2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include "dpif-netdev.h"
24 #include <netinet/in.h>
25 #include <sys/socket.h>
30 #include <sys/ioctl.h>
34 #include "classifier.h"
38 #include "dpif-provider.h"
40 #include "dynamic-string.h"
41 #include "fat-rwlock.h"
46 #include "meta-flow.h"
48 #include "netdev-dpdk.h"
49 #include "netdev-vport.h"
51 #include "odp-execute.h"
53 #include "ofp-print.h"
56 #include "packet-dpif.h"
58 #include "poll-loop.h"
68 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
70 /* By default, choose a priority in the middle. */
71 #define NETDEV_RULE_PRIORITY 0x8000
73 #define FLOW_DUMP_MAX_BATCH 50
74 /* Use per thread recirc_depth to prevent recirculation loop. */
75 #define MAX_RECIRC_DEPTH 5
76 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
78 /* Configuration parameters. */
79 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
81 /* Protects against changes to 'dp_netdevs'. */
82 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
84 /* Contains all 'struct dp_netdev's. */
85 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
86 = SHASH_INITIALIZER(&dp_netdevs);
88 struct dp_netdev_queue {
89 unsigned int packet_count;
91 struct dpif_upcall upcalls[NETDEV_MAX_RX_BATCH];
92 struct ofpbuf bufs[NETDEV_MAX_RX_BATCH];
95 #define DP_NETDEV_QUEUE_INITIALIZER { .packet_count = 0 }
97 /* Datapath based on the network device interface from netdev.h.
103 * Some members, marked 'const', are immutable. Accessing other members
104 * requires synchronization, as noted in more detail below.
106 * Acquisition order is, from outermost to innermost:
108 * dp_netdev_mutex (global)
113 const struct dpif_class *const class;
114 const char *const name;
116 struct ovs_refcount ref_cnt;
117 atomic_flag destroyed;
121 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
122 * changes to 'cls' must be made while still holding the 'flow_mutex'.
124 struct ovs_mutex flow_mutex;
125 struct classifier cls;
126 struct cmap flow_table OVS_GUARDED; /* Flow table. */
130 * ovsthread_stats is internally synchronized. */
131 struct ovsthread_stats stats; /* Contains 'struct dp_netdev_stats *'. */
135 * Protected by RCU. Take the mutex to add or remove ports. */
136 struct ovs_mutex port_mutex;
138 struct seq *port_seq; /* Incremented whenever a port changes. */
140 /* Protects access to ofproto-dpif-upcall interface during revalidator
141 * thread synchronization. */
142 struct fat_rwlock upcall_rwlock;
143 exec_upcall_cb *upcall_cb; /* Callback function for executing upcalls. */
145 /* Forwarding threads. */
146 struct latch exit_latch;
147 struct pmd_thread *pmd_threads;
148 size_t n_pmd_threads;
152 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
156 DP_STAT_HIT, /* Packets that matched in the flow table. */
157 DP_STAT_MISS, /* Packets that did not match. */
158 DP_STAT_LOST, /* Packets not passed up to the client. */
162 /* Contained by struct dp_netdev's 'stats' member. */
163 struct dp_netdev_stats {
164 struct ovs_mutex mutex; /* Protects 'n'. */
166 /* Indexed by DP_STAT_*, protected by 'mutex'. */
167 unsigned long long int n[DP_N_STATS] OVS_GUARDED;
171 /* A port in a netdev-based datapath. */
172 struct dp_netdev_port {
173 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
175 struct netdev *netdev;
176 struct netdev_saved_flags *sf;
177 struct netdev_rxq **rxq;
178 struct ovs_refcount ref_cnt;
179 char *type; /* Port type as requested by user. */
183 /* Stores a miniflow */
185 /* There are fields in the flow structure that we never use. Therefore we can
186 * save a few words of memory */
187 #define NETDEV_KEY_BUF_SIZE_U32 (FLOW_U32S - MINI_N_INLINE \
188 - FLOW_U32_SIZE(regs) \
189 - FLOW_U32_SIZE(metadata) \
191 struct netdev_flow_key {
192 struct miniflow flow;
193 uint32_t buf[NETDEV_KEY_BUF_SIZE_U32];
196 /* A flow in dp_netdev's 'flow_table'.
202 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
203 * its dp_netdev's classifier. The text below calls this classifier 'cls'.
208 * The thread safety rules described here for "struct dp_netdev_flow" are
209 * motivated by two goals:
211 * - Prevent threads that read members of "struct dp_netdev_flow" from
212 * reading bad data due to changes by some thread concurrently modifying
215 * - Prevent two threads making changes to members of a given "struct
216 * dp_netdev_flow" from interfering with each other.
222 * A flow 'flow' may be accessed without a risk of being freed by code that
223 * holds a read-lock or write-lock on 'cls->rwlock' or that owns a reference to
224 * 'flow->ref_cnt' (or both). Code that needs to hold onto a flow for a while
225 * should take 'cls->rwlock', find the flow it needs, increment 'flow->ref_cnt'
226 * with dpif_netdev_flow_ref(), and drop 'cls->rwlock'.
228 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
229 * flow from being deleted from 'cls' (that's 'cls->rwlock') and it doesn't
230 * protect members of 'flow' from modification.
232 * Some members, marked 'const', are immutable. Accessing other members
233 * requires synchronization, as noted in more detail below.
235 struct dp_netdev_flow {
236 /* Packet classification. */
237 const struct cls_rule cr; /* In owning dp_netdev's 'cls'. */
239 /* Hash table index by unmasked flow. */
240 const struct cmap_node node; /* In owning dp_netdev's 'flow_table'. */
241 const struct flow flow; /* The flow that created this entry. */
245 * Reading or writing these members requires 'mutex'. */
246 struct ovsthread_stats stats; /* Contains "struct dp_netdev_flow_stats". */
249 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
252 static void dp_netdev_flow_free(struct dp_netdev_flow *);
254 /* Contained by struct dp_netdev_flow's 'stats' member. */
255 struct dp_netdev_flow_stats {
256 struct ovs_mutex mutex; /* Guards all the other members. */
258 long long int used OVS_GUARDED; /* Last used time, in monotonic msecs. */
259 long long int packet_count OVS_GUARDED; /* Number of packets matched. */
260 long long int byte_count OVS_GUARDED; /* Number of bytes matched. */
261 uint16_t tcp_flags OVS_GUARDED; /* Bitwise-OR of seen tcp_flags values. */
264 /* A set of datapath actions within a "struct dp_netdev_flow".
270 * A struct dp_netdev_actions 'actions' is protected with RCU. */
271 struct dp_netdev_actions {
272 /* These members are immutable: they do not change during the struct's
274 struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */
275 unsigned int size; /* Size of 'actions', in bytes. */
278 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
280 struct dp_netdev_actions *dp_netdev_flow_get_actions(
281 const struct dp_netdev_flow *);
282 static void dp_netdev_actions_free(struct dp_netdev_actions *);
284 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
285 * the performance overhead of interrupt processing. Therefore netdev can
286 * not implement rx-wait for these devices. dpif-netdev needs to poll
287 * these device to check for recv buffer. pmd-thread does polling for
288 * devices assigned to itself thread.
290 * DPDK used PMD for accessing NIC.
292 * A thread that receives packets from PMD ports, looks them up in the flow
293 * table, and executes the actions it finds.
296 struct dp_netdev *dp;
299 atomic_uint change_seq;
302 /* Interface to netdev-based datapath. */
305 struct dp_netdev *dp;
306 uint64_t last_port_seq;
309 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
310 struct dp_netdev_port **portp);
311 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
312 struct dp_netdev_port **portp);
313 static void dp_netdev_free(struct dp_netdev *)
314 OVS_REQUIRES(dp_netdev_mutex);
315 static void dp_netdev_flow_flush(struct dp_netdev *);
316 static int do_add_port(struct dp_netdev *dp, const char *devname,
317 const char *type, odp_port_t port_no)
318 OVS_REQUIRES(dp->port_mutex);
319 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
320 OVS_REQUIRES(dp->port_mutex);
321 static int dpif_netdev_open(const struct dpif_class *, const char *name,
322 bool create, struct dpif **);
323 static int dp_netdev_queue_userspace_packet(struct dp_netdev_queue *,
324 struct ofpbuf *, int type,
325 const struct miniflow *,
326 const struct nlattr *);
327 static void dp_netdev_execute_userspace_queue(struct dp_netdev_queue *,
329 static void dp_netdev_execute_actions(struct dp_netdev *dp,
330 struct dpif_packet **, int c,
331 bool may_steal, struct pkt_metadata *,
332 const struct nlattr *actions,
334 static void dp_netdev_port_input(struct dp_netdev *dp,
335 struct dpif_packet **packets, int cnt,
338 static void dp_netdev_set_pmd_threads(struct dp_netdev *, int n);
339 static void dp_netdev_disable_upcall(struct dp_netdev *);
341 static struct dpif_netdev *
342 dpif_netdev_cast(const struct dpif *dpif)
344 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
345 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
348 static struct dp_netdev *
349 get_dp_netdev(const struct dpif *dpif)
351 return dpif_netdev_cast(dpif)->dp;
355 dpif_netdev_enumerate(struct sset *all_dps,
356 const struct dpif_class *dpif_class)
358 struct shash_node *node;
360 ovs_mutex_lock(&dp_netdev_mutex);
361 SHASH_FOR_EACH(node, &dp_netdevs) {
362 struct dp_netdev *dp = node->data;
363 if (dpif_class != dp->class) {
364 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
365 * If the class doesn't match, skip this dpif. */
368 sset_add(all_dps, node->name);
370 ovs_mutex_unlock(&dp_netdev_mutex);
376 dpif_netdev_class_is_dummy(const struct dpif_class *class)
378 return class != &dpif_netdev_class;
382 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
384 return strcmp(type, "internal") ? type
385 : dpif_netdev_class_is_dummy(class) ? "dummy"
390 create_dpif_netdev(struct dp_netdev *dp)
392 uint16_t netflow_id = hash_string(dp->name, 0);
393 struct dpif_netdev *dpif;
395 ovs_refcount_ref(&dp->ref_cnt);
397 dpif = xmalloc(sizeof *dpif);
398 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
400 dpif->last_port_seq = seq_read(dp->port_seq);
405 /* Choose an unused, non-zero port number and return it on success.
406 * Return ODPP_NONE on failure. */
408 choose_port(struct dp_netdev *dp, const char *name)
409 OVS_REQUIRES(dp->port_mutex)
413 if (dp->class != &dpif_netdev_class) {
417 /* If the port name begins with "br", start the number search at
418 * 100 to make writing tests easier. */
419 if (!strncmp(name, "br", 2)) {
423 /* If the port name contains a number, try to assign that port number.
424 * This can make writing unit tests easier because port numbers are
426 for (p = name; *p != '\0'; p++) {
427 if (isdigit((unsigned char) *p)) {
428 port_no = start_no + strtol(p, NULL, 10);
429 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
430 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
431 return u32_to_odp(port_no);
438 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
439 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
440 return u32_to_odp(port_no);
448 create_dp_netdev(const char *name, const struct dpif_class *class,
449 struct dp_netdev **dpp)
450 OVS_REQUIRES(dp_netdev_mutex)
452 struct dp_netdev *dp;
455 dp = xzalloc(sizeof *dp);
456 shash_add(&dp_netdevs, name, dp);
458 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
459 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
460 ovs_refcount_init(&dp->ref_cnt);
461 atomic_flag_clear(&dp->destroyed);
463 ovs_mutex_init(&dp->flow_mutex);
464 classifier_init(&dp->cls, NULL);
465 cmap_init(&dp->flow_table);
467 ovsthread_stats_init(&dp->stats);
469 ovs_mutex_init(&dp->port_mutex);
470 cmap_init(&dp->ports);
471 dp->port_seq = seq_create();
472 latch_init(&dp->exit_latch);
473 fat_rwlock_init(&dp->upcall_rwlock);
475 /* Disable upcalls by default. */
476 dp_netdev_disable_upcall(dp);
477 dp->upcall_cb = NULL;
479 ovs_mutex_lock(&dp->port_mutex);
480 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
481 ovs_mutex_unlock(&dp->port_mutex);
492 dpif_netdev_open(const struct dpif_class *class, const char *name,
493 bool create, struct dpif **dpifp)
495 struct dp_netdev *dp;
498 ovs_mutex_lock(&dp_netdev_mutex);
499 dp = shash_find_data(&dp_netdevs, name);
501 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
503 error = (dp->class != class ? EINVAL
508 *dpifp = create_dpif_netdev(dp);
511 ovs_mutex_unlock(&dp_netdev_mutex);
516 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
517 * through the 'dp_netdevs' shash while freeing 'dp'. */
519 dp_netdev_free(struct dp_netdev *dp)
520 OVS_REQUIRES(dp_netdev_mutex)
522 struct dp_netdev_port *port;
523 struct dp_netdev_stats *bucket;
526 shash_find_and_delete(&dp_netdevs, dp->name);
528 dp_netdev_set_pmd_threads(dp, 0);
529 free(dp->pmd_threads);
531 dp_netdev_flow_flush(dp);
532 ovs_mutex_lock(&dp->port_mutex);
533 CMAP_FOR_EACH (port, node, &dp->ports) {
534 do_del_port(dp, port);
536 ovs_mutex_unlock(&dp->port_mutex);
538 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
539 ovs_mutex_destroy(&bucket->mutex);
540 free_cacheline(bucket);
542 ovsthread_stats_destroy(&dp->stats);
544 classifier_destroy(&dp->cls);
545 cmap_destroy(&dp->flow_table);
546 ovs_mutex_destroy(&dp->flow_mutex);
547 seq_destroy(dp->port_seq);
548 cmap_destroy(&dp->ports);
549 fat_rwlock_destroy(&dp->upcall_rwlock);
550 latch_destroy(&dp->exit_latch);
551 free(CONST_CAST(char *, dp->name));
556 dp_netdev_unref(struct dp_netdev *dp)
559 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
560 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
561 ovs_mutex_lock(&dp_netdev_mutex);
562 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
565 ovs_mutex_unlock(&dp_netdev_mutex);
570 dpif_netdev_close(struct dpif *dpif)
572 struct dp_netdev *dp = get_dp_netdev(dpif);
579 dpif_netdev_destroy(struct dpif *dpif)
581 struct dp_netdev *dp = get_dp_netdev(dpif);
583 if (!atomic_flag_test_and_set(&dp->destroyed)) {
584 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
585 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
594 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
596 struct dp_netdev *dp = get_dp_netdev(dpif);
597 struct dp_netdev_stats *bucket;
600 stats->n_flows = cmap_count(&dp->flow_table);
602 stats->n_hit = stats->n_missed = stats->n_lost = 0;
603 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
604 ovs_mutex_lock(&bucket->mutex);
605 stats->n_hit += bucket->n[DP_STAT_HIT];
606 stats->n_missed += bucket->n[DP_STAT_MISS];
607 stats->n_lost += bucket->n[DP_STAT_LOST];
608 ovs_mutex_unlock(&bucket->mutex);
610 stats->n_masks = UINT32_MAX;
611 stats->n_mask_hit = UINT64_MAX;
617 dp_netdev_reload_pmd_threads(struct dp_netdev *dp)
621 for (i = 0; i < dp->n_pmd_threads; i++) {
622 struct pmd_thread *f = &dp->pmd_threads[i];
625 atomic_add(&f->change_seq, 1, &id);
630 hash_port_no(odp_port_t port_no)
632 return hash_int(odp_to_u32(port_no), 0);
636 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
638 OVS_REQUIRES(dp->port_mutex)
640 struct netdev_saved_flags *sf;
641 struct dp_netdev_port *port;
642 struct netdev *netdev;
643 enum netdev_flags flags;
644 const char *open_type;
648 /* XXX reject devices already in some dp_netdev. */
650 /* Open and validate network device. */
651 open_type = dpif_netdev_port_open_type(dp->class, type);
652 error = netdev_open(devname, open_type, &netdev);
656 /* XXX reject non-Ethernet devices */
658 netdev_get_flags(netdev, &flags);
659 if (flags & NETDEV_LOOPBACK) {
660 VLOG_ERR("%s: cannot add a loopback device", devname);
661 netdev_close(netdev);
665 port = xzalloc(sizeof *port);
666 port->port_no = port_no;
667 port->netdev = netdev;
668 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
669 port->type = xstrdup(type);
670 for (i = 0; i < netdev_n_rxq(netdev); i++) {
671 error = netdev_rxq_open(netdev, &port->rxq[i], i);
673 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
674 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
675 devname, ovs_strerror(errno));
676 netdev_close(netdev);
681 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
683 for (i = 0; i < netdev_n_rxq(netdev); i++) {
684 netdev_rxq_close(port->rxq[i]);
686 netdev_close(netdev);
693 if (netdev_is_pmd(netdev)) {
695 dp_netdev_set_pmd_threads(dp, NR_PMD_THREADS);
696 dp_netdev_reload_pmd_threads(dp);
698 ovs_refcount_init(&port->ref_cnt);
700 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
701 seq_change(dp->port_seq);
707 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
708 odp_port_t *port_nop)
710 struct dp_netdev *dp = get_dp_netdev(dpif);
711 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
712 const char *dpif_port;
716 ovs_mutex_lock(&dp->port_mutex);
717 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
718 if (*port_nop != ODPP_NONE) {
720 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
722 port_no = choose_port(dp, dpif_port);
723 error = port_no == ODPP_NONE ? EFBIG : 0;
727 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
729 ovs_mutex_unlock(&dp->port_mutex);
735 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
737 struct dp_netdev *dp = get_dp_netdev(dpif);
740 ovs_mutex_lock(&dp->port_mutex);
741 if (port_no == ODPP_LOCAL) {
744 struct dp_netdev_port *port;
746 error = get_port_by_number(dp, port_no, &port);
748 do_del_port(dp, port);
751 ovs_mutex_unlock(&dp->port_mutex);
757 is_valid_port_number(odp_port_t port_no)
759 return port_no != ODPP_NONE;
762 static struct dp_netdev_port *
763 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
765 struct dp_netdev_port *port;
767 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
768 if (port->port_no == port_no) {
776 get_port_by_number(struct dp_netdev *dp,
777 odp_port_t port_no, struct dp_netdev_port **portp)
779 if (!is_valid_port_number(port_no)) {
783 *portp = dp_netdev_lookup_port(dp, port_no);
784 return *portp ? 0 : ENOENT;
789 port_ref(struct dp_netdev_port *port)
792 ovs_refcount_ref(&port->ref_cnt);
797 port_destroy__(struct dp_netdev_port *port)
799 int n_rxq = netdev_n_rxq(port->netdev);
802 netdev_close(port->netdev);
803 netdev_restore_flags(port->sf);
805 for (i = 0; i < n_rxq; i++) {
806 netdev_rxq_close(port->rxq[i]);
814 port_unref(struct dp_netdev_port *port)
816 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
817 ovsrcu_postpone(port_destroy__, port);
822 get_port_by_name(struct dp_netdev *dp,
823 const char *devname, struct dp_netdev_port **portp)
824 OVS_REQUIRES(dp->port_mutex)
826 struct dp_netdev_port *port;
828 CMAP_FOR_EACH (port, node, &dp->ports) {
829 if (!strcmp(netdev_get_name(port->netdev), devname)) {
838 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
839 OVS_REQUIRES(dp->port_mutex)
841 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
842 seq_change(dp->port_seq);
843 if (netdev_is_pmd(port->netdev)) {
844 dp_netdev_reload_pmd_threads(dp);
851 answer_port_query(const struct dp_netdev_port *port,
852 struct dpif_port *dpif_port)
854 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
855 dpif_port->type = xstrdup(port->type);
856 dpif_port->port_no = port->port_no;
860 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
861 struct dpif_port *dpif_port)
863 struct dp_netdev *dp = get_dp_netdev(dpif);
864 struct dp_netdev_port *port;
867 error = get_port_by_number(dp, port_no, &port);
868 if (!error && dpif_port) {
869 answer_port_query(port, dpif_port);
876 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
877 struct dpif_port *dpif_port)
879 struct dp_netdev *dp = get_dp_netdev(dpif);
880 struct dp_netdev_port *port;
883 ovs_mutex_lock(&dp->port_mutex);
884 error = get_port_by_name(dp, devname, &port);
885 if (!error && dpif_port) {
886 answer_port_query(port, dpif_port);
888 ovs_mutex_unlock(&dp->port_mutex);
894 dp_netdev_flow_free(struct dp_netdev_flow *flow)
896 struct dp_netdev_flow_stats *bucket;
899 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
900 ovs_mutex_destroy(&bucket->mutex);
901 free_cacheline(bucket);
903 ovsthread_stats_destroy(&flow->stats);
905 cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
906 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
911 dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
912 OVS_REQUIRES(dp->flow_mutex)
914 struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr);
915 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
917 classifier_remove(&dp->cls, cr);
918 cmap_remove(&dp->flow_table, node, flow_hash(&flow->flow, 0));
919 ovsrcu_postpone(dp_netdev_flow_free, flow);
923 dp_netdev_flow_flush(struct dp_netdev *dp)
925 struct dp_netdev_flow *netdev_flow;
927 ovs_mutex_lock(&dp->flow_mutex);
928 CMAP_FOR_EACH (netdev_flow, node, &dp->flow_table) {
929 dp_netdev_remove_flow(dp, netdev_flow);
931 ovs_mutex_unlock(&dp->flow_mutex);
935 dpif_netdev_flow_flush(struct dpif *dpif)
937 struct dp_netdev *dp = get_dp_netdev(dpif);
939 dp_netdev_flow_flush(dp);
943 struct dp_netdev_port_state {
944 struct cmap_position position;
949 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
951 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
956 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
957 struct dpif_port *dpif_port)
959 struct dp_netdev_port_state *state = state_;
960 struct dp_netdev *dp = get_dp_netdev(dpif);
961 struct cmap_node *node;
964 node = cmap_next_position(&dp->ports, &state->position);
966 struct dp_netdev_port *port;
968 port = CONTAINER_OF(node, struct dp_netdev_port, node);
971 state->name = xstrdup(netdev_get_name(port->netdev));
972 dpif_port->name = state->name;
973 dpif_port->type = port->type;
974 dpif_port->port_no = port->port_no;
985 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
987 struct dp_netdev_port_state *state = state_;
994 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
996 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
997 uint64_t new_port_seq;
1000 new_port_seq = seq_read(dpif->dp->port_seq);
1001 if (dpif->last_port_seq != new_port_seq) {
1002 dpif->last_port_seq = new_port_seq;
1012 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1014 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1016 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1019 static struct dp_netdev_flow *
1020 dp_netdev_flow_cast(const struct cls_rule *cr)
1022 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1025 static struct dp_netdev_flow *
1026 dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct miniflow *key)
1028 struct dp_netdev_flow *netdev_flow;
1029 struct cls_rule *rule;
1031 classifier_lookup_miniflow_batch(&dp->cls, &key, &rule, 1);
1032 netdev_flow = dp_netdev_flow_cast(rule);
1037 static struct dp_netdev_flow *
1038 dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
1040 struct dp_netdev_flow *netdev_flow;
1042 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
1044 if (flow_equal(&netdev_flow->flow, flow)) {
1053 get_dpif_flow_stats(struct dp_netdev_flow *netdev_flow,
1054 struct dpif_flow_stats *stats)
1056 struct dp_netdev_flow_stats *bucket;
1059 memset(stats, 0, sizeof *stats);
1060 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1061 ovs_mutex_lock(&bucket->mutex);
1062 stats->n_packets += bucket->packet_count;
1063 stats->n_bytes += bucket->byte_count;
1064 stats->used = MAX(stats->used, bucket->used);
1065 stats->tcp_flags |= bucket->tcp_flags;
1066 ovs_mutex_unlock(&bucket->mutex);
1071 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1072 const struct nlattr *mask_key,
1073 uint32_t mask_key_len, const struct flow *flow,
1077 enum odp_key_fitness fitness;
1079 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1081 /* This should not happen: it indicates that
1082 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1083 * disagree on the acceptable form of a mask. Log the problem
1084 * as an error, with enough details to enable debugging. */
1085 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1087 if (!VLOG_DROP_ERR(&rl)) {
1091 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1093 VLOG_ERR("internal error parsing flow mask %s (%s)",
1094 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1101 enum mf_field_id id;
1102 /* No mask key, unwildcard everything except fields whose
1103 * prerequisities are not met. */
1104 memset(mask, 0x0, sizeof *mask);
1106 for (id = 0; id < MFF_N_IDS; ++id) {
1107 /* Skip registers and metadata. */
1108 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1109 && id != MFF_METADATA) {
1110 const struct mf_field *mf = mf_from_id(id);
1111 if (mf_are_prereqs_ok(mf, flow)) {
1112 mf_mask_field(mf, mask);
1118 /* Force unwildcard the in_port.
1120 * We need to do this even in the case where we unwildcard "everything"
1121 * above because "everything" only includes the 16-bit OpenFlow port number
1122 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1123 * port number mask->in_port.odp_port. */
1124 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1130 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1135 if (odp_flow_key_to_flow(key, key_len, flow)) {
1136 /* This should not happen: it indicates that odp_flow_key_from_flow()
1137 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1138 * flow. Log the problem as an error, with enough details to enable
1140 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1142 if (!VLOG_DROP_ERR(&rl)) {
1146 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1147 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1154 in_port = flow->in_port.odp_port;
1155 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1163 dpif_netdev_flow_get(const struct dpif *dpif,
1164 const struct nlattr *nl_key, size_t nl_key_len,
1165 struct ofpbuf **bufp,
1166 struct nlattr **maskp, size_t *mask_len,
1167 struct nlattr **actionsp, size_t *actions_len,
1168 struct dpif_flow_stats *stats)
1170 struct dp_netdev *dp = get_dp_netdev(dpif);
1171 struct dp_netdev_flow *netdev_flow;
1175 error = dpif_netdev_flow_from_nlattrs(nl_key, nl_key_len, &key);
1180 netdev_flow = dp_netdev_find_flow(dp, &key);
1184 get_dpif_flow_stats(netdev_flow, stats);
1188 struct flow_wildcards wc;
1190 *bufp = ofpbuf_new(sizeof(struct odputil_keybuf));
1191 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1192 odp_flow_key_from_mask(*bufp, &wc.masks, &netdev_flow->flow,
1193 odp_to_u32(wc.masks.in_port.odp_port),
1195 *maskp = ofpbuf_data(*bufp);
1196 *mask_len = ofpbuf_size(*bufp);
1199 struct dp_netdev_actions *actions;
1201 actions = dp_netdev_flow_get_actions(netdev_flow);
1202 *actionsp = actions->actions;
1203 *actions_len = actions->size;
1213 dp_netdev_flow_add(struct dp_netdev *dp, struct match *match,
1214 const struct nlattr *actions, size_t actions_len)
1215 OVS_REQUIRES(dp->flow_mutex)
1217 struct dp_netdev_flow *netdev_flow;
1219 netdev_flow = xzalloc(sizeof *netdev_flow);
1220 *CONST_CAST(struct flow *, &netdev_flow->flow) = match->flow;
1222 ovsthread_stats_init(&netdev_flow->stats);
1224 ovsrcu_set(&netdev_flow->actions,
1225 dp_netdev_actions_create(actions, actions_len));
1227 cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr),
1228 match, NETDEV_RULE_PRIORITY);
1229 cmap_insert(&dp->flow_table,
1230 CONST_CAST(struct cmap_node *, &netdev_flow->node),
1231 flow_hash(&match->flow, 0));
1232 classifier_insert(&dp->cls,
1233 CONST_CAST(struct cls_rule *, &netdev_flow->cr));
1239 clear_stats(struct dp_netdev_flow *netdev_flow)
1241 struct dp_netdev_flow_stats *bucket;
1244 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1245 ovs_mutex_lock(&bucket->mutex);
1247 bucket->packet_count = 0;
1248 bucket->byte_count = 0;
1249 bucket->tcp_flags = 0;
1250 ovs_mutex_unlock(&bucket->mutex);
1255 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1257 struct dp_netdev *dp = get_dp_netdev(dpif);
1258 struct dp_netdev_flow *netdev_flow;
1259 struct miniflow miniflow;
1263 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
1267 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1268 put->mask, put->mask_len,
1269 &match.flow, &match.wc.masks);
1273 miniflow_init(&miniflow, &match.flow);
1275 ovs_mutex_lock(&dp->flow_mutex);
1276 netdev_flow = dp_netdev_lookup_flow(dp, &miniflow);
1278 if (put->flags & DPIF_FP_CREATE) {
1279 if (cmap_count(&dp->flow_table) < MAX_FLOWS) {
1281 memset(put->stats, 0, sizeof *put->stats);
1283 error = dp_netdev_flow_add(dp, &match, put->actions,
1292 if (put->flags & DPIF_FP_MODIFY
1293 && flow_equal(&match.flow, &netdev_flow->flow)) {
1294 struct dp_netdev_actions *new_actions;
1295 struct dp_netdev_actions *old_actions;
1297 new_actions = dp_netdev_actions_create(put->actions,
1300 old_actions = dp_netdev_flow_get_actions(netdev_flow);
1301 ovsrcu_set(&netdev_flow->actions, new_actions);
1304 get_dpif_flow_stats(netdev_flow, put->stats);
1306 if (put->flags & DPIF_FP_ZERO_STATS) {
1307 clear_stats(netdev_flow);
1310 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
1311 } else if (put->flags & DPIF_FP_CREATE) {
1314 /* Overlapping flow. */
1318 ovs_mutex_unlock(&dp->flow_mutex);
1319 miniflow_destroy(&miniflow);
1325 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
1327 struct dp_netdev *dp = get_dp_netdev(dpif);
1328 struct dp_netdev_flow *netdev_flow;
1332 error = dpif_netdev_flow_from_nlattrs(del->key, del->key_len, &key);
1337 ovs_mutex_lock(&dp->flow_mutex);
1338 netdev_flow = dp_netdev_find_flow(dp, &key);
1341 get_dpif_flow_stats(netdev_flow, del->stats);
1343 dp_netdev_remove_flow(dp, netdev_flow);
1347 ovs_mutex_unlock(&dp->flow_mutex);
1352 struct dpif_netdev_flow_dump {
1353 struct dpif_flow_dump up;
1354 struct cmap_position pos;
1356 struct ovs_mutex mutex;
1359 static struct dpif_netdev_flow_dump *
1360 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
1362 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
1365 static struct dpif_flow_dump *
1366 dpif_netdev_flow_dump_create(const struct dpif *dpif_)
1368 struct dpif_netdev_flow_dump *dump;
1370 dump = xmalloc(sizeof *dump);
1371 dpif_flow_dump_init(&dump->up, dpif_);
1372 memset(&dump->pos, 0, sizeof dump->pos);
1374 ovs_mutex_init(&dump->mutex);
1380 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
1382 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1384 ovs_mutex_destroy(&dump->mutex);
1389 struct dpif_netdev_flow_dump_thread {
1390 struct dpif_flow_dump_thread up;
1391 struct dpif_netdev_flow_dump *dump;
1392 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
1393 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
1396 static struct dpif_netdev_flow_dump_thread *
1397 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
1399 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
1402 static struct dpif_flow_dump_thread *
1403 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
1405 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1406 struct dpif_netdev_flow_dump_thread *thread;
1408 thread = xmalloc(sizeof *thread);
1409 dpif_flow_dump_thread_init(&thread->up, &dump->up);
1410 thread->dump = dump;
1415 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
1417 struct dpif_netdev_flow_dump_thread *thread
1418 = dpif_netdev_flow_dump_thread_cast(thread_);
1424 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
1425 struct dpif_flow *flows, int max_flows)
1427 struct dpif_netdev_flow_dump_thread *thread
1428 = dpif_netdev_flow_dump_thread_cast(thread_);
1429 struct dpif_netdev_flow_dump *dump = thread->dump;
1430 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
1431 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
1432 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
1436 ovs_mutex_lock(&dump->mutex);
1437 if (!dump->status) {
1438 for (n_flows = 0; n_flows < MIN(max_flows, FLOW_DUMP_MAX_BATCH);
1440 struct cmap_node *node;
1442 node = cmap_next_position(&dp->flow_table, &dump->pos);
1447 netdev_flows[n_flows] = CONTAINER_OF(node, struct dp_netdev_flow,
1451 ovs_mutex_unlock(&dump->mutex);
1453 for (i = 0; i < n_flows; i++) {
1454 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
1455 struct odputil_keybuf *keybuf = &thread->keybuf[i];
1456 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
1457 struct dpif_flow *f = &flows[i];
1458 struct dp_netdev_actions *dp_actions;
1459 struct flow_wildcards wc;
1462 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1465 ofpbuf_use_stack(&buf, keybuf, sizeof *keybuf);
1466 odp_flow_key_from_flow(&buf, &netdev_flow->flow, &wc.masks,
1467 netdev_flow->flow.in_port.odp_port, true);
1468 f->key = ofpbuf_data(&buf);
1469 f->key_len = ofpbuf_size(&buf);
1472 ofpbuf_use_stack(&buf, maskbuf, sizeof *maskbuf);
1473 odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
1474 odp_to_u32(wc.masks.in_port.odp_port),
1476 f->mask = ofpbuf_data(&buf);
1477 f->mask_len = ofpbuf_size(&buf);
1480 dp_actions = dp_netdev_flow_get_actions(netdev_flow);
1481 f->actions = dp_actions->actions;
1482 f->actions_len = dp_actions->size;
1485 get_dpif_flow_stats(netdev_flow, &f->stats);
1492 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
1494 struct dp_netdev *dp = get_dp_netdev(dpif);
1495 struct dpif_packet packet, *pp;
1496 struct pkt_metadata *md = &execute->md;
1498 if (ofpbuf_size(execute->packet) < ETH_HEADER_LEN ||
1499 ofpbuf_size(execute->packet) > UINT16_MAX) {
1503 packet.ofpbuf = *execute->packet;
1506 dp_netdev_execute_actions(dp, &pp, 1, false, md,
1507 execute->actions, execute->actions_len);
1509 /* Even though may_steal is set to false, some actions could modify or
1510 * reallocate the ofpbuf memory. We need to pass those changes to the
1512 *execute->packet = packet.ofpbuf;
1518 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
1522 for (i = 0; i < n_ops; i++) {
1523 struct dpif_op *op = ops[i];
1526 case DPIF_OP_FLOW_PUT:
1527 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
1530 case DPIF_OP_FLOW_DEL:
1531 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
1534 case DPIF_OP_EXECUTE:
1535 op->error = dpif_netdev_execute(dpif, &op->u.execute);
1542 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
1543 uint32_t queue_id, uint32_t *priority)
1545 *priority = queue_id;
1550 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
1551 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
1553 struct dp_netdev_actions *
1554 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
1556 struct dp_netdev_actions *netdev_actions;
1558 netdev_actions = xmalloc(sizeof *netdev_actions);
1559 netdev_actions->actions = xmemdup(actions, size);
1560 netdev_actions->size = size;
1562 return netdev_actions;
1565 struct dp_netdev_actions *
1566 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
1568 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
1572 dp_netdev_actions_free(struct dp_netdev_actions *actions)
1574 free(actions->actions);
1580 dp_netdev_process_rxq_port(struct dp_netdev *dp,
1581 struct dp_netdev_port *port,
1582 struct netdev_rxq *rxq)
1584 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
1587 error = netdev_rxq_recv(rxq, packets, &cnt);
1589 dp_netdev_port_input(dp, packets, cnt, port->port_no);
1590 } else if (error != EAGAIN && error != EOPNOTSUPP) {
1591 static struct vlog_rate_limit rl
1592 = VLOG_RATE_LIMIT_INIT(1, 5);
1594 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
1595 netdev_get_name(port->netdev),
1596 ovs_strerror(error));
1601 dpif_netdev_run(struct dpif *dpif)
1603 struct dp_netdev_port *port;
1604 struct dp_netdev *dp = get_dp_netdev(dpif);
1606 CMAP_FOR_EACH (port, node, &dp->ports) {
1607 if (!netdev_is_pmd(port->netdev)) {
1610 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1611 dp_netdev_process_rxq_port(dp, port, port->rxq[i]);
1618 dpif_netdev_wait(struct dpif *dpif)
1620 struct dp_netdev_port *port;
1621 struct dp_netdev *dp = get_dp_netdev(dpif);
1623 ovs_mutex_lock(&dp_netdev_mutex);
1624 CMAP_FOR_EACH (port, node, &dp->ports) {
1625 if (!netdev_is_pmd(port->netdev)) {
1628 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1629 netdev_rxq_wait(port->rxq[i]);
1633 ovs_mutex_unlock(&dp_netdev_mutex);
1637 struct dp_netdev_port *port;
1638 struct netdev_rxq *rx;
1642 pmd_load_queues(struct pmd_thread *f,
1643 struct rxq_poll **ppoll_list, int poll_cnt)
1645 struct dp_netdev *dp = f->dp;
1646 struct rxq_poll *poll_list = *ppoll_list;
1647 struct dp_netdev_port *port;
1652 /* Simple scheduler for netdev rx polling. */
1653 for (i = 0; i < poll_cnt; i++) {
1654 port_unref(poll_list[i].port);
1660 CMAP_FOR_EACH (port, node, &f->dp->ports) {
1661 if (netdev_is_pmd(port->netdev)) {
1664 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1665 if ((index % dp->n_pmd_threads) == id) {
1666 poll_list = xrealloc(poll_list, sizeof *poll_list * (poll_cnt + 1));
1669 poll_list[poll_cnt].port = port;
1670 poll_list[poll_cnt].rx = port->rxq[i];
1678 *ppoll_list = poll_list;
1683 pmd_thread_main(void *f_)
1685 struct pmd_thread *f = f_;
1686 struct dp_netdev *dp = f->dp;
1687 unsigned int lc = 0;
1688 struct rxq_poll *poll_list;
1689 unsigned int port_seq;
1696 pmd_thread_setaffinity_cpu(f->id);
1698 poll_cnt = pmd_load_queues(f, &poll_list, poll_cnt);
1699 atomic_read(&f->change_seq, &port_seq);
1702 unsigned int c_port_seq;
1705 for (i = 0; i < poll_cnt; i++) {
1706 dp_netdev_process_rxq_port(dp, poll_list[i].port, poll_list[i].rx);
1712 /* XXX: need completely userspace based signaling method.
1713 * to keep this thread entirely in userspace.
1714 * For now using atomic counter. */
1716 atomic_read_explicit(&f->change_seq, &c_port_seq, memory_order_consume);
1717 if (c_port_seq != port_seq) {
1723 if (!latch_is_set(&f->dp->exit_latch)){
1727 for (i = 0; i < poll_cnt; i++) {
1728 port_unref(poll_list[i].port);
1736 dp_netdev_disable_upcall(struct dp_netdev *dp)
1737 OVS_ACQUIRES(dp->upcall_rwlock)
1739 fat_rwlock_wrlock(&dp->upcall_rwlock);
1743 dpif_netdev_disable_upcall(struct dpif *dpif)
1744 OVS_NO_THREAD_SAFETY_ANALYSIS
1746 struct dp_netdev *dp = get_dp_netdev(dpif);
1747 dp_netdev_disable_upcall(dp);
1751 dp_netdev_enable_upcall(struct dp_netdev *dp)
1752 OVS_RELEASES(dp->upcall_rwlock)
1754 fat_rwlock_unlock(&dp->upcall_rwlock);
1758 dpif_netdev_enable_upcall(struct dpif *dpif)
1759 OVS_NO_THREAD_SAFETY_ANALYSIS
1761 struct dp_netdev *dp = get_dp_netdev(dpif);
1762 dp_netdev_enable_upcall(dp);
1766 dp_netdev_set_pmd_threads(struct dp_netdev *dp, int n)
1770 if (n == dp->n_pmd_threads) {
1774 /* Stop existing threads. */
1775 latch_set(&dp->exit_latch);
1776 dp_netdev_reload_pmd_threads(dp);
1777 for (i = 0; i < dp->n_pmd_threads; i++) {
1778 struct pmd_thread *f = &dp->pmd_threads[i];
1780 xpthread_join(f->thread, NULL);
1782 latch_poll(&dp->exit_latch);
1783 free(dp->pmd_threads);
1785 /* Start new threads. */
1786 dp->pmd_threads = xmalloc(n * sizeof *dp->pmd_threads);
1787 dp->n_pmd_threads = n;
1789 for (i = 0; i < n; i++) {
1790 struct pmd_thread *f = &dp->pmd_threads[i];
1794 atomic_store(&f->change_seq, 1);
1796 /* Each thread will distribute all devices rx-queues among
1798 f->thread = ovs_thread_create("pmd", pmd_thread_main, f);
1804 dp_netdev_flow_stats_new_cb(void)
1806 struct dp_netdev_flow_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1807 ovs_mutex_init(&bucket->mutex);
1812 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow,
1816 long long int now = time_msec();
1817 struct dp_netdev_flow_stats *bucket;
1819 bucket = ovsthread_stats_bucket_get(&netdev_flow->stats,
1820 dp_netdev_flow_stats_new_cb);
1822 ovs_mutex_lock(&bucket->mutex);
1823 bucket->used = MAX(now, bucket->used);
1824 bucket->packet_count += cnt;
1825 bucket->byte_count += size;
1826 bucket->tcp_flags |= tcp_flags;
1827 ovs_mutex_unlock(&bucket->mutex);
1831 dp_netdev_stats_new_cb(void)
1833 struct dp_netdev_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1834 ovs_mutex_init(&bucket->mutex);
1839 dp_netdev_count_packet(struct dp_netdev *dp, enum dp_stat_type type, int cnt)
1841 struct dp_netdev_stats *bucket;
1843 bucket = ovsthread_stats_bucket_get(&dp->stats, dp_netdev_stats_new_cb);
1844 ovs_mutex_lock(&bucket->mutex);
1845 bucket->n[type] += cnt;
1846 ovs_mutex_unlock(&bucket->mutex);
1849 struct packet_batch {
1850 unsigned int packet_count;
1851 unsigned int byte_count;
1854 struct dp_netdev_flow *flow;
1856 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
1857 struct pkt_metadata md;
1861 packet_batch_update(struct packet_batch *batch,
1862 struct dpif_packet *packet, const struct miniflow *mf)
1864 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
1865 batch->packets[batch->packet_count++] = packet;
1866 batch->byte_count += ofpbuf_size(&packet->ofpbuf);
1870 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow,
1871 struct pkt_metadata *md)
1876 batch->packet_count = 0;
1877 batch->byte_count = 0;
1878 batch->tcp_flags = 0;
1882 packet_batch_execute(struct packet_batch *batch, struct dp_netdev *dp)
1884 struct dp_netdev_actions *actions;
1885 struct dp_netdev_flow *flow = batch->flow;
1887 dp_netdev_flow_used(batch->flow, batch->packet_count, batch->byte_count,
1890 actions = dp_netdev_flow_get_actions(flow);
1892 dp_netdev_execute_actions(dp, batch->packets,
1893 batch->packet_count, true, &batch->md,
1894 actions->actions, actions->size);
1896 dp_netdev_count_packet(dp, DP_STAT_HIT, batch->packet_count);
1900 dp_netdev_input(struct dp_netdev *dp, struct dpif_packet **packets, int cnt,
1901 struct pkt_metadata *md)
1903 struct dp_netdev_queue q = DP_NETDEV_QUEUE_INITIALIZER;
1904 struct packet_batch batches[NETDEV_MAX_RX_BATCH];
1905 struct netdev_flow_key keys[NETDEV_MAX_RX_BATCH];
1906 const struct miniflow *mfs[NETDEV_MAX_RX_BATCH]; /* NULL at bad packets. */
1907 struct cls_rule *rules[NETDEV_MAX_RX_BATCH];
1908 size_t n_batches, i;
1910 for (i = 0; i < cnt; i++) {
1911 if (OVS_UNLIKELY(ofpbuf_size(&packets[i]->ofpbuf) < ETH_HEADER_LEN)) {
1912 dpif_packet_delete(packets[i]);
1917 miniflow_initialize(&keys[i].flow, keys[i].buf);
1918 miniflow_extract(&packets[i]->ofpbuf, md, &keys[i].flow);
1919 mfs[i] = &keys[i].flow;
1922 classifier_lookup_miniflow_batch(&dp->cls, mfs, rules, cnt);
1925 for (i = 0; i < cnt; i++) {
1926 struct dp_netdev_flow *flow;
1927 struct packet_batch *batch;
1930 if (OVS_UNLIKELY(!mfs[i])) {
1934 if (OVS_UNLIKELY(!rules[i])) {
1935 struct ofpbuf *buf = &packets[i]->ofpbuf;
1937 dp_netdev_count_packet(dp, DP_STAT_MISS, 1);
1938 dp_netdev_queue_userspace_packet(&q, buf, DPIF_UC_MISS,
1940 dpif_packet_delete(packets[i]);
1944 /* XXX: This O(n^2) algortihm makes sense if we're operating under the
1945 * assumption that the number of distinct flows (and therefore the
1946 * number of distinct batches) is quite small. If this turns out not
1947 * to be the case, it may make sense to pre sort based on the
1948 * netdev_flow pointer. That done we can get the appropriate batching
1949 * in O(n * log(n)) instead. */
1951 flow = dp_netdev_flow_cast(rules[i]);
1952 for (j = 0; j < n_batches; j++) {
1953 if (batches[j].flow == flow) {
1954 batch = &batches[j];
1960 batch = &batches[n_batches++];
1961 packet_batch_init(batch, flow, md);
1963 packet_batch_update(batch, packets[i], mfs[i]);
1966 for (i = 0; i < n_batches; i++) {
1967 packet_batch_execute(&batches[i], dp);
1970 if (q.packet_count) {
1971 dp_netdev_execute_userspace_queue(&q, dp);
1976 dp_netdev_port_input(struct dp_netdev *dp, struct dpif_packet **packets,
1977 int cnt, odp_port_t port_no)
1979 uint32_t *recirc_depth = recirc_depth_get();
1980 struct pkt_metadata md = PKT_METADATA_INITIALIZER(port_no);
1983 dp_netdev_input(dp, packets, cnt, &md);
1987 dp_netdev_queue_userspace_packet(struct dp_netdev_queue *q,
1988 struct ofpbuf *packet, int type,
1989 const struct miniflow *key,
1990 const struct nlattr *userdata)
1992 if (q->packet_count < NETDEV_MAX_RX_BATCH) {
1993 int cnt = q->packet_count;
1994 struct dpif_upcall *upcall = &q->upcalls[cnt];
1995 struct ofpbuf *buf = &q->bufs[cnt];
2000 upcall->type = type;
2002 /* Allocate buffer big enough for everything. */
2003 buf_size = ODPUTIL_FLOW_KEY_BYTES;
2005 buf_size += NLA_ALIGN(userdata->nla_len);
2007 buf_size += ofpbuf_size(packet);
2008 ofpbuf_init(buf, buf_size);
2011 miniflow_expand(key, &flow);
2012 odp_flow_key_from_flow(buf, &flow, NULL, flow.in_port.odp_port, true);
2013 upcall->key = ofpbuf_data(buf);
2014 upcall->key_len = ofpbuf_size(buf);
2018 upcall->userdata = ofpbuf_put(buf, userdata,
2019 NLA_ALIGN(userdata->nla_len));
2022 /* We have to perform a copy of the packet, because we cannot send DPDK
2023 * mbufs to a non pmd thread. When the upcall processing will be done
2024 * in the pmd thread, this copy can be avoided */
2025 data = ofpbuf_put(buf, ofpbuf_data(packet), ofpbuf_size(packet));
2026 ofpbuf_use_stub(&upcall->packet, data, ofpbuf_size(packet));
2027 ofpbuf_set_size(&upcall->packet, ofpbuf_size(packet));
2037 dp_netdev_execute_userspace_queue(struct dp_netdev_queue *q,
2038 struct dp_netdev *dp)
2040 struct dpif_upcall *upcalls = q->upcalls;
2041 struct ofpbuf *bufs = q->bufs;
2042 int cnt = q->packet_count;
2044 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
2045 ovs_assert(dp->upcall_cb);
2046 dp->upcall_cb(dp->dpif, upcalls, bufs, cnt);
2047 fat_rwlock_unlock(&dp->upcall_rwlock);
2051 for (i = 0; i < cnt; i++) {
2052 ofpbuf_uninit(&bufs[i]);
2053 ofpbuf_uninit(&upcalls[i].packet);
2058 struct dp_netdev_execute_aux {
2059 struct dp_netdev *dp;
2063 dpif_netdev_register_upcall_cb(struct dpif *dpif, exec_upcall_cb *cb)
2065 struct dp_netdev *dp = get_dp_netdev(dpif);
2070 dp_execute_cb(void *aux_, struct dpif_packet **packets, int cnt,
2071 struct pkt_metadata *md,
2072 const struct nlattr *a, bool may_steal)
2073 OVS_NO_THREAD_SAFETY_ANALYSIS
2075 struct dp_netdev_execute_aux *aux = aux_;
2076 int type = nl_attr_type(a);
2077 struct dp_netdev_port *p;
2078 uint32_t *depth = recirc_depth_get();
2081 switch ((enum ovs_action_attr)type) {
2082 case OVS_ACTION_ATTR_OUTPUT:
2083 p = dp_netdev_lookup_port(aux->dp, u32_to_odp(nl_attr_get_u32(a)));
2084 if (OVS_LIKELY(p)) {
2085 netdev_send(p->netdev, packets, cnt, may_steal);
2086 } else if (may_steal) {
2087 for (i = 0; i < cnt; i++) {
2088 dpif_packet_delete(packets[i]);
2093 case OVS_ACTION_ATTR_USERSPACE: {
2094 const struct nlattr *userdata;
2095 struct netdev_flow_key key;
2096 struct dp_netdev_queue q = DP_NETDEV_QUEUE_INITIALIZER;
2098 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
2100 miniflow_initialize(&key.flow, key.buf);
2102 for (i = 0; i < cnt; i++) {
2103 struct ofpbuf *packet;
2105 packet = &packets[i]->ofpbuf;
2107 miniflow_extract(packet, md, &key.flow);
2109 dp_netdev_queue_userspace_packet(&q, packet,
2110 DPIF_UC_ACTION, &key.flow,
2113 dpif_packet_delete(packets[i]);
2117 if (q.packet_count) {
2118 dp_netdev_execute_userspace_queue(&q, aux->dp);
2123 case OVS_ACTION_ATTR_HASH: {
2124 const struct ovs_action_hash *hash_act;
2125 struct netdev_flow_key key;
2128 hash_act = nl_attr_get(a);
2130 miniflow_initialize(&key.flow, key.buf);
2132 for (i = 0; i < cnt; i++) {
2134 /* XXX: this is slow. Use RSS hash in the future */
2135 miniflow_extract(&packets[i]->ofpbuf, md, &key.flow);
2137 if (hash_act->hash_alg == OVS_HASH_ALG_L4) {
2138 /* Hash need not be symmetric, nor does it need to include
2140 hash = miniflow_hash_5tuple(&key.flow, hash_act->hash_basis);
2142 VLOG_WARN("Unknown hash algorithm specified "
2143 "for the hash action.");
2148 hash = 1; /* 0 is not valid */
2154 packets[i]->dp_hash = hash;
2159 case OVS_ACTION_ATTR_RECIRC:
2160 if (*depth < MAX_RECIRC_DEPTH) {
2163 for (i = 0; i < cnt; i++) {
2164 struct dpif_packet *recirc_pkt;
2165 struct pkt_metadata recirc_md = *md;
2167 recirc_pkt = (may_steal) ? packets[i]
2168 : dpif_packet_clone(packets[i]);
2170 recirc_md.recirc_id = nl_attr_get_u32(a);
2172 /* Hash is private to each packet */
2173 recirc_md.dp_hash = packets[i]->dp_hash;
2175 dp_netdev_input(aux->dp, &recirc_pkt, 1, &recirc_md);
2181 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
2183 for (i = 0; i < cnt; i++) {
2184 dpif_packet_delete(packets[i]);
2190 case OVS_ACTION_ATTR_PUSH_VLAN:
2191 case OVS_ACTION_ATTR_POP_VLAN:
2192 case OVS_ACTION_ATTR_PUSH_MPLS:
2193 case OVS_ACTION_ATTR_POP_MPLS:
2194 case OVS_ACTION_ATTR_SET:
2195 case OVS_ACTION_ATTR_SAMPLE:
2196 case OVS_ACTION_ATTR_UNSPEC:
2197 case __OVS_ACTION_ATTR_MAX:
2203 dp_netdev_execute_actions(struct dp_netdev *dp,
2204 struct dpif_packet **packets, int cnt,
2205 bool may_steal, struct pkt_metadata *md,
2206 const struct nlattr *actions, size_t actions_len)
2208 struct dp_netdev_execute_aux aux = {dp};
2210 odp_execute_actions(&aux, packets, cnt, may_steal, md, actions,
2211 actions_len, dp_execute_cb);
2214 const struct dpif_class dpif_netdev_class = {
2216 dpif_netdev_enumerate,
2217 dpif_netdev_port_open_type,
2220 dpif_netdev_destroy,
2223 dpif_netdev_get_stats,
2224 dpif_netdev_port_add,
2225 dpif_netdev_port_del,
2226 dpif_netdev_port_query_by_number,
2227 dpif_netdev_port_query_by_name,
2228 NULL, /* port_get_pid */
2229 dpif_netdev_port_dump_start,
2230 dpif_netdev_port_dump_next,
2231 dpif_netdev_port_dump_done,
2232 dpif_netdev_port_poll,
2233 dpif_netdev_port_poll_wait,
2234 dpif_netdev_flow_get,
2235 dpif_netdev_flow_flush,
2236 dpif_netdev_flow_dump_create,
2237 dpif_netdev_flow_dump_destroy,
2238 dpif_netdev_flow_dump_thread_create,
2239 dpif_netdev_flow_dump_thread_destroy,
2240 dpif_netdev_flow_dump_next,
2241 dpif_netdev_operate,
2242 NULL, /* recv_set */
2243 NULL, /* handlers_set */
2244 dpif_netdev_queue_to_priority,
2246 NULL, /* recv_wait */
2247 NULL, /* recv_purge */
2248 dpif_netdev_register_upcall_cb,
2249 dpif_netdev_enable_upcall,
2250 dpif_netdev_disable_upcall,
2254 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
2255 const char *argv[], void *aux OVS_UNUSED)
2257 struct dp_netdev_port *old_port;
2258 struct dp_netdev_port *new_port;
2259 struct dp_netdev *dp;
2262 ovs_mutex_lock(&dp_netdev_mutex);
2263 dp = shash_find_data(&dp_netdevs, argv[1]);
2264 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2265 ovs_mutex_unlock(&dp_netdev_mutex);
2266 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2269 ovs_refcount_ref(&dp->ref_cnt);
2270 ovs_mutex_unlock(&dp_netdev_mutex);
2272 ovs_mutex_lock(&dp->port_mutex);
2273 if (get_port_by_name(dp, argv[2], &old_port)) {
2274 unixctl_command_reply_error(conn, "unknown port");
2278 port_no = u32_to_odp(atoi(argv[3]));
2279 if (!port_no || port_no == ODPP_NONE) {
2280 unixctl_command_reply_error(conn, "bad port number");
2283 if (dp_netdev_lookup_port(dp, port_no)) {
2284 unixctl_command_reply_error(conn, "port number already in use");
2288 /* Remove old port. */
2289 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
2290 ovsrcu_postpone(free, old_port);
2292 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
2293 new_port = xmemdup(old_port, sizeof *old_port);
2294 new_port->port_no = port_no;
2295 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
2297 seq_change(dp->port_seq);
2298 unixctl_command_reply(conn, NULL);
2301 ovs_mutex_unlock(&dp->port_mutex);
2302 dp_netdev_unref(dp);
2306 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
2307 const char *argv[], void *aux OVS_UNUSED)
2309 struct dp_netdev_port *port;
2310 struct dp_netdev *dp;
2312 ovs_mutex_lock(&dp_netdev_mutex);
2313 dp = shash_find_data(&dp_netdevs, argv[1]);
2314 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2315 ovs_mutex_unlock(&dp_netdev_mutex);
2316 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2319 ovs_refcount_ref(&dp->ref_cnt);
2320 ovs_mutex_unlock(&dp_netdev_mutex);
2322 ovs_mutex_lock(&dp->port_mutex);
2323 if (get_port_by_name(dp, argv[2], &port)) {
2324 unixctl_command_reply_error(conn, "unknown port");
2325 } else if (port->port_no == ODPP_LOCAL) {
2326 unixctl_command_reply_error(conn, "can't delete local port");
2328 do_del_port(dp, port);
2329 unixctl_command_reply(conn, NULL);
2331 ovs_mutex_unlock(&dp->port_mutex);
2333 dp_netdev_unref(dp);
2337 dpif_dummy_register__(const char *type)
2339 struct dpif_class *class;
2341 class = xmalloc(sizeof *class);
2342 *class = dpif_netdev_class;
2343 class->type = xstrdup(type);
2344 dp_register_provider(class);
2348 dpif_dummy_register(bool override)
2355 dp_enumerate_types(&types);
2356 SSET_FOR_EACH (type, &types) {
2357 if (!dp_unregister_provider(type)) {
2358 dpif_dummy_register__(type);
2361 sset_destroy(&types);
2364 dpif_dummy_register__("dummy");
2366 unixctl_command_register("dpif-dummy/change-port-number",
2367 "DP PORT NEW-NUMBER",
2368 3, 3, dpif_dummy_change_port_number, NULL);
2369 unixctl_command_register("dpif-dummy/delete-port", "DP PORT",
2370 2, 2, dpif_dummy_delete_port, NULL);