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 during an RCU
223 * grace period. Code that needs to hold onto a flow for a while
224 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
226 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
227 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
230 * Some members, marked 'const', are immutable. Accessing other members
231 * requires synchronization, as noted in more detail below.
233 struct dp_netdev_flow {
234 /* Packet classification. */
235 const struct cls_rule cr; /* In owning dp_netdev's 'cls'. */
237 /* Hash table index by unmasked flow. */
238 const struct cmap_node node; /* In owning dp_netdev's 'flow_table'. */
239 const struct flow flow; /* The flow that created this entry. */
241 /* Number of references.
242 * The classifier owns one reference.
243 * Any thread trying to keep a rule from being freed should hold its own
245 struct ovs_refcount ref_cnt;
249 * Reading or writing these members requires 'mutex'. */
250 struct ovsthread_stats stats; /* Contains "struct dp_netdev_flow_stats". */
253 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
256 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
258 /* Contained by struct dp_netdev_flow's 'stats' member. */
259 struct dp_netdev_flow_stats {
260 struct ovs_mutex mutex; /* Guards all the other members. */
262 long long int used OVS_GUARDED; /* Last used time, in monotonic msecs. */
263 long long int packet_count OVS_GUARDED; /* Number of packets matched. */
264 long long int byte_count OVS_GUARDED; /* Number of bytes matched. */
265 uint16_t tcp_flags OVS_GUARDED; /* Bitwise-OR of seen tcp_flags values. */
268 /* A set of datapath actions within a "struct dp_netdev_flow".
274 * A struct dp_netdev_actions 'actions' is protected with RCU. */
275 struct dp_netdev_actions {
276 /* These members are immutable: they do not change during the struct's
278 struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */
279 unsigned int size; /* Size of 'actions', in bytes. */
282 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
284 struct dp_netdev_actions *dp_netdev_flow_get_actions(
285 const struct dp_netdev_flow *);
286 static void dp_netdev_actions_free(struct dp_netdev_actions *);
288 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
289 * the performance overhead of interrupt processing. Therefore netdev can
290 * not implement rx-wait for these devices. dpif-netdev needs to poll
291 * these device to check for recv buffer. pmd-thread does polling for
292 * devices assigned to itself thread.
294 * DPDK used PMD for accessing NIC.
296 * A thread that receives packets from PMD ports, looks them up in the flow
297 * table, and executes the actions it finds.
300 struct dp_netdev *dp;
303 atomic_uint change_seq;
306 /* Interface to netdev-based datapath. */
309 struct dp_netdev *dp;
310 uint64_t last_port_seq;
313 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
314 struct dp_netdev_port **portp);
315 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
316 struct dp_netdev_port **portp);
317 static void dp_netdev_free(struct dp_netdev *)
318 OVS_REQUIRES(dp_netdev_mutex);
319 static void dp_netdev_flow_flush(struct dp_netdev *);
320 static int do_add_port(struct dp_netdev *dp, const char *devname,
321 const char *type, odp_port_t port_no)
322 OVS_REQUIRES(dp->port_mutex);
323 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
324 OVS_REQUIRES(dp->port_mutex);
325 static int dpif_netdev_open(const struct dpif_class *, const char *name,
326 bool create, struct dpif **);
327 static int dp_netdev_queue_userspace_packet(struct dp_netdev_queue *,
328 struct ofpbuf *, int type,
329 const struct miniflow *,
330 const struct nlattr *);
331 static void dp_netdev_execute_userspace_queue(struct dp_netdev_queue *,
333 static void dp_netdev_execute_actions(struct dp_netdev *dp,
334 struct dpif_packet **, int c,
335 bool may_steal, struct pkt_metadata *,
336 const struct nlattr *actions,
338 static void dp_netdev_port_input(struct dp_netdev *dp,
339 struct dpif_packet **packets, int cnt,
342 static void dp_netdev_set_pmd_threads(struct dp_netdev *, int n);
343 static void dp_netdev_disable_upcall(struct dp_netdev *);
345 static struct dpif_netdev *
346 dpif_netdev_cast(const struct dpif *dpif)
348 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
349 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
352 static struct dp_netdev *
353 get_dp_netdev(const struct dpif *dpif)
355 return dpif_netdev_cast(dpif)->dp;
359 dpif_netdev_enumerate(struct sset *all_dps,
360 const struct dpif_class *dpif_class)
362 struct shash_node *node;
364 ovs_mutex_lock(&dp_netdev_mutex);
365 SHASH_FOR_EACH(node, &dp_netdevs) {
366 struct dp_netdev *dp = node->data;
367 if (dpif_class != dp->class) {
368 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
369 * If the class doesn't match, skip this dpif. */
372 sset_add(all_dps, node->name);
374 ovs_mutex_unlock(&dp_netdev_mutex);
380 dpif_netdev_class_is_dummy(const struct dpif_class *class)
382 return class != &dpif_netdev_class;
386 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
388 return strcmp(type, "internal") ? type
389 : dpif_netdev_class_is_dummy(class) ? "dummy"
394 create_dpif_netdev(struct dp_netdev *dp)
396 uint16_t netflow_id = hash_string(dp->name, 0);
397 struct dpif_netdev *dpif;
399 ovs_refcount_ref(&dp->ref_cnt);
401 dpif = xmalloc(sizeof *dpif);
402 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
404 dpif->last_port_seq = seq_read(dp->port_seq);
409 /* Choose an unused, non-zero port number and return it on success.
410 * Return ODPP_NONE on failure. */
412 choose_port(struct dp_netdev *dp, const char *name)
413 OVS_REQUIRES(dp->port_mutex)
417 if (dp->class != &dpif_netdev_class) {
421 /* If the port name begins with "br", start the number search at
422 * 100 to make writing tests easier. */
423 if (!strncmp(name, "br", 2)) {
427 /* If the port name contains a number, try to assign that port number.
428 * This can make writing unit tests easier because port numbers are
430 for (p = name; *p != '\0'; p++) {
431 if (isdigit((unsigned char) *p)) {
432 port_no = start_no + strtol(p, NULL, 10);
433 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
434 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
435 return u32_to_odp(port_no);
442 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
443 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
444 return u32_to_odp(port_no);
452 create_dp_netdev(const char *name, const struct dpif_class *class,
453 struct dp_netdev **dpp)
454 OVS_REQUIRES(dp_netdev_mutex)
456 struct dp_netdev *dp;
459 dp = xzalloc(sizeof *dp);
460 shash_add(&dp_netdevs, name, dp);
462 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
463 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
464 ovs_refcount_init(&dp->ref_cnt);
465 atomic_flag_clear(&dp->destroyed);
467 ovs_mutex_init(&dp->flow_mutex);
468 classifier_init(&dp->cls, NULL);
469 cmap_init(&dp->flow_table);
471 ovsthread_stats_init(&dp->stats);
473 ovs_mutex_init(&dp->port_mutex);
474 cmap_init(&dp->ports);
475 dp->port_seq = seq_create();
476 latch_init(&dp->exit_latch);
477 fat_rwlock_init(&dp->upcall_rwlock);
479 /* Disable upcalls by default. */
480 dp_netdev_disable_upcall(dp);
481 dp->upcall_cb = NULL;
483 ovs_mutex_lock(&dp->port_mutex);
484 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
485 ovs_mutex_unlock(&dp->port_mutex);
496 dpif_netdev_open(const struct dpif_class *class, const char *name,
497 bool create, struct dpif **dpifp)
499 struct dp_netdev *dp;
502 ovs_mutex_lock(&dp_netdev_mutex);
503 dp = shash_find_data(&dp_netdevs, name);
505 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
507 error = (dp->class != class ? EINVAL
512 *dpifp = create_dpif_netdev(dp);
515 ovs_mutex_unlock(&dp_netdev_mutex);
520 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
521 * through the 'dp_netdevs' shash while freeing 'dp'. */
523 dp_netdev_free(struct dp_netdev *dp)
524 OVS_REQUIRES(dp_netdev_mutex)
526 struct dp_netdev_port *port;
527 struct dp_netdev_stats *bucket;
530 shash_find_and_delete(&dp_netdevs, dp->name);
532 dp_netdev_set_pmd_threads(dp, 0);
533 free(dp->pmd_threads);
535 dp_netdev_flow_flush(dp);
536 ovs_mutex_lock(&dp->port_mutex);
537 CMAP_FOR_EACH (port, node, &dp->ports) {
538 do_del_port(dp, port);
540 ovs_mutex_unlock(&dp->port_mutex);
542 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
543 ovs_mutex_destroy(&bucket->mutex);
544 free_cacheline(bucket);
546 ovsthread_stats_destroy(&dp->stats);
548 classifier_destroy(&dp->cls);
549 cmap_destroy(&dp->flow_table);
550 ovs_mutex_destroy(&dp->flow_mutex);
551 seq_destroy(dp->port_seq);
552 cmap_destroy(&dp->ports);
553 fat_rwlock_destroy(&dp->upcall_rwlock);
554 latch_destroy(&dp->exit_latch);
555 free(CONST_CAST(char *, dp->name));
560 dp_netdev_unref(struct dp_netdev *dp)
563 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
564 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
565 ovs_mutex_lock(&dp_netdev_mutex);
566 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
569 ovs_mutex_unlock(&dp_netdev_mutex);
574 dpif_netdev_close(struct dpif *dpif)
576 struct dp_netdev *dp = get_dp_netdev(dpif);
583 dpif_netdev_destroy(struct dpif *dpif)
585 struct dp_netdev *dp = get_dp_netdev(dpif);
587 if (!atomic_flag_test_and_set(&dp->destroyed)) {
588 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
589 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
598 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
600 struct dp_netdev *dp = get_dp_netdev(dpif);
601 struct dp_netdev_stats *bucket;
604 stats->n_flows = cmap_count(&dp->flow_table);
606 stats->n_hit = stats->n_missed = stats->n_lost = 0;
607 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
608 ovs_mutex_lock(&bucket->mutex);
609 stats->n_hit += bucket->n[DP_STAT_HIT];
610 stats->n_missed += bucket->n[DP_STAT_MISS];
611 stats->n_lost += bucket->n[DP_STAT_LOST];
612 ovs_mutex_unlock(&bucket->mutex);
614 stats->n_masks = UINT32_MAX;
615 stats->n_mask_hit = UINT64_MAX;
621 dp_netdev_reload_pmd_threads(struct dp_netdev *dp)
625 for (i = 0; i < dp->n_pmd_threads; i++) {
626 struct pmd_thread *f = &dp->pmd_threads[i];
629 atomic_add(&f->change_seq, 1, &id);
634 hash_port_no(odp_port_t port_no)
636 return hash_int(odp_to_u32(port_no), 0);
640 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
642 OVS_REQUIRES(dp->port_mutex)
644 struct netdev_saved_flags *sf;
645 struct dp_netdev_port *port;
646 struct netdev *netdev;
647 enum netdev_flags flags;
648 const char *open_type;
652 /* XXX reject devices already in some dp_netdev. */
654 /* Open and validate network device. */
655 open_type = dpif_netdev_port_open_type(dp->class, type);
656 error = netdev_open(devname, open_type, &netdev);
660 /* XXX reject non-Ethernet devices */
662 netdev_get_flags(netdev, &flags);
663 if (flags & NETDEV_LOOPBACK) {
664 VLOG_ERR("%s: cannot add a loopback device", devname);
665 netdev_close(netdev);
669 port = xzalloc(sizeof *port);
670 port->port_no = port_no;
671 port->netdev = netdev;
672 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
673 port->type = xstrdup(type);
674 for (i = 0; i < netdev_n_rxq(netdev); i++) {
675 error = netdev_rxq_open(netdev, &port->rxq[i], i);
677 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
678 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
679 devname, ovs_strerror(errno));
680 netdev_close(netdev);
685 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
687 for (i = 0; i < netdev_n_rxq(netdev); i++) {
688 netdev_rxq_close(port->rxq[i]);
690 netdev_close(netdev);
697 if (netdev_is_pmd(netdev)) {
699 dp_netdev_set_pmd_threads(dp, NR_PMD_THREADS);
700 dp_netdev_reload_pmd_threads(dp);
702 ovs_refcount_init(&port->ref_cnt);
704 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
705 seq_change(dp->port_seq);
711 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
712 odp_port_t *port_nop)
714 struct dp_netdev *dp = get_dp_netdev(dpif);
715 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
716 const char *dpif_port;
720 ovs_mutex_lock(&dp->port_mutex);
721 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
722 if (*port_nop != ODPP_NONE) {
724 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
726 port_no = choose_port(dp, dpif_port);
727 error = port_no == ODPP_NONE ? EFBIG : 0;
731 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
733 ovs_mutex_unlock(&dp->port_mutex);
739 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
741 struct dp_netdev *dp = get_dp_netdev(dpif);
744 ovs_mutex_lock(&dp->port_mutex);
745 if (port_no == ODPP_LOCAL) {
748 struct dp_netdev_port *port;
750 error = get_port_by_number(dp, port_no, &port);
752 do_del_port(dp, port);
755 ovs_mutex_unlock(&dp->port_mutex);
761 is_valid_port_number(odp_port_t port_no)
763 return port_no != ODPP_NONE;
766 static struct dp_netdev_port *
767 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
769 struct dp_netdev_port *port;
771 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
772 if (port->port_no == port_no) {
780 get_port_by_number(struct dp_netdev *dp,
781 odp_port_t port_no, struct dp_netdev_port **portp)
783 if (!is_valid_port_number(port_no)) {
787 *portp = dp_netdev_lookup_port(dp, port_no);
788 return *portp ? 0 : ENOENT;
793 port_ref(struct dp_netdev_port *port)
796 ovs_refcount_ref(&port->ref_cnt);
801 port_destroy__(struct dp_netdev_port *port)
803 int n_rxq = netdev_n_rxq(port->netdev);
806 netdev_close(port->netdev);
807 netdev_restore_flags(port->sf);
809 for (i = 0; i < n_rxq; i++) {
810 netdev_rxq_close(port->rxq[i]);
818 port_unref(struct dp_netdev_port *port)
820 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
821 ovsrcu_postpone(port_destroy__, port);
826 get_port_by_name(struct dp_netdev *dp,
827 const char *devname, struct dp_netdev_port **portp)
828 OVS_REQUIRES(dp->port_mutex)
830 struct dp_netdev_port *port;
832 CMAP_FOR_EACH (port, node, &dp->ports) {
833 if (!strcmp(netdev_get_name(port->netdev), devname)) {
842 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
843 OVS_REQUIRES(dp->port_mutex)
845 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
846 seq_change(dp->port_seq);
847 if (netdev_is_pmd(port->netdev)) {
848 dp_netdev_reload_pmd_threads(dp);
855 answer_port_query(const struct dp_netdev_port *port,
856 struct dpif_port *dpif_port)
858 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
859 dpif_port->type = xstrdup(port->type);
860 dpif_port->port_no = port->port_no;
864 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
865 struct dpif_port *dpif_port)
867 struct dp_netdev *dp = get_dp_netdev(dpif);
868 struct dp_netdev_port *port;
871 error = get_port_by_number(dp, port_no, &port);
872 if (!error && dpif_port) {
873 answer_port_query(port, dpif_port);
880 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
881 struct dpif_port *dpif_port)
883 struct dp_netdev *dp = get_dp_netdev(dpif);
884 struct dp_netdev_port *port;
887 ovs_mutex_lock(&dp->port_mutex);
888 error = get_port_by_name(dp, devname, &port);
889 if (!error && dpif_port) {
890 answer_port_query(port, dpif_port);
892 ovs_mutex_unlock(&dp->port_mutex);
898 dp_netdev_flow_free(struct dp_netdev_flow *flow)
900 struct dp_netdev_flow_stats *bucket;
903 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
904 ovs_mutex_destroy(&bucket->mutex);
905 free_cacheline(bucket);
907 ovsthread_stats_destroy(&flow->stats);
909 cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
910 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
914 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
916 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
917 ovsrcu_postpone(dp_netdev_flow_free, flow);
922 dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
923 OVS_REQUIRES(dp->flow_mutex)
925 struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr);
926 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
928 classifier_remove(&dp->cls, cr);
929 cmap_remove(&dp->flow_table, node, flow_hash(&flow->flow, 0));
931 dp_netdev_flow_unref(flow);
935 dp_netdev_flow_flush(struct dp_netdev *dp)
937 struct dp_netdev_flow *netdev_flow;
939 ovs_mutex_lock(&dp->flow_mutex);
940 CMAP_FOR_EACH (netdev_flow, node, &dp->flow_table) {
941 dp_netdev_remove_flow(dp, netdev_flow);
943 ovs_mutex_unlock(&dp->flow_mutex);
947 dpif_netdev_flow_flush(struct dpif *dpif)
949 struct dp_netdev *dp = get_dp_netdev(dpif);
951 dp_netdev_flow_flush(dp);
955 struct dp_netdev_port_state {
956 struct cmap_position position;
961 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
963 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
968 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
969 struct dpif_port *dpif_port)
971 struct dp_netdev_port_state *state = state_;
972 struct dp_netdev *dp = get_dp_netdev(dpif);
973 struct cmap_node *node;
976 node = cmap_next_position(&dp->ports, &state->position);
978 struct dp_netdev_port *port;
980 port = CONTAINER_OF(node, struct dp_netdev_port, node);
983 state->name = xstrdup(netdev_get_name(port->netdev));
984 dpif_port->name = state->name;
985 dpif_port->type = port->type;
986 dpif_port->port_no = port->port_no;
997 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
999 struct dp_netdev_port_state *state = state_;
1006 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1008 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1009 uint64_t new_port_seq;
1012 new_port_seq = seq_read(dpif->dp->port_seq);
1013 if (dpif->last_port_seq != new_port_seq) {
1014 dpif->last_port_seq = new_port_seq;
1024 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1026 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1028 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1031 static struct dp_netdev_flow *
1032 dp_netdev_flow_cast(const struct cls_rule *cr)
1034 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1037 static struct dp_netdev_flow *
1038 dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct miniflow *key)
1040 struct dp_netdev_flow *netdev_flow;
1041 struct cls_rule *rule;
1043 classifier_lookup_miniflow_batch(&dp->cls, &key, &rule, 1);
1044 netdev_flow = dp_netdev_flow_cast(rule);
1049 static struct dp_netdev_flow *
1050 dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
1052 struct dp_netdev_flow *netdev_flow;
1054 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
1056 if (flow_equal(&netdev_flow->flow, flow)) {
1065 get_dpif_flow_stats(struct dp_netdev_flow *netdev_flow,
1066 struct dpif_flow_stats *stats)
1068 struct dp_netdev_flow_stats *bucket;
1071 memset(stats, 0, sizeof *stats);
1072 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1073 ovs_mutex_lock(&bucket->mutex);
1074 stats->n_packets += bucket->packet_count;
1075 stats->n_bytes += bucket->byte_count;
1076 stats->used = MAX(stats->used, bucket->used);
1077 stats->tcp_flags |= bucket->tcp_flags;
1078 ovs_mutex_unlock(&bucket->mutex);
1083 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1084 const struct nlattr *mask_key,
1085 uint32_t mask_key_len, const struct flow *flow,
1089 enum odp_key_fitness fitness;
1091 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1093 /* This should not happen: it indicates that
1094 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1095 * disagree on the acceptable form of a mask. Log the problem
1096 * as an error, with enough details to enable debugging. */
1097 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1099 if (!VLOG_DROP_ERR(&rl)) {
1103 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1105 VLOG_ERR("internal error parsing flow mask %s (%s)",
1106 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1113 enum mf_field_id id;
1114 /* No mask key, unwildcard everything except fields whose
1115 * prerequisities are not met. */
1116 memset(mask, 0x0, sizeof *mask);
1118 for (id = 0; id < MFF_N_IDS; ++id) {
1119 /* Skip registers and metadata. */
1120 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1121 && id != MFF_METADATA) {
1122 const struct mf_field *mf = mf_from_id(id);
1123 if (mf_are_prereqs_ok(mf, flow)) {
1124 mf_mask_field(mf, mask);
1130 /* Force unwildcard the in_port.
1132 * We need to do this even in the case where we unwildcard "everything"
1133 * above because "everything" only includes the 16-bit OpenFlow port number
1134 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1135 * port number mask->in_port.odp_port. */
1136 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1142 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1147 if (odp_flow_key_to_flow(key, key_len, flow)) {
1148 /* This should not happen: it indicates that odp_flow_key_from_flow()
1149 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1150 * flow. Log the problem as an error, with enough details to enable
1152 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1154 if (!VLOG_DROP_ERR(&rl)) {
1158 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1159 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1166 in_port = flow->in_port.odp_port;
1167 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1175 dpif_netdev_flow_get(const struct dpif *dpif,
1176 const struct nlattr *nl_key, size_t nl_key_len,
1177 struct ofpbuf **bufp,
1178 struct nlattr **maskp, size_t *mask_len,
1179 struct nlattr **actionsp, size_t *actions_len,
1180 struct dpif_flow_stats *stats)
1182 struct dp_netdev *dp = get_dp_netdev(dpif);
1183 struct dp_netdev_flow *netdev_flow;
1187 error = dpif_netdev_flow_from_nlattrs(nl_key, nl_key_len, &key);
1192 netdev_flow = dp_netdev_find_flow(dp, &key);
1196 get_dpif_flow_stats(netdev_flow, stats);
1200 struct flow_wildcards wc;
1202 *bufp = ofpbuf_new(sizeof(struct odputil_keybuf));
1203 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1204 odp_flow_key_from_mask(*bufp, &wc.masks, &netdev_flow->flow,
1205 odp_to_u32(wc.masks.in_port.odp_port),
1207 *maskp = ofpbuf_data(*bufp);
1208 *mask_len = ofpbuf_size(*bufp);
1211 struct dp_netdev_actions *actions;
1213 actions = dp_netdev_flow_get_actions(netdev_flow);
1214 *actionsp = actions->actions;
1215 *actions_len = actions->size;
1225 dp_netdev_flow_add(struct dp_netdev *dp, struct match *match,
1226 const struct nlattr *actions, size_t actions_len)
1227 OVS_REQUIRES(dp->flow_mutex)
1229 struct dp_netdev_flow *netdev_flow;
1231 netdev_flow = xzalloc(sizeof *netdev_flow);
1232 *CONST_CAST(struct flow *, &netdev_flow->flow) = match->flow;
1234 ovs_refcount_init(&netdev_flow->ref_cnt);
1236 ovsthread_stats_init(&netdev_flow->stats);
1238 ovsrcu_set(&netdev_flow->actions,
1239 dp_netdev_actions_create(actions, actions_len));
1241 cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr),
1242 match, NETDEV_RULE_PRIORITY);
1243 cmap_insert(&dp->flow_table,
1244 CONST_CAST(struct cmap_node *, &netdev_flow->node),
1245 flow_hash(&match->flow, 0));
1246 classifier_insert(&dp->cls,
1247 CONST_CAST(struct cls_rule *, &netdev_flow->cr));
1253 clear_stats(struct dp_netdev_flow *netdev_flow)
1255 struct dp_netdev_flow_stats *bucket;
1258 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1259 ovs_mutex_lock(&bucket->mutex);
1261 bucket->packet_count = 0;
1262 bucket->byte_count = 0;
1263 bucket->tcp_flags = 0;
1264 ovs_mutex_unlock(&bucket->mutex);
1269 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1271 struct dp_netdev *dp = get_dp_netdev(dpif);
1272 struct dp_netdev_flow *netdev_flow;
1273 struct miniflow miniflow;
1277 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
1281 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1282 put->mask, put->mask_len,
1283 &match.flow, &match.wc.masks);
1287 miniflow_init(&miniflow, &match.flow);
1289 ovs_mutex_lock(&dp->flow_mutex);
1290 netdev_flow = dp_netdev_lookup_flow(dp, &miniflow);
1292 if (put->flags & DPIF_FP_CREATE) {
1293 if (cmap_count(&dp->flow_table) < MAX_FLOWS) {
1295 memset(put->stats, 0, sizeof *put->stats);
1297 error = dp_netdev_flow_add(dp, &match, put->actions,
1306 if (put->flags & DPIF_FP_MODIFY
1307 && flow_equal(&match.flow, &netdev_flow->flow)) {
1308 struct dp_netdev_actions *new_actions;
1309 struct dp_netdev_actions *old_actions;
1311 new_actions = dp_netdev_actions_create(put->actions,
1314 old_actions = dp_netdev_flow_get_actions(netdev_flow);
1315 ovsrcu_set(&netdev_flow->actions, new_actions);
1318 get_dpif_flow_stats(netdev_flow, put->stats);
1320 if (put->flags & DPIF_FP_ZERO_STATS) {
1321 clear_stats(netdev_flow);
1324 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
1325 } else if (put->flags & DPIF_FP_CREATE) {
1328 /* Overlapping flow. */
1332 ovs_mutex_unlock(&dp->flow_mutex);
1333 miniflow_destroy(&miniflow);
1339 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
1341 struct dp_netdev *dp = get_dp_netdev(dpif);
1342 struct dp_netdev_flow *netdev_flow;
1346 error = dpif_netdev_flow_from_nlattrs(del->key, del->key_len, &key);
1351 ovs_mutex_lock(&dp->flow_mutex);
1352 netdev_flow = dp_netdev_find_flow(dp, &key);
1355 get_dpif_flow_stats(netdev_flow, del->stats);
1357 dp_netdev_remove_flow(dp, netdev_flow);
1361 ovs_mutex_unlock(&dp->flow_mutex);
1366 struct dpif_netdev_flow_dump {
1367 struct dpif_flow_dump up;
1368 struct cmap_position pos;
1370 struct ovs_mutex mutex;
1373 static struct dpif_netdev_flow_dump *
1374 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
1376 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
1379 static struct dpif_flow_dump *
1380 dpif_netdev_flow_dump_create(const struct dpif *dpif_)
1382 struct dpif_netdev_flow_dump *dump;
1384 dump = xmalloc(sizeof *dump);
1385 dpif_flow_dump_init(&dump->up, dpif_);
1386 memset(&dump->pos, 0, sizeof dump->pos);
1388 ovs_mutex_init(&dump->mutex);
1394 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
1396 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1398 ovs_mutex_destroy(&dump->mutex);
1403 struct dpif_netdev_flow_dump_thread {
1404 struct dpif_flow_dump_thread up;
1405 struct dpif_netdev_flow_dump *dump;
1406 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
1407 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
1410 static struct dpif_netdev_flow_dump_thread *
1411 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
1413 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
1416 static struct dpif_flow_dump_thread *
1417 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
1419 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1420 struct dpif_netdev_flow_dump_thread *thread;
1422 thread = xmalloc(sizeof *thread);
1423 dpif_flow_dump_thread_init(&thread->up, &dump->up);
1424 thread->dump = dump;
1429 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
1431 struct dpif_netdev_flow_dump_thread *thread
1432 = dpif_netdev_flow_dump_thread_cast(thread_);
1438 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
1439 struct dpif_flow *flows, int max_flows)
1441 struct dpif_netdev_flow_dump_thread *thread
1442 = dpif_netdev_flow_dump_thread_cast(thread_);
1443 struct dpif_netdev_flow_dump *dump = thread->dump;
1444 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
1445 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
1446 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
1450 ovs_mutex_lock(&dump->mutex);
1451 if (!dump->status) {
1452 for (n_flows = 0; n_flows < MIN(max_flows, FLOW_DUMP_MAX_BATCH);
1454 struct cmap_node *node;
1456 node = cmap_next_position(&dp->flow_table, &dump->pos);
1461 netdev_flows[n_flows] = CONTAINER_OF(node, struct dp_netdev_flow,
1465 ovs_mutex_unlock(&dump->mutex);
1467 for (i = 0; i < n_flows; i++) {
1468 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
1469 struct odputil_keybuf *keybuf = &thread->keybuf[i];
1470 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
1471 struct dpif_flow *f = &flows[i];
1472 struct dp_netdev_actions *dp_actions;
1473 struct flow_wildcards wc;
1476 minimask_expand(&netdev_flow->cr.match.mask, &wc);
1479 ofpbuf_use_stack(&buf, keybuf, sizeof *keybuf);
1480 odp_flow_key_from_flow(&buf, &netdev_flow->flow, &wc.masks,
1481 netdev_flow->flow.in_port.odp_port, true);
1482 f->key = ofpbuf_data(&buf);
1483 f->key_len = ofpbuf_size(&buf);
1486 ofpbuf_use_stack(&buf, maskbuf, sizeof *maskbuf);
1487 odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
1488 odp_to_u32(wc.masks.in_port.odp_port),
1490 f->mask = ofpbuf_data(&buf);
1491 f->mask_len = ofpbuf_size(&buf);
1494 dp_actions = dp_netdev_flow_get_actions(netdev_flow);
1495 f->actions = dp_actions->actions;
1496 f->actions_len = dp_actions->size;
1499 get_dpif_flow_stats(netdev_flow, &f->stats);
1506 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
1508 struct dp_netdev *dp = get_dp_netdev(dpif);
1509 struct dpif_packet packet, *pp;
1510 struct pkt_metadata *md = &execute->md;
1512 if (ofpbuf_size(execute->packet) < ETH_HEADER_LEN ||
1513 ofpbuf_size(execute->packet) > UINT16_MAX) {
1517 packet.ofpbuf = *execute->packet;
1520 dp_netdev_execute_actions(dp, &pp, 1, false, md,
1521 execute->actions, execute->actions_len);
1523 /* Even though may_steal is set to false, some actions could modify or
1524 * reallocate the ofpbuf memory. We need to pass those changes to the
1526 *execute->packet = packet.ofpbuf;
1532 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
1536 for (i = 0; i < n_ops; i++) {
1537 struct dpif_op *op = ops[i];
1540 case DPIF_OP_FLOW_PUT:
1541 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
1544 case DPIF_OP_FLOW_DEL:
1545 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
1548 case DPIF_OP_EXECUTE:
1549 op->error = dpif_netdev_execute(dpif, &op->u.execute);
1556 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
1557 uint32_t queue_id, uint32_t *priority)
1559 *priority = queue_id;
1564 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
1565 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
1567 struct dp_netdev_actions *
1568 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
1570 struct dp_netdev_actions *netdev_actions;
1572 netdev_actions = xmalloc(sizeof *netdev_actions);
1573 netdev_actions->actions = xmemdup(actions, size);
1574 netdev_actions->size = size;
1576 return netdev_actions;
1579 struct dp_netdev_actions *
1580 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
1582 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
1586 dp_netdev_actions_free(struct dp_netdev_actions *actions)
1588 free(actions->actions);
1594 dp_netdev_process_rxq_port(struct dp_netdev *dp,
1595 struct dp_netdev_port *port,
1596 struct netdev_rxq *rxq)
1598 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
1601 error = netdev_rxq_recv(rxq, packets, &cnt);
1603 dp_netdev_port_input(dp, packets, cnt, port->port_no);
1604 } else if (error != EAGAIN && error != EOPNOTSUPP) {
1605 static struct vlog_rate_limit rl
1606 = VLOG_RATE_LIMIT_INIT(1, 5);
1608 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
1609 netdev_get_name(port->netdev),
1610 ovs_strerror(error));
1615 dpif_netdev_run(struct dpif *dpif)
1617 struct dp_netdev_port *port;
1618 struct dp_netdev *dp = get_dp_netdev(dpif);
1620 CMAP_FOR_EACH (port, node, &dp->ports) {
1621 if (!netdev_is_pmd(port->netdev)) {
1624 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1625 dp_netdev_process_rxq_port(dp, port, port->rxq[i]);
1632 dpif_netdev_wait(struct dpif *dpif)
1634 struct dp_netdev_port *port;
1635 struct dp_netdev *dp = get_dp_netdev(dpif);
1637 ovs_mutex_lock(&dp_netdev_mutex);
1638 CMAP_FOR_EACH (port, node, &dp->ports) {
1639 if (!netdev_is_pmd(port->netdev)) {
1642 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1643 netdev_rxq_wait(port->rxq[i]);
1647 ovs_mutex_unlock(&dp_netdev_mutex);
1651 struct dp_netdev_port *port;
1652 struct netdev_rxq *rx;
1656 pmd_load_queues(struct pmd_thread *f,
1657 struct rxq_poll **ppoll_list, int poll_cnt)
1659 struct dp_netdev *dp = f->dp;
1660 struct rxq_poll *poll_list = *ppoll_list;
1661 struct dp_netdev_port *port;
1666 /* Simple scheduler for netdev rx polling. */
1667 for (i = 0; i < poll_cnt; i++) {
1668 port_unref(poll_list[i].port);
1674 CMAP_FOR_EACH (port, node, &f->dp->ports) {
1675 if (netdev_is_pmd(port->netdev)) {
1678 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
1679 if ((index % dp->n_pmd_threads) == id) {
1680 poll_list = xrealloc(poll_list, sizeof *poll_list * (poll_cnt + 1));
1683 poll_list[poll_cnt].port = port;
1684 poll_list[poll_cnt].rx = port->rxq[i];
1692 *ppoll_list = poll_list;
1697 pmd_thread_main(void *f_)
1699 struct pmd_thread *f = f_;
1700 struct dp_netdev *dp = f->dp;
1701 unsigned int lc = 0;
1702 struct rxq_poll *poll_list;
1703 unsigned int port_seq;
1710 pmd_thread_setaffinity_cpu(f->id);
1712 poll_cnt = pmd_load_queues(f, &poll_list, poll_cnt);
1713 atomic_read(&f->change_seq, &port_seq);
1716 unsigned int c_port_seq;
1719 for (i = 0; i < poll_cnt; i++) {
1720 dp_netdev_process_rxq_port(dp, poll_list[i].port, poll_list[i].rx);
1726 /* XXX: need completely userspace based signaling method.
1727 * to keep this thread entirely in userspace.
1728 * For now using atomic counter. */
1730 atomic_read_explicit(&f->change_seq, &c_port_seq, memory_order_consume);
1731 if (c_port_seq != port_seq) {
1737 if (!latch_is_set(&f->dp->exit_latch)){
1741 for (i = 0; i < poll_cnt; i++) {
1742 port_unref(poll_list[i].port);
1750 dp_netdev_disable_upcall(struct dp_netdev *dp)
1751 OVS_ACQUIRES(dp->upcall_rwlock)
1753 fat_rwlock_wrlock(&dp->upcall_rwlock);
1757 dpif_netdev_disable_upcall(struct dpif *dpif)
1758 OVS_NO_THREAD_SAFETY_ANALYSIS
1760 struct dp_netdev *dp = get_dp_netdev(dpif);
1761 dp_netdev_disable_upcall(dp);
1765 dp_netdev_enable_upcall(struct dp_netdev *dp)
1766 OVS_RELEASES(dp->upcall_rwlock)
1768 fat_rwlock_unlock(&dp->upcall_rwlock);
1772 dpif_netdev_enable_upcall(struct dpif *dpif)
1773 OVS_NO_THREAD_SAFETY_ANALYSIS
1775 struct dp_netdev *dp = get_dp_netdev(dpif);
1776 dp_netdev_enable_upcall(dp);
1780 dp_netdev_set_pmd_threads(struct dp_netdev *dp, int n)
1784 if (n == dp->n_pmd_threads) {
1788 /* Stop existing threads. */
1789 latch_set(&dp->exit_latch);
1790 dp_netdev_reload_pmd_threads(dp);
1791 for (i = 0; i < dp->n_pmd_threads; i++) {
1792 struct pmd_thread *f = &dp->pmd_threads[i];
1794 xpthread_join(f->thread, NULL);
1796 latch_poll(&dp->exit_latch);
1797 free(dp->pmd_threads);
1799 /* Start new threads. */
1800 dp->pmd_threads = xmalloc(n * sizeof *dp->pmd_threads);
1801 dp->n_pmd_threads = n;
1803 for (i = 0; i < n; i++) {
1804 struct pmd_thread *f = &dp->pmd_threads[i];
1808 atomic_store(&f->change_seq, 1);
1810 /* Each thread will distribute all devices rx-queues among
1812 f->thread = ovs_thread_create("pmd", pmd_thread_main, f);
1818 dp_netdev_flow_stats_new_cb(void)
1820 struct dp_netdev_flow_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1821 ovs_mutex_init(&bucket->mutex);
1826 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow,
1830 long long int now = time_msec();
1831 struct dp_netdev_flow_stats *bucket;
1833 bucket = ovsthread_stats_bucket_get(&netdev_flow->stats,
1834 dp_netdev_flow_stats_new_cb);
1836 ovs_mutex_lock(&bucket->mutex);
1837 bucket->used = MAX(now, bucket->used);
1838 bucket->packet_count += cnt;
1839 bucket->byte_count += size;
1840 bucket->tcp_flags |= tcp_flags;
1841 ovs_mutex_unlock(&bucket->mutex);
1845 dp_netdev_stats_new_cb(void)
1847 struct dp_netdev_stats *bucket = xzalloc_cacheline(sizeof *bucket);
1848 ovs_mutex_init(&bucket->mutex);
1853 dp_netdev_count_packet(struct dp_netdev *dp, enum dp_stat_type type, int cnt)
1855 struct dp_netdev_stats *bucket;
1857 bucket = ovsthread_stats_bucket_get(&dp->stats, dp_netdev_stats_new_cb);
1858 ovs_mutex_lock(&bucket->mutex);
1859 bucket->n[type] += cnt;
1860 ovs_mutex_unlock(&bucket->mutex);
1863 struct packet_batch {
1864 unsigned int packet_count;
1865 unsigned int byte_count;
1868 struct dp_netdev_flow *flow;
1870 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
1871 struct pkt_metadata md;
1875 packet_batch_update(struct packet_batch *batch,
1876 struct dpif_packet *packet, const struct miniflow *mf)
1878 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
1879 batch->packets[batch->packet_count++] = packet;
1880 batch->byte_count += ofpbuf_size(&packet->ofpbuf);
1884 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow,
1885 struct pkt_metadata *md)
1890 batch->packet_count = 0;
1891 batch->byte_count = 0;
1892 batch->tcp_flags = 0;
1896 packet_batch_execute(struct packet_batch *batch, struct dp_netdev *dp)
1898 struct dp_netdev_actions *actions;
1899 struct dp_netdev_flow *flow = batch->flow;
1901 dp_netdev_flow_used(batch->flow, batch->packet_count, batch->byte_count,
1904 actions = dp_netdev_flow_get_actions(flow);
1906 dp_netdev_execute_actions(dp, batch->packets,
1907 batch->packet_count, true, &batch->md,
1908 actions->actions, actions->size);
1910 dp_netdev_count_packet(dp, DP_STAT_HIT, batch->packet_count);
1914 dp_netdev_input(struct dp_netdev *dp, struct dpif_packet **packets, int cnt,
1915 struct pkt_metadata *md)
1917 struct dp_netdev_queue q = DP_NETDEV_QUEUE_INITIALIZER;
1918 struct packet_batch batches[NETDEV_MAX_RX_BATCH];
1919 struct netdev_flow_key keys[NETDEV_MAX_RX_BATCH];
1920 const struct miniflow *mfs[NETDEV_MAX_RX_BATCH]; /* NULL at bad packets. */
1921 struct cls_rule *rules[NETDEV_MAX_RX_BATCH];
1922 size_t n_batches, i;
1924 for (i = 0; i < cnt; i++) {
1925 if (OVS_UNLIKELY(ofpbuf_size(&packets[i]->ofpbuf) < ETH_HEADER_LEN)) {
1926 dpif_packet_delete(packets[i]);
1931 miniflow_initialize(&keys[i].flow, keys[i].buf);
1932 miniflow_extract(&packets[i]->ofpbuf, md, &keys[i].flow);
1933 mfs[i] = &keys[i].flow;
1936 classifier_lookup_miniflow_batch(&dp->cls, mfs, rules, cnt);
1939 for (i = 0; i < cnt; i++) {
1940 struct dp_netdev_flow *flow;
1941 struct packet_batch *batch;
1944 if (OVS_UNLIKELY(!mfs[i])) {
1948 if (OVS_UNLIKELY(!rules[i])) {
1949 struct ofpbuf *buf = &packets[i]->ofpbuf;
1951 dp_netdev_count_packet(dp, DP_STAT_MISS, 1);
1952 dp_netdev_queue_userspace_packet(&q, buf, DPIF_UC_MISS,
1954 dpif_packet_delete(packets[i]);
1958 /* XXX: This O(n^2) algortihm makes sense if we're operating under the
1959 * assumption that the number of distinct flows (and therefore the
1960 * number of distinct batches) is quite small. If this turns out not
1961 * to be the case, it may make sense to pre sort based on the
1962 * netdev_flow pointer. That done we can get the appropriate batching
1963 * in O(n * log(n)) instead. */
1965 flow = dp_netdev_flow_cast(rules[i]);
1966 for (j = 0; j < n_batches; j++) {
1967 if (batches[j].flow == flow) {
1968 batch = &batches[j];
1974 batch = &batches[n_batches++];
1975 packet_batch_init(batch, flow, md);
1977 packet_batch_update(batch, packets[i], mfs[i]);
1980 for (i = 0; i < n_batches; i++) {
1981 packet_batch_execute(&batches[i], dp);
1984 if (q.packet_count) {
1985 dp_netdev_execute_userspace_queue(&q, dp);
1990 dp_netdev_port_input(struct dp_netdev *dp, struct dpif_packet **packets,
1991 int cnt, odp_port_t port_no)
1993 uint32_t *recirc_depth = recirc_depth_get();
1994 struct pkt_metadata md = PKT_METADATA_INITIALIZER(port_no);
1997 dp_netdev_input(dp, packets, cnt, &md);
2001 dp_netdev_queue_userspace_packet(struct dp_netdev_queue *q,
2002 struct ofpbuf *packet, int type,
2003 const struct miniflow *key,
2004 const struct nlattr *userdata)
2006 if (q->packet_count < NETDEV_MAX_RX_BATCH) {
2007 int cnt = q->packet_count;
2008 struct dpif_upcall *upcall = &q->upcalls[cnt];
2009 struct ofpbuf *buf = &q->bufs[cnt];
2014 upcall->type = type;
2016 /* Allocate buffer big enough for everything. */
2017 buf_size = ODPUTIL_FLOW_KEY_BYTES;
2019 buf_size += NLA_ALIGN(userdata->nla_len);
2021 buf_size += ofpbuf_size(packet);
2022 ofpbuf_init(buf, buf_size);
2025 miniflow_expand(key, &flow);
2026 odp_flow_key_from_flow(buf, &flow, NULL, flow.in_port.odp_port, true);
2027 upcall->key = ofpbuf_data(buf);
2028 upcall->key_len = ofpbuf_size(buf);
2032 upcall->userdata = ofpbuf_put(buf, userdata,
2033 NLA_ALIGN(userdata->nla_len));
2036 /* We have to perform a copy of the packet, because we cannot send DPDK
2037 * mbufs to a non pmd thread. When the upcall processing will be done
2038 * in the pmd thread, this copy can be avoided */
2039 data = ofpbuf_put(buf, ofpbuf_data(packet), ofpbuf_size(packet));
2040 ofpbuf_use_stub(&upcall->packet, data, ofpbuf_size(packet));
2041 ofpbuf_set_size(&upcall->packet, ofpbuf_size(packet));
2051 dp_netdev_execute_userspace_queue(struct dp_netdev_queue *q,
2052 struct dp_netdev *dp)
2054 struct dpif_upcall *upcalls = q->upcalls;
2055 struct ofpbuf *bufs = q->bufs;
2056 int cnt = q->packet_count;
2058 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
2059 ovs_assert(dp->upcall_cb);
2060 dp->upcall_cb(dp->dpif, upcalls, bufs, cnt);
2061 fat_rwlock_unlock(&dp->upcall_rwlock);
2065 for (i = 0; i < cnt; i++) {
2066 ofpbuf_uninit(&bufs[i]);
2067 ofpbuf_uninit(&upcalls[i].packet);
2072 struct dp_netdev_execute_aux {
2073 struct dp_netdev *dp;
2077 dpif_netdev_register_upcall_cb(struct dpif *dpif, exec_upcall_cb *cb)
2079 struct dp_netdev *dp = get_dp_netdev(dpif);
2084 dp_execute_cb(void *aux_, struct dpif_packet **packets, int cnt,
2085 struct pkt_metadata *md,
2086 const struct nlattr *a, bool may_steal)
2087 OVS_NO_THREAD_SAFETY_ANALYSIS
2089 struct dp_netdev_execute_aux *aux = aux_;
2090 int type = nl_attr_type(a);
2091 struct dp_netdev_port *p;
2092 uint32_t *depth = recirc_depth_get();
2095 switch ((enum ovs_action_attr)type) {
2096 case OVS_ACTION_ATTR_OUTPUT:
2097 p = dp_netdev_lookup_port(aux->dp, u32_to_odp(nl_attr_get_u32(a)));
2098 if (OVS_LIKELY(p)) {
2099 netdev_send(p->netdev, packets, cnt, may_steal);
2100 } else if (may_steal) {
2101 for (i = 0; i < cnt; i++) {
2102 dpif_packet_delete(packets[i]);
2107 case OVS_ACTION_ATTR_USERSPACE: {
2108 const struct nlattr *userdata;
2109 struct netdev_flow_key key;
2110 struct dp_netdev_queue q = DP_NETDEV_QUEUE_INITIALIZER;
2112 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
2114 miniflow_initialize(&key.flow, key.buf);
2116 for (i = 0; i < cnt; i++) {
2117 struct ofpbuf *packet;
2119 packet = &packets[i]->ofpbuf;
2121 miniflow_extract(packet, md, &key.flow);
2123 dp_netdev_queue_userspace_packet(&q, packet,
2124 DPIF_UC_ACTION, &key.flow,
2127 dpif_packet_delete(packets[i]);
2131 if (q.packet_count) {
2132 dp_netdev_execute_userspace_queue(&q, aux->dp);
2137 case OVS_ACTION_ATTR_HASH: {
2138 const struct ovs_action_hash *hash_act;
2139 struct netdev_flow_key key;
2142 hash_act = nl_attr_get(a);
2144 miniflow_initialize(&key.flow, key.buf);
2146 for (i = 0; i < cnt; i++) {
2148 /* XXX: this is slow. Use RSS hash in the future */
2149 miniflow_extract(&packets[i]->ofpbuf, md, &key.flow);
2151 if (hash_act->hash_alg == OVS_HASH_ALG_L4) {
2152 /* Hash need not be symmetric, nor does it need to include
2154 hash = miniflow_hash_5tuple(&key.flow, hash_act->hash_basis);
2156 VLOG_WARN("Unknown hash algorithm specified "
2157 "for the hash action.");
2162 hash = 1; /* 0 is not valid */
2168 packets[i]->dp_hash = hash;
2173 case OVS_ACTION_ATTR_RECIRC:
2174 if (*depth < MAX_RECIRC_DEPTH) {
2177 for (i = 0; i < cnt; i++) {
2178 struct dpif_packet *recirc_pkt;
2179 struct pkt_metadata recirc_md = *md;
2181 recirc_pkt = (may_steal) ? packets[i]
2182 : dpif_packet_clone(packets[i]);
2184 recirc_md.recirc_id = nl_attr_get_u32(a);
2186 /* Hash is private to each packet */
2187 recirc_md.dp_hash = packets[i]->dp_hash;
2189 dp_netdev_input(aux->dp, &recirc_pkt, 1, &recirc_md);
2195 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
2197 for (i = 0; i < cnt; i++) {
2198 dpif_packet_delete(packets[i]);
2204 case OVS_ACTION_ATTR_PUSH_VLAN:
2205 case OVS_ACTION_ATTR_POP_VLAN:
2206 case OVS_ACTION_ATTR_PUSH_MPLS:
2207 case OVS_ACTION_ATTR_POP_MPLS:
2208 case OVS_ACTION_ATTR_SET:
2209 case OVS_ACTION_ATTR_SAMPLE:
2210 case OVS_ACTION_ATTR_UNSPEC:
2211 case __OVS_ACTION_ATTR_MAX:
2217 dp_netdev_execute_actions(struct dp_netdev *dp,
2218 struct dpif_packet **packets, int cnt,
2219 bool may_steal, struct pkt_metadata *md,
2220 const struct nlattr *actions, size_t actions_len)
2222 struct dp_netdev_execute_aux aux = {dp};
2224 odp_execute_actions(&aux, packets, cnt, may_steal, md, actions,
2225 actions_len, dp_execute_cb);
2228 const struct dpif_class dpif_netdev_class = {
2230 dpif_netdev_enumerate,
2231 dpif_netdev_port_open_type,
2234 dpif_netdev_destroy,
2237 dpif_netdev_get_stats,
2238 dpif_netdev_port_add,
2239 dpif_netdev_port_del,
2240 dpif_netdev_port_query_by_number,
2241 dpif_netdev_port_query_by_name,
2242 NULL, /* port_get_pid */
2243 dpif_netdev_port_dump_start,
2244 dpif_netdev_port_dump_next,
2245 dpif_netdev_port_dump_done,
2246 dpif_netdev_port_poll,
2247 dpif_netdev_port_poll_wait,
2248 dpif_netdev_flow_get,
2249 dpif_netdev_flow_flush,
2250 dpif_netdev_flow_dump_create,
2251 dpif_netdev_flow_dump_destroy,
2252 dpif_netdev_flow_dump_thread_create,
2253 dpif_netdev_flow_dump_thread_destroy,
2254 dpif_netdev_flow_dump_next,
2255 dpif_netdev_operate,
2256 NULL, /* recv_set */
2257 NULL, /* handlers_set */
2258 dpif_netdev_queue_to_priority,
2260 NULL, /* recv_wait */
2261 NULL, /* recv_purge */
2262 dpif_netdev_register_upcall_cb,
2263 dpif_netdev_enable_upcall,
2264 dpif_netdev_disable_upcall,
2268 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
2269 const char *argv[], void *aux OVS_UNUSED)
2271 struct dp_netdev_port *old_port;
2272 struct dp_netdev_port *new_port;
2273 struct dp_netdev *dp;
2276 ovs_mutex_lock(&dp_netdev_mutex);
2277 dp = shash_find_data(&dp_netdevs, argv[1]);
2278 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2279 ovs_mutex_unlock(&dp_netdev_mutex);
2280 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2283 ovs_refcount_ref(&dp->ref_cnt);
2284 ovs_mutex_unlock(&dp_netdev_mutex);
2286 ovs_mutex_lock(&dp->port_mutex);
2287 if (get_port_by_name(dp, argv[2], &old_port)) {
2288 unixctl_command_reply_error(conn, "unknown port");
2292 port_no = u32_to_odp(atoi(argv[3]));
2293 if (!port_no || port_no == ODPP_NONE) {
2294 unixctl_command_reply_error(conn, "bad port number");
2297 if (dp_netdev_lookup_port(dp, port_no)) {
2298 unixctl_command_reply_error(conn, "port number already in use");
2302 /* Remove old port. */
2303 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
2304 ovsrcu_postpone(free, old_port);
2306 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
2307 new_port = xmemdup(old_port, sizeof *old_port);
2308 new_port->port_no = port_no;
2309 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
2311 seq_change(dp->port_seq);
2312 unixctl_command_reply(conn, NULL);
2315 ovs_mutex_unlock(&dp->port_mutex);
2316 dp_netdev_unref(dp);
2320 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
2321 const char *argv[], void *aux OVS_UNUSED)
2323 struct dp_netdev_port *port;
2324 struct dp_netdev *dp;
2326 ovs_mutex_lock(&dp_netdev_mutex);
2327 dp = shash_find_data(&dp_netdevs, argv[1]);
2328 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
2329 ovs_mutex_unlock(&dp_netdev_mutex);
2330 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
2333 ovs_refcount_ref(&dp->ref_cnt);
2334 ovs_mutex_unlock(&dp_netdev_mutex);
2336 ovs_mutex_lock(&dp->port_mutex);
2337 if (get_port_by_name(dp, argv[2], &port)) {
2338 unixctl_command_reply_error(conn, "unknown port");
2339 } else if (port->port_no == ODPP_LOCAL) {
2340 unixctl_command_reply_error(conn, "can't delete local port");
2342 do_del_port(dp, port);
2343 unixctl_command_reply(conn, NULL);
2345 ovs_mutex_unlock(&dp->port_mutex);
2347 dp_netdev_unref(dp);
2351 dpif_dummy_register__(const char *type)
2353 struct dpif_class *class;
2355 class = xmalloc(sizeof *class);
2356 *class = dpif_netdev_class;
2357 class->type = xstrdup(type);
2358 dp_register_provider(class);
2362 dpif_dummy_register(bool override)
2369 dp_enumerate_types(&types);
2370 SSET_FOR_EACH (type, &types) {
2371 if (!dp_unregister_provider(type)) {
2372 dpif_dummy_register__(type);
2375 sset_destroy(&types);
2378 dpif_dummy_register__("dummy");
2380 unixctl_command_register("dpif-dummy/change-port-number",
2381 "DP PORT NEW-NUMBER",
2382 3, 3, dpif_dummy_change_port_number, NULL);
2383 unixctl_command_register("dpif-dummy/delete-port", "DP PORT",
2384 2, 2, dpif_dummy_delete_port, NULL);