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 static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600);
90 /* Datapath based on the network device interface from netdev.h.
96 * Some members, marked 'const', are immutable. Accessing other members
97 * requires synchronization, as noted in more detail below.
99 * Acquisition order is, from outermost to innermost:
101 * dp_netdev_mutex (global)
106 const struct dpif_class *const class;
107 const char *const name;
109 struct ovs_refcount ref_cnt;
110 atomic_flag destroyed;
114 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
115 * changes to 'cls' must be made while still holding the 'flow_mutex'.
117 struct ovs_mutex flow_mutex;
118 struct classifier cls;
119 struct cmap flow_table OVS_GUARDED; /* Flow table. */
123 * ovsthread_stats is internally synchronized. */
124 struct ovsthread_stats stats; /* Contains 'struct dp_netdev_stats *'. */
128 * Protected by RCU. Take the mutex to add or remove ports. */
129 struct ovs_mutex port_mutex;
131 struct seq *port_seq; /* Incremented whenever a port changes. */
133 /* Protects access to ofproto-dpif-upcall interface during revalidator
134 * thread synchronization. */
135 struct fat_rwlock upcall_rwlock;
136 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
139 /* Forwarding threads. */
140 struct latch exit_latch;
141 struct pmd_thread *pmd_threads;
142 size_t n_pmd_threads;
146 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
150 DP_STAT_HIT, /* Packets that matched in the flow table. */
151 DP_STAT_MISS, /* Packets that did not match. */
152 DP_STAT_LOST, /* Packets not passed up to the client. */
156 /* Contained by struct dp_netdev's 'stats' member. */
157 struct dp_netdev_stats {
158 struct ovs_mutex mutex; /* Protects 'n'. */
160 /* Indexed by DP_STAT_*, protected by 'mutex'. */
161 unsigned long long int n[DP_N_STATS] OVS_GUARDED;
165 /* A port in a netdev-based datapath. */
166 struct dp_netdev_port {
167 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
169 struct netdev *netdev;
170 struct netdev_saved_flags *sf;
171 struct netdev_rxq **rxq;
172 struct ovs_refcount ref_cnt;
173 char *type; /* Port type as requested by user. */
177 /* Stores a miniflow */
179 /* There are fields in the flow structure that we never use. Therefore we can
180 * save a few words of memory */
181 #define NETDEV_KEY_BUF_SIZE_U32 (FLOW_U32S - MINI_N_INLINE \
182 - FLOW_U32_SIZE(regs) \
183 - FLOW_U32_SIZE(metadata) \
185 struct netdev_flow_key {
186 struct miniflow flow;
187 uint32_t buf[NETDEV_KEY_BUF_SIZE_U32];
190 /* A flow in dp_netdev's 'flow_table'.
196 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
197 * its dp_netdev's classifier. The text below calls this classifier 'cls'.
202 * The thread safety rules described here for "struct dp_netdev_flow" are
203 * motivated by two goals:
205 * - Prevent threads that read members of "struct dp_netdev_flow" from
206 * reading bad data due to changes by some thread concurrently modifying
209 * - Prevent two threads making changes to members of a given "struct
210 * dp_netdev_flow" from interfering with each other.
216 * A flow 'flow' may be accessed without a risk of being freed during an RCU
217 * grace period. Code that needs to hold onto a flow for a while
218 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
220 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
221 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
224 * Some members, marked 'const', are immutable. Accessing other members
225 * requires synchronization, as noted in more detail below.
227 struct dp_netdev_flow {
228 /* Packet classification. */
229 const struct cls_rule cr; /* In owning dp_netdev's 'cls'. */
231 /* Hash table index by unmasked flow. */
232 const struct cmap_node node; /* In owning dp_netdev's 'flow_table'. */
233 const struct flow flow; /* The flow that created this entry. */
235 /* Number of references.
236 * The classifier owns one reference.
237 * Any thread trying to keep a rule from being freed should hold its own
239 struct ovs_refcount ref_cnt;
243 * Reading or writing these members requires 'mutex'. */
244 struct ovsthread_stats stats; /* Contains "struct dp_netdev_flow_stats". */
247 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
250 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
252 /* Contained by struct dp_netdev_flow's 'stats' member. */
253 struct dp_netdev_flow_stats {
254 struct ovs_mutex mutex; /* Guards all the other members. */
256 long long int used OVS_GUARDED; /* Last used time, in monotonic msecs. */
257 long long int packet_count OVS_GUARDED; /* Number of packets matched. */
258 long long int byte_count OVS_GUARDED; /* Number of bytes matched. */
259 uint16_t tcp_flags OVS_GUARDED; /* Bitwise-OR of seen tcp_flags values. */
262 /* A set of datapath actions within a "struct dp_netdev_flow".
268 * A struct dp_netdev_actions 'actions' is protected with RCU. */
269 struct dp_netdev_actions {
270 /* These members are immutable: they do not change during the struct's
272 struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */
273 unsigned int size; /* Size of 'actions', in bytes. */
276 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
278 struct dp_netdev_actions *dp_netdev_flow_get_actions(
279 const struct dp_netdev_flow *);
280 static void dp_netdev_actions_free(struct dp_netdev_actions *);
282 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
283 * the performance overhead of interrupt processing. Therefore netdev can
284 * not implement rx-wait for these devices. dpif-netdev needs to poll
285 * these device to check for recv buffer. pmd-thread does polling for
286 * devices assigned to itself thread.
288 * DPDK used PMD for accessing NIC.
290 * A thread that receives packets from PMD ports, looks them up in the flow
291 * table, and executes the actions it finds.
294 struct dp_netdev *dp;
297 atomic_uint change_seq;
300 #define PMD_INITIAL_SEQ 1
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 void dp_netdev_execute_actions(struct dp_netdev *dp,
324 struct dpif_packet **, int c,
325 bool may_steal, struct pkt_metadata *,
326 const struct nlattr *actions,
328 static void dp_netdev_port_input(struct dp_netdev *dp,
329 struct dpif_packet **packets, int cnt,
332 static void dp_netdev_set_pmd_threads(struct dp_netdev *, int n);
333 static void dp_netdev_disable_upcall(struct dp_netdev *);
335 static struct dpif_netdev *
336 dpif_netdev_cast(const struct dpif *dpif)
338 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
339 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
342 static struct dp_netdev *
343 get_dp_netdev(const struct dpif *dpif)
345 return dpif_netdev_cast(dpif)->dp;
349 dpif_netdev_enumerate(struct sset *all_dps,
350 const struct dpif_class *dpif_class)
352 struct shash_node *node;
354 ovs_mutex_lock(&dp_netdev_mutex);
355 SHASH_FOR_EACH(node, &dp_netdevs) {
356 struct dp_netdev *dp = node->data;
357 if (dpif_class != dp->class) {
358 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
359 * If the class doesn't match, skip this dpif. */
362 sset_add(all_dps, node->name);
364 ovs_mutex_unlock(&dp_netdev_mutex);
370 dpif_netdev_class_is_dummy(const struct dpif_class *class)
372 return class != &dpif_netdev_class;
376 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
378 return strcmp(type, "internal") ? type
379 : dpif_netdev_class_is_dummy(class) ? "dummy"
384 create_dpif_netdev(struct dp_netdev *dp)
386 uint16_t netflow_id = hash_string(dp->name, 0);
387 struct dpif_netdev *dpif;
389 ovs_refcount_ref(&dp->ref_cnt);
391 dpif = xmalloc(sizeof *dpif);
392 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
394 dpif->last_port_seq = seq_read(dp->port_seq);
399 /* Choose an unused, non-zero port number and return it on success.
400 * Return ODPP_NONE on failure. */
402 choose_port(struct dp_netdev *dp, const char *name)
403 OVS_REQUIRES(dp->port_mutex)
407 if (dp->class != &dpif_netdev_class) {
411 /* If the port name begins with "br", start the number search at
412 * 100 to make writing tests easier. */
413 if (!strncmp(name, "br", 2)) {
417 /* If the port name contains a number, try to assign that port number.
418 * This can make writing unit tests easier because port numbers are
420 for (p = name; *p != '\0'; p++) {
421 if (isdigit((unsigned char) *p)) {
422 port_no = start_no + strtol(p, NULL, 10);
423 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
424 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
425 return u32_to_odp(port_no);
432 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
433 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
434 return u32_to_odp(port_no);
442 create_dp_netdev(const char *name, const struct dpif_class *class,
443 struct dp_netdev **dpp)
444 OVS_REQUIRES(dp_netdev_mutex)
446 struct dp_netdev *dp;
449 dp = xzalloc(sizeof *dp);
450 shash_add(&dp_netdevs, name, dp);
452 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
453 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
454 ovs_refcount_init(&dp->ref_cnt);
455 atomic_flag_clear(&dp->destroyed);
457 ovs_mutex_init(&dp->flow_mutex);
458 classifier_init(&dp->cls, NULL);
459 cmap_init(&dp->flow_table);
461 ovsthread_stats_init(&dp->stats);
463 ovs_mutex_init(&dp->port_mutex);
464 cmap_init(&dp->ports);
465 dp->port_seq = seq_create();
466 latch_init(&dp->exit_latch);
467 fat_rwlock_init(&dp->upcall_rwlock);
469 /* Disable upcalls by default. */
470 dp_netdev_disable_upcall(dp);
471 dp->upcall_aux = NULL;
472 dp->upcall_cb = NULL;
474 ovs_mutex_lock(&dp->port_mutex);
475 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
476 ovs_mutex_unlock(&dp->port_mutex);
487 dpif_netdev_open(const struct dpif_class *class, const char *name,
488 bool create, struct dpif **dpifp)
490 struct dp_netdev *dp;
493 ovs_mutex_lock(&dp_netdev_mutex);
494 dp = shash_find_data(&dp_netdevs, name);
496 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
498 error = (dp->class != class ? EINVAL
503 *dpifp = create_dpif_netdev(dp);
506 ovs_mutex_unlock(&dp_netdev_mutex);
511 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
512 * through the 'dp_netdevs' shash while freeing 'dp'. */
514 dp_netdev_free(struct dp_netdev *dp)
515 OVS_REQUIRES(dp_netdev_mutex)
517 struct dp_netdev_port *port;
518 struct dp_netdev_stats *bucket;
521 shash_find_and_delete(&dp_netdevs, dp->name);
523 dp_netdev_set_pmd_threads(dp, 0);
524 free(dp->pmd_threads);
526 dp_netdev_flow_flush(dp);
527 ovs_mutex_lock(&dp->port_mutex);
528 CMAP_FOR_EACH (port, node, &dp->ports) {
529 do_del_port(dp, port);
531 ovs_mutex_unlock(&dp->port_mutex);
533 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
534 ovs_mutex_destroy(&bucket->mutex);
535 free_cacheline(bucket);
537 ovsthread_stats_destroy(&dp->stats);
539 classifier_destroy(&dp->cls);
540 cmap_destroy(&dp->flow_table);
541 ovs_mutex_destroy(&dp->flow_mutex);
542 seq_destroy(dp->port_seq);
543 cmap_destroy(&dp->ports);
544 fat_rwlock_destroy(&dp->upcall_rwlock);
545 latch_destroy(&dp->exit_latch);
546 free(CONST_CAST(char *, dp->name));
551 dp_netdev_unref(struct dp_netdev *dp)
554 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
555 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
556 ovs_mutex_lock(&dp_netdev_mutex);
557 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
560 ovs_mutex_unlock(&dp_netdev_mutex);
565 dpif_netdev_close(struct dpif *dpif)
567 struct dp_netdev *dp = get_dp_netdev(dpif);
574 dpif_netdev_destroy(struct dpif *dpif)
576 struct dp_netdev *dp = get_dp_netdev(dpif);
578 if (!atomic_flag_test_and_set(&dp->destroyed)) {
579 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
580 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
589 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
591 struct dp_netdev *dp = get_dp_netdev(dpif);
592 struct dp_netdev_stats *bucket;
595 stats->n_flows = cmap_count(&dp->flow_table);
597 stats->n_hit = stats->n_missed = stats->n_lost = 0;
598 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
599 ovs_mutex_lock(&bucket->mutex);
600 stats->n_hit += bucket->n[DP_STAT_HIT];
601 stats->n_missed += bucket->n[DP_STAT_MISS];
602 stats->n_lost += bucket->n[DP_STAT_LOST];
603 ovs_mutex_unlock(&bucket->mutex);
605 stats->n_masks = UINT32_MAX;
606 stats->n_mask_hit = UINT64_MAX;
612 dp_netdev_reload_pmd_threads(struct dp_netdev *dp)
616 for (i = 0; i < dp->n_pmd_threads; i++) {
617 struct pmd_thread *f = &dp->pmd_threads[i];
620 atomic_add_relaxed(&f->change_seq, 1, &old_seq);
625 hash_port_no(odp_port_t port_no)
627 return hash_int(odp_to_u32(port_no), 0);
631 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
633 OVS_REQUIRES(dp->port_mutex)
635 struct netdev_saved_flags *sf;
636 struct dp_netdev_port *port;
637 struct netdev *netdev;
638 enum netdev_flags flags;
639 const char *open_type;
643 /* XXX reject devices already in some dp_netdev. */
645 /* Open and validate network device. */
646 open_type = dpif_netdev_port_open_type(dp->class, type);
647 error = netdev_open(devname, open_type, &netdev);
651 /* XXX reject non-Ethernet devices */
653 netdev_get_flags(netdev, &flags);
654 if (flags & NETDEV_LOOPBACK) {
655 VLOG_ERR("%s: cannot add a loopback device", devname);
656 netdev_close(netdev);
660 port = xzalloc(sizeof *port);
661 port->port_no = port_no;
662 port->netdev = netdev;
663 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
664 port->type = xstrdup(type);
665 for (i = 0; i < netdev_n_rxq(netdev); i++) {
666 error = netdev_rxq_open(netdev, &port->rxq[i], i);
668 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
669 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
670 devname, ovs_strerror(errno));
671 netdev_close(netdev);
679 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
681 for (i = 0; i < netdev_n_rxq(netdev); i++) {
682 netdev_rxq_close(port->rxq[i]);
684 netdev_close(netdev);
692 if (netdev_is_pmd(netdev)) {
694 dp_netdev_set_pmd_threads(dp, NR_PMD_THREADS);
695 dp_netdev_reload_pmd_threads(dp);
697 ovs_refcount_init(&port->ref_cnt);
699 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
700 seq_change(dp->port_seq);
706 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
707 odp_port_t *port_nop)
709 struct dp_netdev *dp = get_dp_netdev(dpif);
710 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
711 const char *dpif_port;
715 ovs_mutex_lock(&dp->port_mutex);
716 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
717 if (*port_nop != ODPP_NONE) {
719 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
721 port_no = choose_port(dp, dpif_port);
722 error = port_no == ODPP_NONE ? EFBIG : 0;
726 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
728 ovs_mutex_unlock(&dp->port_mutex);
734 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
736 struct dp_netdev *dp = get_dp_netdev(dpif);
739 ovs_mutex_lock(&dp->port_mutex);
740 if (port_no == ODPP_LOCAL) {
743 struct dp_netdev_port *port;
745 error = get_port_by_number(dp, port_no, &port);
747 do_del_port(dp, port);
750 ovs_mutex_unlock(&dp->port_mutex);
756 is_valid_port_number(odp_port_t port_no)
758 return port_no != ODPP_NONE;
761 static struct dp_netdev_port *
762 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
764 struct dp_netdev_port *port;
766 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
767 if (port->port_no == port_no) {
775 get_port_by_number(struct dp_netdev *dp,
776 odp_port_t port_no, struct dp_netdev_port **portp)
778 if (!is_valid_port_number(port_no)) {
782 *portp = dp_netdev_lookup_port(dp, port_no);
783 return *portp ? 0 : ENOENT;
788 port_ref(struct dp_netdev_port *port)
791 ovs_refcount_ref(&port->ref_cnt);
796 port_destroy__(struct dp_netdev_port *port)
798 int n_rxq = netdev_n_rxq(port->netdev);
801 netdev_close(port->netdev);
802 netdev_restore_flags(port->sf);
804 for (i = 0; i < n_rxq; i++) {
805 netdev_rxq_close(port->rxq[i]);
813 port_unref(struct dp_netdev_port *port)
815 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
816 ovsrcu_postpone(port_destroy__, port);
821 get_port_by_name(struct dp_netdev *dp,
822 const char *devname, struct dp_netdev_port **portp)
823 OVS_REQUIRES(dp->port_mutex)
825 struct dp_netdev_port *port;
827 CMAP_FOR_EACH (port, node, &dp->ports) {
828 if (!strcmp(netdev_get_name(port->netdev), devname)) {
837 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
838 OVS_REQUIRES(dp->port_mutex)
840 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
841 seq_change(dp->port_seq);
842 if (netdev_is_pmd(port->netdev)) {
843 dp_netdev_reload_pmd_threads(dp);
850 answer_port_query(const struct dp_netdev_port *port,
851 struct dpif_port *dpif_port)
853 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
854 dpif_port->type = xstrdup(port->type);
855 dpif_port->port_no = port->port_no;
859 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
860 struct dpif_port *dpif_port)
862 struct dp_netdev *dp = get_dp_netdev(dpif);
863 struct dp_netdev_port *port;
866 error = get_port_by_number(dp, port_no, &port);
867 if (!error && dpif_port) {
868 answer_port_query(port, dpif_port);
875 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
876 struct dpif_port *dpif_port)
878 struct dp_netdev *dp = get_dp_netdev(dpif);
879 struct dp_netdev_port *port;
882 ovs_mutex_lock(&dp->port_mutex);
883 error = get_port_by_name(dp, devname, &port);
884 if (!error && dpif_port) {
885 answer_port_query(port, dpif_port);
887 ovs_mutex_unlock(&dp->port_mutex);
893 dp_netdev_flow_free(struct dp_netdev_flow *flow)
895 struct dp_netdev_flow_stats *bucket;
898 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
899 ovs_mutex_destroy(&bucket->mutex);
900 free_cacheline(bucket);
902 ovsthread_stats_destroy(&flow->stats);
904 cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
905 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
909 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
911 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
912 ovsrcu_postpone(dp_netdev_flow_free, flow);
917 dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
918 OVS_REQUIRES(dp->flow_mutex)
920 struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr);
921 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
923 classifier_remove(&dp->cls, cr);
924 cmap_remove(&dp->flow_table, node, flow_hash(&flow->flow, 0));
926 dp_netdev_flow_unref(flow);
930 dp_netdev_flow_flush(struct dp_netdev *dp)
932 struct dp_netdev_flow *netdev_flow;
934 ovs_mutex_lock(&dp->flow_mutex);
935 CMAP_FOR_EACH (netdev_flow, node, &dp->flow_table) {
936 dp_netdev_remove_flow(dp, netdev_flow);
938 ovs_mutex_unlock(&dp->flow_mutex);
942 dpif_netdev_flow_flush(struct dpif *dpif)
944 struct dp_netdev *dp = get_dp_netdev(dpif);
946 dp_netdev_flow_flush(dp);
950 struct dp_netdev_port_state {
951 struct cmap_position position;
956 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
958 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
963 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
964 struct dpif_port *dpif_port)
966 struct dp_netdev_port_state *state = state_;
967 struct dp_netdev *dp = get_dp_netdev(dpif);
968 struct cmap_node *node;
971 node = cmap_next_position(&dp->ports, &state->position);
973 struct dp_netdev_port *port;
975 port = CONTAINER_OF(node, struct dp_netdev_port, node);
978 state->name = xstrdup(netdev_get_name(port->netdev));
979 dpif_port->name = state->name;
980 dpif_port->type = port->type;
981 dpif_port->port_no = port->port_no;
992 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
994 struct dp_netdev_port_state *state = state_;
1001 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1003 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1004 uint64_t new_port_seq;
1007 new_port_seq = seq_read(dpif->dp->port_seq);
1008 if (dpif->last_port_seq != new_port_seq) {
1009 dpif->last_port_seq = new_port_seq;
1019 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1021 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1023 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1026 static struct dp_netdev_flow *
1027 dp_netdev_flow_cast(const struct cls_rule *cr)
1029 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1032 static struct dp_netdev_flow *
1033 dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct miniflow *key)
1035 struct dp_netdev_flow *netdev_flow;
1036 struct cls_rule *rule;
1038 classifier_lookup_miniflow_batch(&dp->cls, &key, &rule, 1);
1039 netdev_flow = dp_netdev_flow_cast(rule);
1044 static struct dp_netdev_flow *
1045 dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
1047 struct dp_netdev_flow *netdev_flow;
1049 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
1051 if (flow_equal(&netdev_flow->flow, flow)) {
1060 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow,
1061 struct dpif_flow_stats *stats)
1063 struct dp_netdev_flow_stats *bucket;
1066 memset(stats, 0, sizeof *stats);
1067 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1068 ovs_mutex_lock(&bucket->mutex);
1069 stats->n_packets += bucket->packet_count;
1070 stats->n_bytes += bucket->byte_count;
1071 stats->used = MAX(stats->used, bucket->used);
1072 stats->tcp_flags |= bucket->tcp_flags;
1073 ovs_mutex_unlock(&bucket->mutex);
1078 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1079 struct ofpbuf *buffer, struct dpif_flow *flow)
1081 struct flow_wildcards wc;
1082 struct dp_netdev_actions *actions;
1084 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1085 odp_flow_key_from_mask(buffer, &wc.masks, &netdev_flow->flow,
1086 odp_to_u32(wc.masks.in_port.odp_port),
1088 flow->mask = ofpbuf_data(buffer);
1089 flow->mask_len = ofpbuf_size(buffer);
1091 actions = dp_netdev_flow_get_actions(netdev_flow);
1092 flow->actions = actions->actions;
1093 flow->actions_len = actions->size;
1095 get_dpif_flow_stats(netdev_flow, &flow->stats);
1099 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1100 const struct nlattr *mask_key,
1101 uint32_t mask_key_len, const struct flow *flow,
1105 enum odp_key_fitness fitness;
1107 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1109 /* This should not happen: it indicates that
1110 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1111 * disagree on the acceptable form of a mask. Log the problem
1112 * as an error, with enough details to enable debugging. */
1113 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1115 if (!VLOG_DROP_ERR(&rl)) {
1119 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1121 VLOG_ERR("internal error parsing flow mask %s (%s)",
1122 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1129 enum mf_field_id id;
1130 /* No mask key, unwildcard everything except fields whose
1131 * prerequisities are not met. */
1132 memset(mask, 0x0, sizeof *mask);
1134 for (id = 0; id < MFF_N_IDS; ++id) {
1135 /* Skip registers and metadata. */
1136 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1137 && id != MFF_METADATA) {
1138 const struct mf_field *mf = mf_from_id(id);
1139 if (mf_are_prereqs_ok(mf, flow)) {
1140 mf_mask_field(mf, mask);
1146 /* Force unwildcard the in_port.
1148 * We need to do this even in the case where we unwildcard "everything"
1149 * above because "everything" only includes the 16-bit OpenFlow port number
1150 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1151 * port number mask->in_port.odp_port. */
1152 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1158 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1163 if (odp_flow_key_to_flow(key, key_len, flow)) {
1164 /* This should not happen: it indicates that odp_flow_key_from_flow()
1165 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1166 * flow. Log the problem as an error, with enough details to enable
1168 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1170 if (!VLOG_DROP_ERR(&rl)) {
1174 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1175 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1182 in_port = flow->in_port.odp_port;
1183 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1191 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1193 struct dp_netdev *dp = get_dp_netdev(dpif);
1194 struct dp_netdev_flow *netdev_flow;
1198 error = dpif_netdev_flow_from_nlattrs(get->key, get->key_len, &key);
1203 netdev_flow = dp_netdev_find_flow(dp, &key);
1206 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->flow);
1215 dp_netdev_flow_add(struct dp_netdev *dp, struct match *match,
1216 const struct nlattr *actions, size_t actions_len)
1217 OVS_REQUIRES(dp->flow_mutex)
1219 struct dp_netdev_flow *netdev_flow;
1221 netdev_flow = xzalloc(sizeof *netdev_flow);
1222 *CONST_CAST(struct flow *, &netdev_flow->flow) = match->flow;
1224 ovs_refcount_init(&netdev_flow->ref_cnt);
1226 ovsthread_stats_init(&netdev_flow->stats);
1228 ovsrcu_set(&netdev_flow->actions,
1229 dp_netdev_actions_create(actions, actions_len));
1231 cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr),
1232 match, NETDEV_RULE_PRIORITY);
1233 cmap_insert(&dp->flow_table,
1234 CONST_CAST(struct cmap_node *, &netdev_flow->node),
1235 flow_hash(&match->flow, 0));
1236 classifier_insert(&dp->cls,
1237 CONST_CAST(struct cls_rule *, &netdev_flow->cr));
1239 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
1240 struct ds ds = DS_EMPTY_INITIALIZER;
1242 ds_put_cstr(&ds, "flow_add: ");
1243 match_format(match, &ds, OFP_DEFAULT_PRIORITY);
1244 ds_put_cstr(&ds, ", actions:");
1245 format_odp_actions(&ds, actions, actions_len);
1247 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
1256 clear_stats(struct dp_netdev_flow *netdev_flow)
1258 struct dp_netdev_flow_stats *bucket;
1261 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1262 ovs_mutex_lock(&bucket->mutex);
1264 bucket->packet_count = 0;
1265 bucket->byte_count = 0;
1266 bucket->tcp_flags = 0;
1267 ovs_mutex_unlock(&bucket->mutex);
1272 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1274 struct dp_netdev *dp = get_dp_netdev(dpif);
1275 struct dp_netdev_flow *netdev_flow;
1276 struct miniflow miniflow;
1280 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
1284 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1285 put->mask, put->mask_len,
1286 &match.flow, &match.wc.masks);
1290 miniflow_init(&miniflow, &match.flow);
1292 ovs_mutex_lock(&dp->flow_mutex);
1293 netdev_flow = dp_netdev_lookup_flow(dp, &miniflow);
1295 if (put->flags & DPIF_FP_CREATE) {
1296 if (cmap_count(&dp->flow_table) < MAX_FLOWS) {
1298 memset(put->stats, 0, sizeof *put->stats);
1300 error = dp_netdev_flow_add(dp, &match, put->actions,
1309 if (put->flags & DPIF_FP_MODIFY
1310 && flow_equal(&match.flow, &netdev_flow->flow)) {
1311 struct dp_netdev_actions *new_actions;
1312 struct dp_netdev_actions *old_actions;
1314 new_actions = dp_netdev_actions_create(put->actions,
1317 old_actions = dp_netdev_flow_get_actions(netdev_flow);
1318 ovsrcu_set(&netdev_flow->actions, new_actions);
1321 get_dpif_flow_stats(netdev_flow, put->stats);
1323 if (put->flags & DPIF_FP_ZERO_STATS) {
1324 clear_stats(netdev_flow);
1327 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
1328 } else if (put->flags & DPIF_FP_CREATE) {
1331 /* Overlapping flow. */
1335 ovs_mutex_unlock(&dp->flow_mutex);
1336 miniflow_destroy(&miniflow);
1342 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
1344 struct dp_netdev *dp = get_dp_netdev(dpif);
1345 struct dp_netdev_flow *netdev_flow;
1349 error = dpif_netdev_flow_from_nlattrs(del->key, del->key_len, &key);
1354 ovs_mutex_lock(&dp->flow_mutex);
1355 netdev_flow = dp_netdev_find_flow(dp, &key);
1358 get_dpif_flow_stats(netdev_flow, del->stats);
1360 dp_netdev_remove_flow(dp, netdev_flow);
1364 ovs_mutex_unlock(&dp->flow_mutex);
1369 struct dpif_netdev_flow_dump {
1370 struct dpif_flow_dump up;
1371 struct cmap_position pos;
1373 struct ovs_mutex mutex;
1376 static struct dpif_netdev_flow_dump *
1377 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
1379 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
1382 static struct dpif_flow_dump *
1383 dpif_netdev_flow_dump_create(const struct dpif *dpif_)
1385 struct dpif_netdev_flow_dump *dump;
1387 dump = xmalloc(sizeof *dump);
1388 dpif_flow_dump_init(&dump->up, dpif_);
1389 memset(&dump->pos, 0, sizeof dump->pos);
1391 ovs_mutex_init(&dump->mutex);
1397 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
1399 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1401 ovs_mutex_destroy(&dump->mutex);
1406 struct dpif_netdev_flow_dump_thread {
1407 struct dpif_flow_dump_thread up;
1408 struct dpif_netdev_flow_dump *dump;
1409 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
1410 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
1413 static struct dpif_netdev_flow_dump_thread *
1414 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
1416 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
1419 static struct dpif_flow_dump_thread *
1420 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
1422 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1423 struct dpif_netdev_flow_dump_thread *thread;
1425 thread = xmalloc(sizeof *thread);
1426 dpif_flow_dump_thread_init(&thread->up, &dump->up);
1427 thread->dump = dump;
1432 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
1434 struct dpif_netdev_flow_dump_thread *thread
1435 = dpif_netdev_flow_dump_thread_cast(thread_);
1441 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
1442 struct dpif_flow *flows, int max_flows)
1444 struct dpif_netdev_flow_dump_thread *thread
1445 = dpif_netdev_flow_dump_thread_cast(thread_);
1446 struct dpif_netdev_flow_dump *dump = thread->dump;
1447 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
1448 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
1449 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
1453 ovs_mutex_lock(&dump->mutex);
1454 if (!dump->status) {
1455 for (n_flows = 0; n_flows < MIN(max_flows, FLOW_DUMP_MAX_BATCH);
1457 struct cmap_node *node;
1459 node = cmap_next_position(&dp->flow_table, &dump->pos);
1464 netdev_flows[n_flows] = CONTAINER_OF(node, struct dp_netdev_flow,
1468 ovs_mutex_unlock(&dump->mutex);
1470 for (i = 0; i < n_flows; i++) {
1471 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
1472 struct odputil_keybuf *keybuf = &thread->keybuf[i];
1473 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
1474 struct dpif_flow *f = &flows[i];
1475 struct dp_netdev_actions *dp_actions;
1476 struct flow_wildcards wc;
1479 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1482 ofpbuf_use_stack(&buf, keybuf, sizeof *keybuf);
1483 odp_flow_key_from_flow(&buf, &netdev_flow->flow, &wc.masks,
1484 netdev_flow->flow.in_port.odp_port, true);
1485 f->key = ofpbuf_data(&buf);
1486 f->key_len = ofpbuf_size(&buf);
1489 ofpbuf_use_stack(&buf, maskbuf, sizeof *maskbuf);
1490 odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
1491 odp_to_u32(wc.masks.in_port.odp_port),
1493 f->mask = ofpbuf_data(&buf);
1494 f->mask_len = ofpbuf_size(&buf);
1497 dp_actions = dp_netdev_flow_get_actions(netdev_flow);
1498 f->actions = dp_actions->actions;
1499 f->actions_len = dp_actions->size;
1502 get_dpif_flow_stats(netdev_flow, &f->stats);
1509 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
1511 struct dp_netdev *dp = get_dp_netdev(dpif);
1512 struct dpif_packet packet, *pp;
1513 struct pkt_metadata *md = &execute->md;
1515 if (ofpbuf_size(execute->packet) < ETH_HEADER_LEN ||
1516 ofpbuf_size(execute->packet) > UINT16_MAX) {
1520 packet.ofpbuf = *execute->packet;
1523 dp_netdev_execute_actions(dp, &pp, 1, false, md,
1524 execute->actions, execute->actions_len);
1526 /* Even though may_steal is set to false, some actions could modify or
1527 * reallocate the ofpbuf memory. We need to pass those changes to the
1529 *execute->packet = packet.ofpbuf;
1535 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
1539 for (i = 0; i < n_ops; i++) {
1540 struct dpif_op *op = ops[i];
1543 case DPIF_OP_FLOW_PUT:
1544 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
1547 case DPIF_OP_FLOW_DEL:
1548 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
1551 case DPIF_OP_EXECUTE:
1552 op->error = dpif_netdev_execute(dpif, &op->u.execute);
1555 case DPIF_OP_FLOW_GET:
1556 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
1563 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
1564 uint32_t queue_id, uint32_t *priority)
1566 *priority = queue_id;
1571 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
1572 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
1574 struct dp_netdev_actions *
1575 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
1577 struct dp_netdev_actions *netdev_actions;
1579 netdev_actions = xmalloc(sizeof *netdev_actions);
1580 netdev_actions->actions = xmemdup(actions, size);
1581 netdev_actions->size = size;
1583 return netdev_actions;
1586 struct dp_netdev_actions *
1587 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
1589 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
1593 dp_netdev_actions_free(struct dp_netdev_actions *actions)
1595 free(actions->actions);
1601 dp_netdev_process_rxq_port(struct dp_netdev *dp,
1602 struct dp_netdev_port *port,
1603 struct netdev_rxq *rxq)
1605 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
1608 error = netdev_rxq_recv(rxq, packets, &cnt);
1610 dp_netdev_port_input(dp, packets, cnt, port->port_no);
1611 } else if (error != EAGAIN && error != EOPNOTSUPP) {
1612 static struct vlog_rate_limit rl
1613 = VLOG_RATE_LIMIT_INIT(1, 5);
1615 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
1616 netdev_get_name(port->netdev),
1617 ovs_strerror(error));
1622 dpif_netdev_run(struct dpif *dpif)
1624 struct dp_netdev_port *port;
1625 struct dp_netdev *dp = get_dp_netdev(dpif);
1627 CMAP_FOR_EACH (port, node, &dp->ports) {
1628 if (!netdev_is_pmd(port->netdev)) {
1631 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1632 dp_netdev_process_rxq_port(dp, port, port->rxq[i]);
1639 dpif_netdev_wait(struct dpif *dpif)
1641 struct dp_netdev_port *port;
1642 struct dp_netdev *dp = get_dp_netdev(dpif);
1644 ovs_mutex_lock(&dp_netdev_mutex);
1645 CMAP_FOR_EACH (port, node, &dp->ports) {
1646 if (!netdev_is_pmd(port->netdev)) {
1649 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1650 netdev_rxq_wait(port->rxq[i]);
1654 ovs_mutex_unlock(&dp_netdev_mutex);
1658 struct dp_netdev_port *port;
1659 struct netdev_rxq *rx;
1663 pmd_load_queues(struct pmd_thread *f,
1664 struct rxq_poll **ppoll_list, int poll_cnt)
1666 struct dp_netdev *dp = f->dp;
1667 struct rxq_poll *poll_list = *ppoll_list;
1668 struct dp_netdev_port *port;
1673 /* Simple scheduler for netdev rx polling. */
1674 for (i = 0; i < poll_cnt; i++) {
1675 port_unref(poll_list[i].port);
1681 CMAP_FOR_EACH (port, node, &f->dp->ports) {
1682 if (netdev_is_pmd(port->netdev)) {
1685 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1686 if ((index % dp->n_pmd_threads) == id) {
1687 poll_list = xrealloc(poll_list, sizeof *poll_list * (poll_cnt + 1));
1690 poll_list[poll_cnt].port = port;
1691 poll_list[poll_cnt].rx = port->rxq[i];
1699 *ppoll_list = poll_list;
1704 pmd_thread_main(void *f_)
1706 struct pmd_thread *f = f_;
1707 struct dp_netdev *dp = f->dp;
1708 unsigned int lc = 0;
1709 struct rxq_poll *poll_list;
1710 unsigned int port_seq = PMD_INITIAL_SEQ;
1717 pmd_thread_setaffinity_cpu(f->id);
1719 poll_cnt = pmd_load_queues(f, &poll_list, poll_cnt);
1724 for (i = 0; i < poll_cnt; i++) {
1725 dp_netdev_process_rxq_port(dp, poll_list[i].port, poll_list[i].rx);
1735 atomic_read_relaxed(&f->change_seq, &seq);
1736 if (seq != port_seq) {
1743 if (!latch_is_set(&f->dp->exit_latch)){
1747 for (i = 0; i < poll_cnt; i++) {
1748 port_unref(poll_list[i].port);
1756 dp_netdev_disable_upcall(struct dp_netdev *dp)
1757 OVS_ACQUIRES(dp->upcall_rwlock)
1759 fat_rwlock_wrlock(&dp->upcall_rwlock);
1763 dpif_netdev_disable_upcall(struct dpif *dpif)
1764 OVS_NO_THREAD_SAFETY_ANALYSIS
1766 struct dp_netdev *dp = get_dp_netdev(dpif);
1767 dp_netdev_disable_upcall(dp);
1771 dp_netdev_enable_upcall(struct dp_netdev *dp)
1772 OVS_RELEASES(dp->upcall_rwlock)
1774 fat_rwlock_unlock(&dp->upcall_rwlock);
1778 dpif_netdev_enable_upcall(struct dpif *dpif)
1779 OVS_NO_THREAD_SAFETY_ANALYSIS
1781 struct dp_netdev *dp = get_dp_netdev(dpif);
1782 dp_netdev_enable_upcall(dp);
1786 dp_netdev_set_pmd_threads(struct dp_netdev *dp, int n)
1790 if (n == dp->n_pmd_threads) {
1794 /* Stop existing threads. */
1795 latch_set(&dp->exit_latch);
1796 dp_netdev_reload_pmd_threads(dp);
1797 for (i = 0; i < dp->n_pmd_threads; i++) {
1798 struct pmd_thread *f = &dp->pmd_threads[i];
1800 xpthread_join(f->thread, NULL);
1802 latch_poll(&dp->exit_latch);
1803 free(dp->pmd_threads);
1805 /* Start new threads. */
1806 dp->pmd_threads = xmalloc(n * sizeof *dp->pmd_threads);
1807 dp->n_pmd_threads = n;
1809 for (i = 0; i < n; i++) {
1810 struct pmd_thread *f = &dp->pmd_threads[i];
1814 atomic_init(&f->change_seq, PMD_INITIAL_SEQ);
1816 /* Each thread will distribute all devices rx-queues among
1818 f->thread = ovs_thread_create("pmd", pmd_thread_main, f);
1824 dp_netdev_flow_stats_new_cb(void)
1826 struct dp_netdev_flow_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1827 ovs_mutex_init(&bucket->mutex);
1832 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow,
1836 long long int now = time_msec();
1837 struct dp_netdev_flow_stats *bucket;
1839 bucket = ovsthread_stats_bucket_get(&netdev_flow->stats,
1840 dp_netdev_flow_stats_new_cb);
1842 ovs_mutex_lock(&bucket->mutex);
1843 bucket->used = MAX(now, bucket->used);
1844 bucket->packet_count += cnt;
1845 bucket->byte_count += size;
1846 bucket->tcp_flags |= tcp_flags;
1847 ovs_mutex_unlock(&bucket->mutex);
1851 dp_netdev_stats_new_cb(void)
1853 struct dp_netdev_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1854 ovs_mutex_init(&bucket->mutex);
1859 dp_netdev_count_packet(struct dp_netdev *dp, enum dp_stat_type type, int cnt)
1861 struct dp_netdev_stats *bucket;
1863 bucket = ovsthread_stats_bucket_get(&dp->stats, dp_netdev_stats_new_cb);
1864 ovs_mutex_lock(&bucket->mutex);
1865 bucket->n[type] += cnt;
1866 ovs_mutex_unlock(&bucket->mutex);
1870 dp_netdev_upcall(struct dp_netdev *dp, struct dpif_packet *packet_,
1871 struct flow *flow, struct flow_wildcards *wc,
1872 enum dpif_upcall_type type, const struct nlattr *userdata,
1873 struct ofpbuf *actions, struct ofpbuf *put_actions)
1875 struct ofpbuf *packet = &packet_->ofpbuf;
1877 if (type == DPIF_UC_MISS) {
1878 dp_netdev_count_packet(dp, DP_STAT_MISS, 1);
1881 if (OVS_UNLIKELY(!dp->upcall_cb)) {
1885 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
1886 struct ds ds = DS_EMPTY_INITIALIZER;
1890 ofpbuf_init(&key, 0);
1891 odp_flow_key_from_flow(&key, flow, &wc->masks, flow->in_port.odp_port,
1894 packet_str = ofp_packet_to_string(ofpbuf_data(packet),
1895 ofpbuf_size(packet));
1897 odp_flow_key_format(ofpbuf_data(&key), ofpbuf_size(&key), &ds);
1899 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
1900 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
1902 ofpbuf_uninit(&key);
1907 return dp->upcall_cb(packet, flow, type, userdata, actions, wc,
1908 put_actions, dp->upcall_aux);
1911 struct packet_batch {
1912 unsigned int packet_count;
1913 unsigned int byte_count;
1916 struct dp_netdev_flow *flow;
1918 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
1919 struct pkt_metadata md;
1923 packet_batch_update(struct packet_batch *batch,
1924 struct dpif_packet *packet, const struct miniflow *mf)
1926 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
1927 batch->packets[batch->packet_count++] = packet;
1928 batch->byte_count += ofpbuf_size(&packet->ofpbuf);
1932 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow,
1933 struct pkt_metadata *md)
1938 batch->packet_count = 0;
1939 batch->byte_count = 0;
1940 batch->tcp_flags = 0;
1944 packet_batch_execute(struct packet_batch *batch, struct dp_netdev *dp)
1946 struct dp_netdev_actions *actions;
1947 struct dp_netdev_flow *flow = batch->flow;
1949 dp_netdev_flow_used(batch->flow, batch->packet_count, batch->byte_count,
1952 actions = dp_netdev_flow_get_actions(flow);
1954 dp_netdev_execute_actions(dp, batch->packets,
1955 batch->packet_count, true, &batch->md,
1956 actions->actions, actions->size);
1958 dp_netdev_count_packet(dp, DP_STAT_HIT, batch->packet_count);
1962 dp_netdev_input(struct dp_netdev *dp, struct dpif_packet **packets, int cnt,
1963 struct pkt_metadata *md)
1965 struct packet_batch batches[NETDEV_MAX_RX_BATCH];
1966 struct netdev_flow_key keys[NETDEV_MAX_RX_BATCH];
1967 const struct miniflow *mfs[NETDEV_MAX_RX_BATCH]; /* NULL at bad packets. */
1968 struct cls_rule *rules[NETDEV_MAX_RX_BATCH];
1969 size_t n_batches, i;
1972 for (i = 0; i < cnt; i++) {
1973 if (OVS_UNLIKELY(ofpbuf_size(&packets[i]->ofpbuf) < ETH_HEADER_LEN)) {
1974 dpif_packet_delete(packets[i]);
1979 miniflow_initialize(&keys[i].flow, keys[i].buf);
1980 miniflow_extract(&packets[i]->ofpbuf, md, &keys[i].flow);
1981 mfs[i] = &keys[i].flow;
1984 any_miss = !classifier_lookup_miniflow_batch(&dp->cls, mfs, rules, cnt);
1985 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
1986 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
1987 struct ofpbuf actions, put_actions;
1990 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
1991 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
1993 for (i = 0; i < cnt; i++) {
1994 const struct dp_netdev_flow *netdev_flow;
1995 struct ofpbuf *add_actions;
1998 if (OVS_LIKELY(rules[i] || !mfs[i])) {
2002 /* It's possible that an earlier slow path execution installed
2003 * the rule this flow needs. In this case, it's a lot cheaper
2004 * to catch it here than execute a miss. */
2005 netdev_flow = dp_netdev_lookup_flow(dp, mfs[i]);
2007 rules[i] = CONST_CAST(struct cls_rule *, &netdev_flow->cr);
2011 miniflow_expand(mfs[i], &match.flow);
2013 ofpbuf_clear(&actions);
2014 ofpbuf_clear(&put_actions);
2016 error = dp_netdev_upcall(dp, packets[i], &match.flow, &match.wc,
2017 DPIF_UC_MISS, NULL, &actions,
2019 if (OVS_UNLIKELY(error && error != ENOSPC)) {
2023 /* We can't allow the packet batching in the next loop to execute
2024 * the actions. Otherwise, if there are any slow path actions,
2025 * we'll send the packet up twice. */
2026 dp_netdev_execute_actions(dp, &packets[i], 1, false, md,
2027 ofpbuf_data(&actions),
2028 ofpbuf_size(&actions));
2030 add_actions = ofpbuf_size(&put_actions)
2034 ovs_mutex_lock(&dp->flow_mutex);
2035 /* XXX: There's a brief race where this flow could have already
2036 * been installed since we last did the flow lookup. This could be
2037 * solved by moving the mutex lock outside the loop, but that's an
2038 * awful long time to be locking everyone out of making flow
2039 * installs. If we move to a per-core classifier, it would be
2041 if (OVS_LIKELY(error != ENOSPC)
2042 && !dp_netdev_lookup_flow(dp, mfs[i])) {
2043 dp_netdev_flow_add(dp, &match, ofpbuf_data(add_actions),
2044 ofpbuf_size(add_actions));
2046 ovs_mutex_unlock(&dp->flow_mutex);
2049 ofpbuf_uninit(&actions);
2050 ofpbuf_uninit(&put_actions);
2051 fat_rwlock_unlock(&dp->upcall_rwlock);
2055 for (i = 0; i < cnt; i++) {
2056 struct dp_netdev_flow *flow;
2057 struct packet_batch *batch;
2060 if (OVS_UNLIKELY(!rules[i] || !mfs[i])) {
2064 /* XXX: This O(n^2) algortihm makes sense if we're operating under the
2065 * assumption that the number of distinct flows (and therefore the
2066 * number of distinct batches) is quite small. If this turns out not
2067 * to be the case, it may make sense to pre sort based on the
2068 * netdev_flow pointer. That done we can get the appropriate batching
2069 * in O(n * log(n)) instead. */
2071 flow = dp_netdev_flow_cast(rules[i]);
2072 for (j = 0; j < n_batches; j++) {
2073 if (batches[j].flow == flow) {
2074 batch = &batches[j];
2080 batch = &batches[n_batches++];
2081 packet_batch_init(batch, flow, md);
2083 packet_batch_update(batch, packets[i], mfs[i]);
2086 for (i = 0; i < n_batches; i++) {
2087 packet_batch_execute(&batches[i], dp);
2092 dp_netdev_port_input(struct dp_netdev *dp, struct dpif_packet **packets,
2093 int cnt, odp_port_t port_no)
2095 uint32_t *recirc_depth = recirc_depth_get();
2096 struct pkt_metadata md = PKT_METADATA_INITIALIZER(port_no);
2099 dp_netdev_input(dp, packets, cnt, &md);
2102 struct dp_netdev_execute_aux {
2103 struct dp_netdev *dp;
2107 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
2110 struct dp_netdev *dp = get_dp_netdev(dpif);
2111 dp->upcall_aux = aux;
2116 dp_execute_cb(void *aux_, struct dpif_packet **packets, int cnt,
2117 struct pkt_metadata *md,
2118 const struct nlattr *a, bool may_steal)
2119 OVS_NO_THREAD_SAFETY_ANALYSIS
2121 struct dp_netdev_execute_aux *aux = aux_;
2122 uint32_t *depth = recirc_depth_get();
2123 struct dp_netdev *dp = aux->dp;
2124 int type = nl_attr_type(a);
2125 struct dp_netdev_port *p;
2128 switch ((enum ovs_action_attr)type) {
2129 case OVS_ACTION_ATTR_OUTPUT:
2130 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
2131 if (OVS_LIKELY(p)) {
2132 netdev_send(p->netdev, packets, cnt, may_steal);
2133 } else if (may_steal) {
2134 for (i = 0; i < cnt; i++) {
2135 dpif_packet_delete(packets[i]);
2140 case OVS_ACTION_ATTR_USERSPACE:
2141 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
2142 const struct nlattr *userdata;
2143 struct ofpbuf actions;
2146 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
2147 ofpbuf_init(&actions, 0);
2149 for (i = 0; i < cnt; i++) {
2152 ofpbuf_clear(&actions);
2154 flow_extract(&packets[i]->ofpbuf, md, &flow);
2155 error = dp_netdev_upcall(dp, packets[i], &flow, NULL,
2156 DPIF_UC_ACTION, userdata, &actions,
2158 if (!error || error == ENOSPC) {
2159 dp_netdev_execute_actions(dp, &packets[i], 1, false, md,
2160 ofpbuf_data(&actions),
2161 ofpbuf_size(&actions));
2165 dpif_packet_delete(packets[i]);
2168 ofpbuf_uninit(&actions);
2169 fat_rwlock_unlock(&dp->upcall_rwlock);
2174 case OVS_ACTION_ATTR_HASH: {
2175 const struct ovs_action_hash *hash_act;
2176 struct netdev_flow_key key;
2179 hash_act = nl_attr_get(a);
2181 miniflow_initialize(&key.flow, key.buf);
2183 for (i = 0; i < cnt; i++) {
2185 /* XXX: this is slow. Use RSS hash in the future */
2186 miniflow_extract(&packets[i]->ofpbuf, md, &key.flow);
2188 if (hash_act->hash_alg == OVS_HASH_ALG_L4) {
2189 /* Hash need not be symmetric, nor does it need to include
2191 hash = miniflow_hash_5tuple(&key.flow, hash_act->hash_basis);
2193 VLOG_WARN("Unknown hash algorithm specified "
2194 "for the hash action.");
2199 hash = 1; /* 0 is not valid */
2205 packets[i]->dp_hash = hash;
2210 case OVS_ACTION_ATTR_RECIRC:
2211 if (*depth < MAX_RECIRC_DEPTH) {
2214 for (i = 0; i < cnt; i++) {
2215 struct dpif_packet *recirc_pkt;
2216 struct pkt_metadata recirc_md = *md;
2218 recirc_pkt = (may_steal) ? packets[i]
2219 : dpif_packet_clone(packets[i]);
2221 recirc_md.recirc_id = nl_attr_get_u32(a);
2223 /* Hash is private to each packet */
2224 recirc_md.dp_hash = dpif_packet_get_dp_hash(packets[i]);
2226 dp_netdev_input(dp, &recirc_pkt, 1, &recirc_md);
2232 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
2234 for (i = 0; i < cnt; i++) {
2235 dpif_packet_delete(packets[i]);
2241 case OVS_ACTION_ATTR_PUSH_VLAN:
2242 case OVS_ACTION_ATTR_POP_VLAN:
2243 case OVS_ACTION_ATTR_PUSH_MPLS:
2244 case OVS_ACTION_ATTR_POP_MPLS:
2245 case OVS_ACTION_ATTR_SET:
2246 case OVS_ACTION_ATTR_SAMPLE:
2247 case OVS_ACTION_ATTR_UNSPEC:
2248 case __OVS_ACTION_ATTR_MAX:
2254 dp_netdev_execute_actions(struct dp_netdev *dp,
2255 struct dpif_packet **packets, int cnt,
2256 bool may_steal, struct pkt_metadata *md,
2257 const struct nlattr *actions, size_t actions_len)
2259 struct dp_netdev_execute_aux aux = {dp};
2261 odp_execute_actions(&aux, packets, cnt, may_steal, md, actions,
2262 actions_len, dp_execute_cb);
2265 const struct dpif_class dpif_netdev_class = {
2267 dpif_netdev_enumerate,
2268 dpif_netdev_port_open_type,
2271 dpif_netdev_destroy,
2274 dpif_netdev_get_stats,
2275 dpif_netdev_port_add,
2276 dpif_netdev_port_del,
2277 dpif_netdev_port_query_by_number,
2278 dpif_netdev_port_query_by_name,
2279 NULL, /* port_get_pid */
2280 dpif_netdev_port_dump_start,
2281 dpif_netdev_port_dump_next,
2282 dpif_netdev_port_dump_done,
2283 dpif_netdev_port_poll,
2284 dpif_netdev_port_poll_wait,
2285 dpif_netdev_flow_flush,
2286 dpif_netdev_flow_dump_create,
2287 dpif_netdev_flow_dump_destroy,
2288 dpif_netdev_flow_dump_thread_create,
2289 dpif_netdev_flow_dump_thread_destroy,
2290 dpif_netdev_flow_dump_next,
2291 dpif_netdev_operate,
2292 NULL, /* recv_set */
2293 NULL, /* handlers_set */
2294 dpif_netdev_queue_to_priority,
2296 NULL, /* recv_wait */
2297 NULL, /* recv_purge */
2298 dpif_netdev_register_upcall_cb,
2299 dpif_netdev_enable_upcall,
2300 dpif_netdev_disable_upcall,
2304 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
2305 const char *argv[], void *aux OVS_UNUSED)
2307 struct dp_netdev_port *old_port;
2308 struct dp_netdev_port *new_port;
2309 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], &old_port)) {
2324 unixctl_command_reply_error(conn, "unknown port");
2328 port_no = u32_to_odp(atoi(argv[3]));
2329 if (!port_no || port_no == ODPP_NONE) {
2330 unixctl_command_reply_error(conn, "bad port number");
2333 if (dp_netdev_lookup_port(dp, port_no)) {
2334 unixctl_command_reply_error(conn, "port number already in use");
2338 /* Remove old port. */
2339 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
2340 ovsrcu_postpone(free, old_port);
2342 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
2343 new_port = xmemdup(old_port, sizeof *old_port);
2344 new_port->port_no = port_no;
2345 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
2347 seq_change(dp->port_seq);
2348 unixctl_command_reply(conn, NULL);
2351 ovs_mutex_unlock(&dp->port_mutex);
2352 dp_netdev_unref(dp);
2356 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
2357 const char *argv[], void *aux OVS_UNUSED)
2359 struct dp_netdev_port *port;
2360 struct dp_netdev *dp;
2362 ovs_mutex_lock(&dp_netdev_mutex);
2363 dp = shash_find_data(&dp_netdevs, argv[1]);
2364 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2365 ovs_mutex_unlock(&dp_netdev_mutex);
2366 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2369 ovs_refcount_ref(&dp->ref_cnt);
2370 ovs_mutex_unlock(&dp_netdev_mutex);
2372 ovs_mutex_lock(&dp->port_mutex);
2373 if (get_port_by_name(dp, argv[2], &port)) {
2374 unixctl_command_reply_error(conn, "unknown port");
2375 } else if (port->port_no == ODPP_LOCAL) {
2376 unixctl_command_reply_error(conn, "can't delete local port");
2378 do_del_port(dp, port);
2379 unixctl_command_reply(conn, NULL);
2381 ovs_mutex_unlock(&dp->port_mutex);
2383 dp_netdev_unref(dp);
2387 dpif_dummy_register__(const char *type)
2389 struct dpif_class *class;
2391 class = xmalloc(sizeof *class);
2392 *class = dpif_netdev_class;
2393 class->type = xstrdup(type);
2394 dp_register_provider(class);
2398 dpif_dummy_register(bool override)
2405 dp_enumerate_types(&types);
2406 SSET_FOR_EACH (type, &types) {
2407 if (!dp_unregister_provider(type)) {
2408 dpif_dummy_register__(type);
2411 sset_destroy(&types);
2414 dpif_dummy_register__("dummy");
2416 unixctl_command_register("dpif-dummy/change-port-number",
2417 "DP PORT NEW-NUMBER",
2418 3, 3, dpif_dummy_change_port_number, NULL);
2419 unixctl_command_register("dpif-dummy/delete-port", "DP PORT",
2420 2, 2, dpif_dummy_delete_port, NULL);