_done(). */
720 dpif_port_dump_next(struct dpif_port_dump *dump, struct dpif_port *port)
722 const struct dpif *dpif = dump->dpif;
728 dump->error = dpif->dpif_class->port_dump_next(dpif, dump->state, port);
729 if (dump->error == EOF) {
730 VLOG_DBG_RL(&dpmsg_rl, "%s: dumped all ports", dpif_name(dpif));
732 log_operation(dpif, "port_dump_next", dump->error);
736 dpif->dpif_class->port_dump_done(dpif, dump->state);
742 /* Completes port table dump operation 'dump', which must have been initialized
743 * with dpif_port_dump_start(). Returns 0 if the dump operation was
744 * error-free, otherwise a positive errno value describing the problem. */
746 dpif_port_dump_done(struct dpif_port_dump *dump)
748 const struct dpif *dpif = dump->dpif;
750 dump->error = dpif->dpif_class->port_dump_done(dpif, dump->state);
751 log_operation(dpif, "port_dump_done", dump->error);
753 return dump->error == EOF ? 0 : dump->error;
756 /* Polls for changes in the set of ports in 'dpif'. If the set of ports in
757 * 'dpif' has changed, this function does one of the following:
759 * - Stores the name of the device that was added to or deleted from 'dpif' in
760 * '*devnamep' and returns 0. The caller is responsible for freeing
761 * '*devnamep' (with free()) when it no longer needs it.
763 * - Returns ENOBUFS and sets '*devnamep' to NULL.
765 * This function may also return 'false positives', where it returns 0 and
766 * '*devnamep' names a device that was not actually added or deleted or it
767 * returns ENOBUFS without any change.
769 * Returns EAGAIN if the set of ports in 'dpif' has not changed. May also
770 * return other positive errno values to indicate that something has gone
773 dpif_port_poll(const struct dpif *dpif, char **devnamep)
775 int error = dpif->dpif_class->port_poll(dpif, devnamep);
782 /* Arranges for the poll loop to wake up when port_poll(dpif) will return a
783 * value other than EAGAIN. */
785 dpif_port_poll_wait(const struct dpif *dpif)
787 dpif->dpif_class->port_poll_wait(dpif);
790 /* Extracts the flow stats for a packet. The 'flow' and 'packet'
791 * arguments must have been initialized through a call to flow_extract().
792 * 'used' is stored into stats->used. */
794 dpif_flow_stats_extract(const struct flow *flow, const struct ofpbuf *packet,
795 long long int used, struct dpif_flow_stats *stats)
797 stats->tcp_flags = ntohs(flow->tcp_flags);
798 stats->n_bytes = ofpbuf_size(packet);
799 stats->n_packets = 1;
803 /* Appends a human-readable representation of 'stats' to 's'. */
805 dpif_flow_stats_format(const struct dpif_flow_stats *stats, struct ds *s)
807 ds_put_format(s, "packets:%"PRIu64", bytes:%"PRIu64", used:",
808 stats->n_packets, stats->n_bytes);
810 ds_put_format(s, "%.3fs", (time_msec() - stats->used) / 1000.0);
812 ds_put_format(s, "never");
814 if (stats->tcp_flags) {
815 ds_put_cstr(s, ", flags:");
816 packet_format_tcp_flags(s, stats->tcp_flags);
820 /* Deletes all flows from 'dpif'. Returns 0 if successful, otherwise a
821 * positive errno value. */
823 dpif_flow_flush(struct dpif *dpif)
827 COVERAGE_INC(dpif_flow_flush);
829 error = dpif->dpif_class->flow_flush(dpif);
830 log_operation(dpif, "flow_flush", error);
834 /* A dpif_operate() wrapper for performing a single DPIF_OP_FLOW_GET. */
836 dpif_flow_get(struct dpif *dpif,
837 const struct nlattr *key, size_t key_len,
838 struct ofpbuf *buf, struct dpif_flow *flow)
843 op.type = DPIF_OP_FLOW_GET;
844 op.u.flow_get.key = key;
845 op.u.flow_get.key_len = key_len;
846 op.u.flow_get.buffer = buf;
847 op.u.flow_get.flow = flow;
848 op.u.flow_get.flow->key = key;
849 op.u.flow_get.flow->key_len = key_len;
852 dpif_operate(dpif, &opp, 1);
857 /* A dpif_operate() wrapper for performing a single DPIF_OP_FLOW_PUT. */
859 dpif_flow_put(struct dpif *dpif, enum dpif_flow_put_flags flags,
860 const struct nlattr *key, size_t key_len,
861 const struct nlattr *mask, size_t mask_len,
862 const struct nlattr *actions, size_t actions_len,
863 struct dpif_flow_stats *stats)
868 op.type = DPIF_OP_FLOW_PUT;
869 op.u.flow_put.flags = flags;
870 op.u.flow_put.key = key;
871 op.u.flow_put.key_len = key_len;
872 op.u.flow_put.mask = mask;
873 op.u.flow_put.mask_len = mask_len;
874 op.u.flow_put.actions = actions;
875 op.u.flow_put.actions_len = actions_len;
876 op.u.flow_put.stats = stats;
879 dpif_operate(dpif, &opp, 1);
884 /* A dpif_operate() wrapper for performing a single DPIF_OP_FLOW_DEL. */
886 dpif_flow_del(struct dpif *dpif,
887 const struct nlattr *key, size_t key_len,
888 struct dpif_flow_stats *stats)
893 op.type = DPIF_OP_FLOW_DEL;
894 op.u.flow_del.key = key;
895 op.u.flow_del.key_len = key_len;
896 op.u.flow_del.stats = stats;
899 dpif_operate(dpif, &opp, 1);
904 /* Creates and returns a new 'struct dpif_flow_dump' for iterating through the
907 * This function always successfully returns a dpif_flow_dump. Error
908 * reporting is deferred to dpif_flow_dump_destroy(). */
909 struct dpif_flow_dump *
910 dpif_flow_dump_create(const struct dpif *dpif)
912 return dpif->dpif_class->flow_dump_create(dpif);
915 /* Destroys 'dump', which must have been created with dpif_flow_dump_create().
916 * All dpif_flow_dump_thread structures previously created for 'dump' must
917 * previously have been destroyed.
919 * Returns 0 if the dump operation was error-free, otherwise a positive errno
920 * value describing the problem. */
922 dpif_flow_dump_destroy(struct dpif_flow_dump *dump)
924 const struct dpif *dpif = dump->dpif;
925 int error = dpif->dpif_class->flow_dump_destroy(dump);
926 log_operation(dpif, "flow_dump_destroy", error);
927 return error == EOF ? 0 : error;
930 /* Returns new thread-local state for use with dpif_flow_dump_next(). */
931 struct dpif_flow_dump_thread *
932 dpif_flow_dump_thread_create(struct dpif_flow_dump *dump)
934 return dump->dpif->dpif_class->flow_dump_thread_create(dump);
937 /* Releases 'thread'. */
939 dpif_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread)
941 thread->dpif->dpif_class->flow_dump_thread_destroy(thread);
944 /* Attempts to retrieve up to 'max_flows' more flows from 'thread'. Returns 0
945 * if and only if no flows remained to be retrieved, otherwise a positive
946 * number reflecting the number of elements in 'flows[]' that were updated.
947 * The number of flows returned might be less than 'max_flows' because
948 * fewer than 'max_flows' remained, because this particular datapath does not
949 * benefit from batching, or because an error occurred partway through
950 * retrieval. Thus, the caller should continue calling until a 0 return value,
951 * even if intermediate return values are less than 'max_flows'.
953 * No error status is immediately provided. An error status for the entire
954 * dump operation is provided when it is completed by calling
955 * dpif_flow_dump_destroy().
957 * All of the data stored into 'flows' is owned by the datapath, not by the
958 * caller, and the caller must not modify or free it. The datapath guarantees
959 * that it remains accessible and unchanged until the first of:
960 * - The next call to dpif_flow_dump_next() for 'thread', or
961 * - The next rcu quiescent period. */
963 dpif_flow_dump_next(struct dpif_flow_dump_thread *thread,
964 struct dpif_flow *flows, int max_flows)
966 struct dpif *dpif = thread->dpif;
969 ovs_assert(max_flows > 0);
970 n = dpif->dpif_class->flow_dump_next(thread, flows, max_flows);
974 for (f = flows; f < &flows[n] && should_log_flow_message(0); f++) {
975 log_flow_message(dpif, 0, "flow_dump",
976 f->key, f->key_len, f->mask, f->mask_len,
977 &f->stats, f->actions, f->actions_len);
980 VLOG_DBG_RL(&dpmsg_rl, "%s: dumped all flows", dpif_name(dpif));
985 struct dpif_execute_helper_aux {
990 /* This is called for actions that need the context of the datapath to be
993 dpif_execute_helper_cb(void *aux_, struct dpif_packet **packets, int cnt,
994 const struct nlattr *action, bool may_steal OVS_UNUSED)
996 struct dpif_execute_helper_aux *aux = aux_;
997 int type = nl_attr_type(action);
998 struct ofpbuf *packet = &packets[0]->ofpbuf;
999 struct pkt_metadata *md = &packets[0]->md;
1001 ovs_assert(cnt == 1);
1003 switch ((enum ovs_action_attr)type) {
1004 case OVS_ACTION_ATTR_OUTPUT:
1005 case OVS_ACTION_ATTR_USERSPACE:
1006 case OVS_ACTION_ATTR_RECIRC: {
1007 struct dpif_execute execute;
1008 struct ofpbuf execute_actions;
1009 uint64_t stub[256 / 8];
1011 if (md->tunnel.ip_dst) {
1012 /* The Linux kernel datapath throws away the tunnel information
1013 * that we supply as metadata. We have to use a "set" action to
1015 ofpbuf_use_stub(&execute_actions, stub, sizeof stub);
1016 odp_put_tunnel_action(&md->tunnel, &execute_actions);
1017 ofpbuf_put(&execute_actions, action, NLA_ALIGN(action->nla_len));
1019 execute.actions = ofpbuf_data(&execute_actions);
1020 execute.actions_len = ofpbuf_size(&execute_actions);
1022 execute.actions = action;
1023 execute.actions_len = NLA_ALIGN(action->nla_len);
1026 execute.packet = packet;
1028 execute.needs_help = false;
1029 execute.probe = false;
1030 aux->error = dpif_execute(aux->dpif, &execute);
1031 log_execute_message(aux->dpif, &execute, true, aux->error);
1033 if (md->tunnel.ip_dst) {
1034 ofpbuf_uninit(&execute_actions);
1039 case OVS_ACTION_ATTR_HASH:
1040 case OVS_ACTION_ATTR_PUSH_VLAN:
1041 case OVS_ACTION_ATTR_POP_VLAN:
1042 case OVS_ACTION_ATTR_PUSH_MPLS:
1043 case OVS_ACTION_ATTR_POP_MPLS:
1044 case OVS_ACTION_ATTR_SET:
1045 case OVS_ACTION_ATTR_SET_MASKED:
1046 case OVS_ACTION_ATTR_SAMPLE:
1047 case OVS_ACTION_ATTR_UNSPEC:
1048 case __OVS_ACTION_ATTR_MAX:
1053 /* Executes 'execute' by performing most of the actions in userspace and
1054 * passing the fully constructed packets to 'dpif' for output and userspace
1057 * This helps with actions that a given 'dpif' doesn't implement directly. */
1059 dpif_execute_with_help(struct dpif *dpif, struct dpif_execute *execute)
1061 struct dpif_execute_helper_aux aux = {dpif, 0};
1062 struct dpif_packet packet, *pp;
1064 COVERAGE_INC(dpif_execute_with_help);
1066 packet.ofpbuf = *execute->packet;
1067 packet.md = execute->md;
1070 odp_execute_actions(&aux, &pp, 1, false, execute->actions,
1071 execute->actions_len, dpif_execute_helper_cb);
1073 /* Even though may_steal is set to false, some actions could modify or
1074 * reallocate the ofpbuf memory. We need to pass those changes to the
1076 *execute->packet = packet.ofpbuf;
1077 execute->md = packet.md;
1082 /* Returns true if the datapath needs help executing 'execute'. */
1084 dpif_execute_needs_help(const struct dpif_execute *execute)
1086 return execute->needs_help || nl_attr_oversized(execute->actions_len);
1089 /* A dpif_operate() wrapper for performing a single DPIF_OP_EXECUTE. */
1091 dpif_execute(struct dpif *dpif, struct dpif_execute *execute)
1093 if (execute->actions_len) {
1094 struct dpif_op *opp;
1097 op.type = DPIF_OP_EXECUTE;
1098 op.u.execute = *execute;
1101 dpif_operate(dpif, &opp, 1);
1109 /* Executes each of the 'n_ops' operations in 'ops' on 'dpif', in the order in
1110 * which they are specified. Places each operation's results in the "output"
1111 * members documented in comments, and 0 in the 'error' member on success or a
1112 * positive errno on failure. */
1114 dpif_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
1119 /* Count 'chunk', the number of ops that can be executed without
1120 * needing any help. Ops that need help should be rare, so we
1121 * expect this to ordinarily be 'n_ops', that is, all the ops. */
1122 for (chunk = 0; chunk < n_ops; chunk++) {
1123 struct dpif_op *op = ops[chunk];
1125 if (op->type == DPIF_OP_EXECUTE
1126 && dpif_execute_needs_help(&op->u.execute)) {
1132 /* Execute a chunk full of ops that the dpif provider can
1133 * handle itself, without help. */
1136 dpif->dpif_class->operate(dpif, ops, chunk);
1138 for (i = 0; i < chunk; i++) {
1139 struct dpif_op *op = ops[i];
1140 int error = op->error;
1143 case DPIF_OP_FLOW_PUT: {
1144 struct dpif_flow_put *put = &op->u.flow_put;
1146 COVERAGE_INC(dpif_flow_put);
1147 log_flow_put_message(dpif, put, error);
1148 if (error && put->stats) {
1149 memset(put->stats, 0, sizeof *put->stats);
1154 case DPIF_OP_FLOW_GET: {
1155 struct dpif_flow_get *get = &op->u.flow_get;
1157 COVERAGE_INC(dpif_flow_get);
1158 log_flow_get_message(dpif, get, error);
1161 memset(get->flow, 0, sizeof *get->flow);
1166 case DPIF_OP_FLOW_DEL: {
1167 struct dpif_flow_del *del = &op->u.flow_del;
1169 COVERAGE_INC(dpif_flow_del);
1170 log_flow_del_message(dpif, del, error);
1171 if (error && del->stats) {
1172 memset(del->stats, 0, sizeof *del->stats);
1177 case DPIF_OP_EXECUTE:
1178 COVERAGE_INC(dpif_execute);
1179 log_execute_message(dpif, &op->u.execute, false, error);
1187 /* Help the dpif provider to execute one op. */
1188 struct dpif_op *op = ops[0];
1190 COVERAGE_INC(dpif_execute);
1191 op->error = dpif_execute_with_help(dpif, &op->u.execute);
1198 /* Returns a string that represents 'type', for use in log messages. */
1200 dpif_upcall_type_to_string(enum dpif_upcall_type type)
1203 case DPIF_UC_MISS: return "miss";
1204 case DPIF_UC_ACTION: return "action";
1205 case DPIF_N_UC_TYPES: default: return "<unknown>";
1209 /* Enables or disables receiving packets with dpif_recv() on 'dpif'. Returns 0
1210 * if successful, otherwise a positive errno value.
1212 * Turning packet receive off and then back on may change the Netlink PID
1213 * assignments returned by dpif_port_get_pid(). If the client does this, it
1214 * must update all of the flows that have OVS_ACTION_ATTR_USERSPACE actions
1215 * using the new PID assignment. */
1217 dpif_recv_set(struct dpif *dpif, bool enable)
1221 if (dpif->dpif_class->recv_set) {
1222 error = dpif->dpif_class->recv_set(dpif, enable);
1223 log_operation(dpif, "recv_set", error);
1228 /* Refreshes the poll loops and Netlink sockets associated to each port,
1229 * when the number of upcall handlers (upcall receiving thread) is changed
1230 * to 'n_handlers' and receiving packets for 'dpif' is enabled by
1233 * Since multiple upcall handlers can read upcalls simultaneously from
1234 * 'dpif', each port can have multiple Netlink sockets, one per upcall
1235 * handler. So, handlers_set() is responsible for the following tasks:
1237 * When receiving upcall is enabled, extends or creates the
1238 * configuration to support:
1240 * - 'n_handlers' Netlink sockets for each port.
1242 * - 'n_handlers' poll loops, one for each upcall handler.
1244 * - registering the Netlink sockets for the same upcall handler to
1245 * the corresponding poll loop.
1247 * Returns 0 if successful, otherwise a positive errno value. */
1249 dpif_handlers_set(struct dpif *dpif, uint32_t n_handlers)
1253 if (dpif->dpif_class->handlers_set) {
1254 error = dpif->dpif_class->handlers_set(dpif, n_handlers);
1255 log_operation(dpif, "handlers_set", error);
1261 dpif_register_upcall_cb(struct dpif *dpif, upcall_callback *cb, void *aux)
1263 if (dpif->dpif_class->register_upcall_cb) {
1264 dpif->dpif_class->register_upcall_cb(dpif, cb, aux);
1269 dpif_enable_upcall(struct dpif *dpif)
1271 if (dpif->dpif_class->enable_upcall) {
1272 dpif->dpif_class->enable_upcall(dpif);
1277 dpif_disable_upcall(struct dpif *dpif)
1279 if (dpif->dpif_class->disable_upcall) {
1280 dpif->dpif_class->disable_upcall(dpif);
1285 dpif_print_packet(struct dpif *dpif, struct dpif_upcall *upcall)
1287 if (!VLOG_DROP_DBG(&dpmsg_rl)) {
1291 packet = ofp_packet_to_string(ofpbuf_data(&upcall->packet),
1292 ofpbuf_size(&upcall->packet));
1295 odp_flow_key_format(upcall->key, upcall->key_len, &flow);
1297 VLOG_DBG("%s: %s upcall:\n%s\n%s",
1298 dpif_name(dpif), dpif_upcall_type_to_string(upcall->type),
1299 ds_cstr(&flow), packet);
1306 /* If 'dpif' creates its own I/O polling threads, refreshes poll threads
1309 dpif_poll_threads_set(struct dpif *dpif, unsigned int n_rxqs,
1314 if (dpif->dpif_class->poll_threads_set) {
1315 error = dpif->dpif_class->poll_threads_set(dpif, n_rxqs, cmask);
1317 log_operation(dpif, "poll_threads_set", error);
1324 /* Polls for an upcall from 'dpif' for an upcall handler. Since there
1325 * there can be multiple poll loops, 'handler_id' is needed as index to
1326 * identify the corresponding poll loop. If successful, stores the upcall
1327 * into '*upcall', using 'buf' for storage. Should only be called if
1328 * 'recv_set' has been used to enable receiving packets from 'dpif'.
1330 * 'upcall->key' and 'upcall->userdata' point into data in the caller-provided
1331 * 'buf', so their memory cannot be freed separately from 'buf'.
1333 * The caller owns the data of 'upcall->packet' and may modify it. If
1334 * packet's headroom is exhausted as it is manipulated, 'upcall->packet'
1335 * will be reallocated. This requires the data of 'upcall->packet' to be
1336 * released with ofpbuf_uninit() before 'upcall' is destroyed. However,
1337 * when an error is returned, the 'upcall->packet' may be uninitialized
1338 * and should not be released.
1340 * Returns 0 if successful, otherwise a positive errno value. Returns EAGAIN
1341 * if no upcall is immediately available. */
1343 dpif_recv(struct dpif *dpif, uint32_t handler_id, struct dpif_upcall *upcall,
1348 if (dpif->dpif_class->recv) {
1349 error = dpif->dpif_class->recv(dpif, handler_id, upcall, buf);
1351 dpif_print_packet(dpif, upcall);
1352 } else if (error != EAGAIN) {
1353 log_operation(dpif, "recv", error);
1359 /* Discards all messages that would otherwise be received by dpif_recv() on
1362 dpif_recv_purge(struct dpif *dpif)
1364 COVERAGE_INC(dpif_purge);
1365 if (dpif->dpif_class->recv_purge) {
1366 dpif->dpif_class->recv_purge(dpif);
1370 /* Arranges for the poll loop for an upcall handler to wake up when 'dpif'
1371 * 'dpif' has a message queued to be received with the recv member
1372 * function. Since there can be multiple poll loops, 'handler_id' is
1373 * needed as index to identify the corresponding poll loop. */
1375 dpif_recv_wait(struct dpif *dpif, uint32_t handler_id)
1377 if (dpif->dpif_class->recv_wait) {
1378 dpif->dpif_class->recv_wait(dpif, handler_id);
1382 /* Obtains the NetFlow engine type and engine ID for 'dpif' into '*engine_type'
1383 * and '*engine_id', respectively. */
1385 dpif_get_netflow_ids(const struct dpif *dpif,
1386 uint8_t *engine_type, uint8_t *engine_id)
1388 *engine_type = dpif->netflow_engine_type;
1389 *engine_id = dpif->netflow_engine_id;
1392 /* Translates OpenFlow queue ID 'queue_id' (in host byte order) into a priority
1393 * value used for setting packet priority.
1394 * On success, returns 0 and stores the priority into '*priority'.
1395 * On failure, returns a positive errno value and stores 0 into '*priority'. */
1397 dpif_queue_to_priority(const struct dpif *dpif, uint32_t queue_id,
1400 int error = (dpif->dpif_class->queue_to_priority
1401 ? dpif->dpif_class->queue_to_priority(dpif, queue_id,
1407 log_operation(dpif, "queue_to_priority", error);
1412 dpif_init(struct dpif *dpif, const struct dpif_class *dpif_class,
1414 uint8_t netflow_engine_type, uint8_t netflow_engine_id)
1416 dpif->dpif_class = dpif_class;
1417 dpif->base_name = xstrdup(name);
1418 dpif->full_name = xasprintf("%s@%s", dpif_class->type, name);
1419 dpif->netflow_engine_type = netflow_engine_type;
1420 dpif->netflow_engine_id = netflow_engine_id;
1423 /* Undoes the results of initialization.
1425 * Normally this function only needs to be called from dpif_close().
1426 * However, it may be called by providers due to an error on opening
1427 * that occurs after initialization. It this case dpif_close() would
1428 * never be called. */
1430 dpif_uninit(struct dpif *dpif, bool close)
1432 char *base_name = dpif->base_name;
1433 char *full_name = dpif->full_name;
1436 dpif->dpif_class->close(dpif);
1444 log_operation(const struct dpif *dpif, const char *operation, int error)
1447 VLOG_DBG_RL(&dpmsg_rl, "%s: %s success", dpif_name(dpif), operation);
1448 } else if (ofperr_is_valid(error)) {
1449 VLOG_WARN_RL(&error_rl, "%s: %s failed (%s)",
1450 dpif_name(dpif), operation, ofperr_get_name(error));
1452 VLOG_WARN_RL(&error_rl, "%s: %s failed (%s)",
1453 dpif_name(dpif), operation, ovs_strerror(error));
1457 static enum vlog_level
1458 flow_message_log_level(int error)
1460 /* If flows arrive in a batch, userspace may push down multiple
1461 * unique flow definitions that overlap when wildcards are applied.
1462 * Kernels that support flow wildcarding will reject these flows as
1463 * duplicates (EEXIST), so lower the log level to debug for these
1464 * types of messages. */
1465 return (error && error != EEXIST) ? VLL_WARN : VLL_DBG;
1469 should_log_flow_message(int error)
1471 return !vlog_should_drop(THIS_MODULE, flow_message_log_level(error),
1472 error ? &error_rl : &dpmsg_rl);
1476 log_flow_message(const struct dpif *dpif, int error, const char *operation,
1477 const struct nlattr *key, size_t key_len,
1478 const struct nlattr *mask, size_t mask_len,
1479 const struct dpif_flow_stats *stats,
1480 const struct nlattr *actions, size_t actions_len)
1482 struct ds ds = DS_EMPTY_INITIALIZER;
1483 ds_put_format(&ds, "%s: ", dpif_name(dpif));
1485 ds_put_cstr(&ds, "failed to ");
1487 ds_put_format(&ds, "%s ", operation);
1489 ds_put_format(&ds, "(%s) ", ovs_strerror(error));
1491 odp_flow_format(key, key_len, mask, mask_len, NULL, &ds, true);
1493 ds_put_cstr(&ds, ", ");
1494 dpif_flow_stats_format(stats, &ds);
1496 if (actions || actions_len) {
1497 ds_put_cstr(&ds, ", actions:");
1498 format_odp_actions(&ds, actions, actions_len);
1500 vlog(THIS_MODULE, flow_message_log_level(error), "%s", ds_cstr(&ds));
1505 log_flow_put_message(struct dpif *dpif, const struct dpif_flow_put *put,
1508 if (should_log_flow_message(error) && !(put->flags & DPIF_FP_PROBE)) {
1512 ds_put_cstr(&s, "put");
1513 if (put->flags & DPIF_FP_CREATE) {
1514 ds_put_cstr(&s, "[create]");
1516 if (put->flags & DPIF_FP_MODIFY) {
1517 ds_put_cstr(&s, "[modify]");
1519 if (put->flags & DPIF_FP_ZERO_STATS) {
1520 ds_put_cstr(&s, "[zero]");
1522 log_flow_message(dpif, error, ds_cstr(&s),
1523 put->key, put->key_len, put->mask, put->mask_len,
1524 put->stats, put->actions, put->actions_len);
1530 log_flow_del_message(struct dpif *dpif, const struct dpif_flow_del *del,
1533 if (should_log_flow_message(error)) {
1534 log_flow_message(dpif, error, "flow_del", del->key, del->key_len,
1535 NULL, 0, !error ? del->stats : NULL, NULL, 0);
1539 /* Logs that 'execute' was executed on 'dpif' and completed with errno 'error'
1540 * (0 for success). 'subexecute' should be true if the execution is a result
1541 * of breaking down a larger execution that needed help, false otherwise.
1544 * XXX In theory, the log message could be deceptive because this function is
1545 * called after the dpif_provider's '->execute' function, which is allowed to
1546 * modify execute->packet and execute->md. In practice, though:
1548 * - dpif-netlink doesn't modify execute->packet or execute->md.
1550 * - dpif-netdev does modify them but it is less likely to have problems
1551 * because it is built into ovs-vswitchd and cannot have version skew,
1554 * It would still be better to avoid the potential problem. I don't know of a
1555 * good way to do that, though, that isn't expensive. */
1557 log_execute_message(struct dpif *dpif, const struct dpif_execute *execute,
1558 bool subexecute, int error)
1560 if (!(error ? VLOG_DROP_WARN(&error_rl) : VLOG_DROP_DBG(&dpmsg_rl))
1561 && !execute->probe) {
1562 struct ds ds = DS_EMPTY_INITIALIZER;
1565 packet = ofp_packet_to_string(ofpbuf_data(execute->packet),
1566 ofpbuf_size(execute->packet));
1567 ds_put_format(&ds, "%s: %sexecute ",
1569 (subexecute ? "sub-"
1570 : dpif_execute_needs_help(execute) ? "super-"
1572 format_odp_actions(&ds, execute->actions, execute->actions_len);
1574 ds_put_format(&ds, " failed (%s)", ovs_strerror(error));
1576 ds_put_format(&ds, " on packet %s", packet);
1577 vlog(THIS_MODULE, error ? VLL_WARN : VLL_DBG, "%s", ds_cstr(&ds));
1584 log_flow_get_message(const struct dpif *dpif, const struct dpif_flow_get *get,
1587 if (should_log_flow_message(error)) {
1588 log_flow_message(dpif, error, "flow_get",
1589 get->key, get->key_len,
1590 get->flow->mask, get->flow->mask_len,
1592 get->flow->actions, get->flow->actions_len);