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>
37 #include "dpif-provider.h"
39 #include "dynamic-string.h"
40 #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"
57 #include "packet-dpif.h"
59 #include "poll-loop.h"
70 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
72 #define FLOW_DUMP_MAX_BATCH 50
73 /* Use per thread recirc_depth to prevent recirculation loop. */
74 #define MAX_RECIRC_DEPTH 5
75 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
77 /* Configuration parameters. */
78 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
80 /* Protects against changes to 'dp_netdevs'. */
81 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
83 /* Contains all 'struct dp_netdev's. */
84 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
85 = SHASH_INITIALIZER(&dp_netdevs);
87 static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600);
89 /* Stores a miniflow with inline values */
91 struct netdev_flow_key {
92 uint32_t hash; /* Hash function differs for different users. */
93 uint32_t len; /* Length of the following miniflow (incl. map). */
95 uint32_t buf[FLOW_MAX_PACKET_U32S - MINI_N_INLINE];
98 /* Exact match cache for frequently used flows
100 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
101 * search its entries for a miniflow that matches exactly the miniflow of the
102 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
104 * A cache entry holds a reference to its 'dp_netdev_flow'.
106 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
107 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
108 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
109 * value is the index of a cache entry where the miniflow could be.
115 * Each pmd_thread has its own private exact match cache.
116 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
119 #define EM_FLOW_HASH_SHIFT 10
120 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
121 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
122 #define EM_FLOW_HASH_SEGS 2
125 struct dp_netdev_flow *flow;
126 struct netdev_flow_key key; /* key.hash used for emc hash value. */
130 struct emc_entry entries[EM_FLOW_HASH_ENTRIES];
133 /* Iterate in the exact match cache through every entry that might contain a
134 * miniflow with hash 'HASH'. */
135 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
136 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
137 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
138 i__ < EM_FLOW_HASH_SEGS; \
139 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
141 /* Simple non-wildcarding single-priority classifier. */
144 struct cmap subtables_map;
145 struct pvector subtables;
148 /* A rule to be inserted to the classifier. */
150 struct cmap_node cmap_node; /* Within struct dpcls_subtable 'rules'. */
151 struct netdev_flow_key *mask; /* Subtable's mask. */
152 struct netdev_flow_key flow; /* Matching key. */
153 /* 'flow' must be the last field, additional space is allocated here. */
156 static void dpcls_init(struct dpcls *);
157 static void dpcls_destroy(struct dpcls *);
158 static void dpcls_insert(struct dpcls *, struct dpcls_rule *,
159 const struct netdev_flow_key *mask);
160 static void dpcls_remove(struct dpcls *, struct dpcls_rule *);
161 static bool dpcls_lookup(const struct dpcls *cls,
162 const struct netdev_flow_key keys[],
163 struct dpcls_rule **rules, size_t cnt);
165 /* Datapath based on the network device interface from netdev.h.
171 * Some members, marked 'const', are immutable. Accessing other members
172 * requires synchronization, as noted in more detail below.
174 * Acquisition order is, from outermost to innermost:
176 * dp_netdev_mutex (global)
181 const struct dpif_class *const class;
182 const char *const name;
184 struct ovs_refcount ref_cnt;
185 atomic_flag destroyed;
189 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
190 * changes to 'cls' must be made while still holding the 'flow_mutex'.
192 struct ovs_mutex flow_mutex;
194 struct cmap flow_table OVS_GUARDED; /* Flow table. */
198 * ovsthread_stats is internally synchronized. */
199 struct ovsthread_stats stats; /* Contains 'struct dp_netdev_stats *'. */
203 * Protected by RCU. Take the mutex to add or remove ports. */
204 struct ovs_mutex port_mutex;
206 struct seq *port_seq; /* Incremented whenever a port changes. */
208 /* Protects access to ofproto-dpif-upcall interface during revalidator
209 * thread synchronization. */
210 struct fat_rwlock upcall_rwlock;
211 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
214 /* Stores all 'struct dp_netdev_pmd_thread's. */
215 struct cmap poll_threads;
217 /* Protects the access of the 'struct dp_netdev_pmd_thread'
218 * instance for non-pmd thread. */
219 struct ovs_mutex non_pmd_mutex;
221 /* Each pmd thread will store its pointer to
222 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
223 ovsthread_key_t per_pmd_key;
225 /* Number of rx queues for each dpdk interface and the cpu mask
226 * for pin of pmd threads. */
231 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
235 DP_STAT_HIT, /* Packets that matched in the flow table. */
236 DP_STAT_MISS, /* Packets that did not match. */
237 DP_STAT_LOST, /* Packets not passed up to the client. */
241 /* Contained by struct dp_netdev's 'stats' member. */
242 struct dp_netdev_stats {
243 struct ovs_mutex mutex; /* Protects 'n'. */
245 /* Indexed by DP_STAT_*, protected by 'mutex'. */
246 unsigned long long int n[DP_N_STATS] OVS_GUARDED;
250 /* A port in a netdev-based datapath. */
251 struct dp_netdev_port {
252 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
254 struct netdev *netdev;
255 struct netdev_saved_flags *sf;
256 struct netdev_rxq **rxq;
257 struct ovs_refcount ref_cnt;
258 char *type; /* Port type as requested by user. */
262 /* A flow in dp_netdev's 'flow_table'.
268 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
269 * its dp_netdev's classifier. The text below calls this classifier 'cls'.
274 * The thread safety rules described here for "struct dp_netdev_flow" are
275 * motivated by two goals:
277 * - Prevent threads that read members of "struct dp_netdev_flow" from
278 * reading bad data due to changes by some thread concurrently modifying
281 * - Prevent two threads making changes to members of a given "struct
282 * dp_netdev_flow" from interfering with each other.
288 * A flow 'flow' may be accessed without a risk of being freed during an RCU
289 * grace period. Code that needs to hold onto a flow for a while
290 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
292 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
293 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
296 * Some members, marked 'const', are immutable. Accessing other members
297 * requires synchronization, as noted in more detail below.
299 struct dp_netdev_flow {
302 /* Hash table index by unmasked flow. */
303 const struct cmap_node node; /* In owning dp_netdev's 'flow_table'. */
304 const struct flow flow; /* Unmasked flow that created this entry. */
306 /* Number of references.
307 * The classifier owns one reference.
308 * Any thread trying to keep a rule from being freed should hold its own
310 struct ovs_refcount ref_cnt;
314 * Reading or writing these members requires 'mutex'. */
315 struct ovsthread_stats stats; /* Contains "struct dp_netdev_flow_stats". */
318 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
320 /* Packet classification. */
321 struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
322 /* 'cr' must be the last member. */
325 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
326 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
328 /* Contained by struct dp_netdev_flow's 'stats' member. */
329 struct dp_netdev_flow_stats {
330 struct ovs_mutex mutex; /* Guards all the other members. */
332 long long int used OVS_GUARDED; /* Last used time, in monotonic msecs. */
333 long long int packet_count OVS_GUARDED; /* Number of packets matched. */
334 long long int byte_count OVS_GUARDED; /* Number of bytes matched. */
335 uint16_t tcp_flags OVS_GUARDED; /* Bitwise-OR of seen tcp_flags values. */
338 /* A set of datapath actions within a "struct dp_netdev_flow".
344 * A struct dp_netdev_actions 'actions' is protected with RCU. */
345 struct dp_netdev_actions {
346 /* These members are immutable: they do not change during the struct's
348 struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */
349 unsigned int size; /* Size of 'actions', in bytes. */
352 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
354 struct dp_netdev_actions *dp_netdev_flow_get_actions(
355 const struct dp_netdev_flow *);
356 static void dp_netdev_actions_free(struct dp_netdev_actions *);
358 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
359 * the performance overhead of interrupt processing. Therefore netdev can
360 * not implement rx-wait for these devices. dpif-netdev needs to poll
361 * these device to check for recv buffer. pmd-thread does polling for
362 * devices assigned to itself thread.
364 * DPDK used PMD for accessing NIC.
366 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
367 * I/O of all non-pmd threads. There will be no actual thread created
370 struct dp_netdev_pmd_thread {
371 struct dp_netdev *dp;
372 struct cmap_node node; /* In 'dp->poll_threads'. */
373 /* Per thread exact-match cache. Note, the instance for cpu core
374 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
375 * need to be protected (e.g. by 'dp_netdev_mutex'). All other
376 * instances will only be accessed by its own pmd thread. */
377 struct emc_cache flow_cache;
378 struct latch exit_latch; /* For terminating the pmd thread. */
379 atomic_uint change_seq; /* For reloading pmd ports. */
381 int index; /* Idx of this pmd thread among pmd*/
382 /* threads on same numa node. */
383 int core_id; /* CPU core id of this pmd thread. */
384 int numa_id; /* numa node id of this pmd thread. */
387 #define PMD_INITIAL_SEQ 1
389 /* Interface to netdev-based datapath. */
392 struct dp_netdev *dp;
393 uint64_t last_port_seq;
396 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
397 struct dp_netdev_port **portp);
398 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
399 struct dp_netdev_port **portp);
400 static void dp_netdev_free(struct dp_netdev *)
401 OVS_REQUIRES(dp_netdev_mutex);
402 static void dp_netdev_flow_flush(struct dp_netdev *);
403 static int do_add_port(struct dp_netdev *dp, const char *devname,
404 const char *type, odp_port_t port_no)
405 OVS_REQUIRES(dp->port_mutex);
406 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
407 OVS_REQUIRES(dp->port_mutex);
408 static int dpif_netdev_open(const struct dpif_class *, const char *name,
409 bool create, struct dpif **);
410 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
411 struct dpif_packet **, int c,
413 const struct nlattr *actions,
415 static void dp_netdev_input(struct dp_netdev_pmd_thread *,
416 struct dpif_packet **, int cnt);
418 static void dp_netdev_disable_upcall(struct dp_netdev *);
419 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
420 struct dp_netdev *dp, int index,
421 int core_id, int numa_id);
422 static void dp_netdev_set_nonpmd(struct dp_netdev *dp);
423 static struct dp_netdev_pmd_thread *dp_netdev_get_nonpmd(struct dp_netdev *dp);
424 static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
425 static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
426 static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id);
427 static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp);
429 static void emc_clear_entry(struct emc_entry *ce);
432 emc_cache_init(struct emc_cache *flow_cache)
436 BUILD_ASSERT(offsetof(struct miniflow, inline_values) == sizeof(uint64_t));
438 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
439 flow_cache->entries[i].flow = NULL;
440 flow_cache->entries[i].key.hash = 0;
441 flow_cache->entries[i].key.len
442 = offsetof(struct miniflow, inline_values);
443 miniflow_initialize(&flow_cache->entries[i].key.mf,
444 flow_cache->entries[i].key.buf);
449 emc_cache_uninit(struct emc_cache *flow_cache)
453 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
454 emc_clear_entry(&flow_cache->entries[i]);
458 static struct dpif_netdev *
459 dpif_netdev_cast(const struct dpif *dpif)
461 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
462 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
465 static struct dp_netdev *
466 get_dp_netdev(const struct dpif *dpif)
468 return dpif_netdev_cast(dpif)->dp;
472 dpif_netdev_enumerate(struct sset *all_dps,
473 const struct dpif_class *dpif_class)
475 struct shash_node *node;
477 ovs_mutex_lock(&dp_netdev_mutex);
478 SHASH_FOR_EACH(node, &dp_netdevs) {
479 struct dp_netdev *dp = node->data;
480 if (dpif_class != dp->class) {
481 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
482 * If the class doesn't match, skip this dpif. */
485 sset_add(all_dps, node->name);
487 ovs_mutex_unlock(&dp_netdev_mutex);
493 dpif_netdev_class_is_dummy(const struct dpif_class *class)
495 return class != &dpif_netdev_class;
499 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
501 return strcmp(type, "internal") ? type
502 : dpif_netdev_class_is_dummy(class) ? "dummy"
507 create_dpif_netdev(struct dp_netdev *dp)
509 uint16_t netflow_id = hash_string(dp->name, 0);
510 struct dpif_netdev *dpif;
512 ovs_refcount_ref(&dp->ref_cnt);
514 dpif = xmalloc(sizeof *dpif);
515 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
517 dpif->last_port_seq = seq_read(dp->port_seq);
522 /* Choose an unused, non-zero port number and return it on success.
523 * Return ODPP_NONE on failure. */
525 choose_port(struct dp_netdev *dp, const char *name)
526 OVS_REQUIRES(dp->port_mutex)
530 if (dp->class != &dpif_netdev_class) {
534 /* If the port name begins with "br", start the number search at
535 * 100 to make writing tests easier. */
536 if (!strncmp(name, "br", 2)) {
540 /* If the port name contains a number, try to assign that port number.
541 * This can make writing unit tests easier because port numbers are
543 for (p = name; *p != '\0'; p++) {
544 if (isdigit((unsigned char) *p)) {
545 port_no = start_no + strtol(p, NULL, 10);
546 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
547 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
548 return u32_to_odp(port_no);
555 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
556 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
557 return u32_to_odp(port_no);
565 create_dp_netdev(const char *name, const struct dpif_class *class,
566 struct dp_netdev **dpp)
567 OVS_REQUIRES(dp_netdev_mutex)
569 struct dp_netdev *dp;
572 dp = xzalloc(sizeof *dp);
573 shash_add(&dp_netdevs, name, dp);
575 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
576 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
577 ovs_refcount_init(&dp->ref_cnt);
578 atomic_flag_clear(&dp->destroyed);
580 ovs_mutex_init(&dp->flow_mutex);
581 dpcls_init(&dp->cls);
582 cmap_init(&dp->flow_table);
584 ovsthread_stats_init(&dp->stats);
586 ovs_mutex_init(&dp->port_mutex);
587 cmap_init(&dp->ports);
588 dp->port_seq = seq_create();
589 fat_rwlock_init(&dp->upcall_rwlock);
591 /* Disable upcalls by default. */
592 dp_netdev_disable_upcall(dp);
593 dp->upcall_aux = NULL;
594 dp->upcall_cb = NULL;
596 cmap_init(&dp->poll_threads);
597 ovs_mutex_init_recursive(&dp->non_pmd_mutex);
598 ovsthread_key_create(&dp->per_pmd_key, NULL);
600 /* Reserves the core NON_PMD_CORE_ID for all non-pmd threads. */
601 ovs_numa_try_pin_core_specific(NON_PMD_CORE_ID);
602 dp_netdev_set_nonpmd(dp);
603 dp->n_dpdk_rxqs = NR_QUEUE;
605 ovs_mutex_lock(&dp->port_mutex);
606 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
607 ovs_mutex_unlock(&dp->port_mutex);
618 dpif_netdev_open(const struct dpif_class *class, const char *name,
619 bool create, struct dpif **dpifp)
621 struct dp_netdev *dp;
624 ovs_mutex_lock(&dp_netdev_mutex);
625 dp = shash_find_data(&dp_netdevs, name);
627 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
629 error = (dp->class != class ? EINVAL
634 *dpifp = create_dpif_netdev(dp);
637 ovs_mutex_unlock(&dp_netdev_mutex);
643 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
644 OVS_NO_THREAD_SAFETY_ANALYSIS
646 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
647 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
649 /* Before freeing a lock we should release it */
650 fat_rwlock_unlock(&dp->upcall_rwlock);
651 fat_rwlock_destroy(&dp->upcall_rwlock);
654 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
655 * through the 'dp_netdevs' shash while freeing 'dp'. */
657 dp_netdev_free(struct dp_netdev *dp)
658 OVS_REQUIRES(dp_netdev_mutex)
660 struct dp_netdev_port *port;
661 struct dp_netdev_stats *bucket;
664 shash_find_and_delete(&dp_netdevs, dp->name);
666 dp_netdev_destroy_all_pmds(dp);
667 cmap_destroy(&dp->poll_threads);
668 ovs_mutex_destroy(&dp->non_pmd_mutex);
669 ovsthread_key_delete(dp->per_pmd_key);
671 dp_netdev_flow_flush(dp);
672 ovs_mutex_lock(&dp->port_mutex);
673 CMAP_FOR_EACH (port, node, &dp->ports) {
674 do_del_port(dp, port);
676 ovs_mutex_unlock(&dp->port_mutex);
678 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
679 ovs_mutex_destroy(&bucket->mutex);
680 free_cacheline(bucket);
682 ovsthread_stats_destroy(&dp->stats);
684 dpcls_destroy(&dp->cls);
685 cmap_destroy(&dp->flow_table);
686 ovs_mutex_destroy(&dp->flow_mutex);
687 seq_destroy(dp->port_seq);
688 cmap_destroy(&dp->ports);
690 /* Upcalls must be disabled at this point */
691 dp_netdev_destroy_upcall_lock(dp);
694 free(CONST_CAST(char *, dp->name));
699 dp_netdev_unref(struct dp_netdev *dp)
702 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
703 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
704 ovs_mutex_lock(&dp_netdev_mutex);
705 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
708 ovs_mutex_unlock(&dp_netdev_mutex);
713 dpif_netdev_close(struct dpif *dpif)
715 struct dp_netdev *dp = get_dp_netdev(dpif);
722 dpif_netdev_destroy(struct dpif *dpif)
724 struct dp_netdev *dp = get_dp_netdev(dpif);
726 if (!atomic_flag_test_and_set(&dp->destroyed)) {
727 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
728 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
737 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
739 struct dp_netdev *dp = get_dp_netdev(dpif);
740 struct dp_netdev_stats *bucket;
743 stats->n_flows = cmap_count(&dp->flow_table);
745 stats->n_hit = stats->n_missed = stats->n_lost = 0;
746 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
747 ovs_mutex_lock(&bucket->mutex);
748 stats->n_hit += bucket->n[DP_STAT_HIT];
749 stats->n_missed += bucket->n[DP_STAT_MISS];
750 stats->n_lost += bucket->n[DP_STAT_LOST];
751 ovs_mutex_unlock(&bucket->mutex);
753 stats->n_masks = UINT32_MAX;
754 stats->n_mask_hit = UINT64_MAX;
760 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
764 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
767 /* Causes all pmd threads to reload its tx/rx devices.
768 * Must be called after adding/removing ports. */
770 dp_netdev_reload_pmds(struct dp_netdev *dp)
772 struct dp_netdev_pmd_thread *pmd;
774 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
775 dp_netdev_reload_pmd__(pmd);
780 hash_port_no(odp_port_t port_no)
782 return hash_int(odp_to_u32(port_no), 0);
786 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
788 OVS_REQUIRES(dp->port_mutex)
790 struct netdev_saved_flags *sf;
791 struct dp_netdev_port *port;
792 struct netdev *netdev;
793 enum netdev_flags flags;
794 const char *open_type;
798 /* XXX reject devices already in some dp_netdev. */
800 /* Open and validate network device. */
801 open_type = dpif_netdev_port_open_type(dp->class, type);
802 error = netdev_open(devname, open_type, &netdev);
806 /* XXX reject non-Ethernet devices */
808 netdev_get_flags(netdev, &flags);
809 if (flags & NETDEV_LOOPBACK) {
810 VLOG_ERR("%s: cannot add a loopback device", devname);
811 netdev_close(netdev);
815 if (netdev_is_pmd(netdev)) {
816 int n_cores = ovs_numa_get_n_cores();
818 if (n_cores == OVS_CORE_UNSPEC) {
819 VLOG_ERR("%s, cannot get cpu core info", devname);
822 /* There can only be ovs_numa_get_n_cores() pmd threads,
823 * so creates a txq for each. */
824 error = netdev_set_multiq(netdev, n_cores, dp->n_dpdk_rxqs);
825 if (error && (error != EOPNOTSUPP)) {
826 VLOG_ERR("%s, cannot set multiq", devname);
830 port = xzalloc(sizeof *port);
831 port->port_no = port_no;
832 port->netdev = netdev;
833 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
834 port->type = xstrdup(type);
835 for (i = 0; i < netdev_n_rxq(netdev); i++) {
836 error = netdev_rxq_open(netdev, &port->rxq[i], i);
838 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
839 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
840 devname, ovs_strerror(errno));
841 netdev_close(netdev);
849 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
851 for (i = 0; i < netdev_n_rxq(netdev); i++) {
852 netdev_rxq_close(port->rxq[i]);
854 netdev_close(netdev);
862 if (netdev_is_pmd(netdev)) {
863 dp_netdev_set_pmds_on_numa(dp, netdev_get_numa_id(netdev));
864 dp_netdev_reload_pmds(dp);
866 ovs_refcount_init(&port->ref_cnt);
868 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
869 seq_change(dp->port_seq);
875 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
876 odp_port_t *port_nop)
878 struct dp_netdev *dp = get_dp_netdev(dpif);
879 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
880 const char *dpif_port;
884 ovs_mutex_lock(&dp->port_mutex);
885 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
886 if (*port_nop != ODPP_NONE) {
888 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
890 port_no = choose_port(dp, dpif_port);
891 error = port_no == ODPP_NONE ? EFBIG : 0;
895 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
897 ovs_mutex_unlock(&dp->port_mutex);
903 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
905 struct dp_netdev *dp = get_dp_netdev(dpif);
908 ovs_mutex_lock(&dp->port_mutex);
909 if (port_no == ODPP_LOCAL) {
912 struct dp_netdev_port *port;
914 error = get_port_by_number(dp, port_no, &port);
916 do_del_port(dp, port);
919 ovs_mutex_unlock(&dp->port_mutex);
925 is_valid_port_number(odp_port_t port_no)
927 return port_no != ODPP_NONE;
930 static struct dp_netdev_port *
931 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
933 struct dp_netdev_port *port;
935 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
936 if (port->port_no == port_no) {
944 get_port_by_number(struct dp_netdev *dp,
945 odp_port_t port_no, struct dp_netdev_port **portp)
947 if (!is_valid_port_number(port_no)) {
951 *portp = dp_netdev_lookup_port(dp, port_no);
952 return *portp ? 0 : ENOENT;
957 port_ref(struct dp_netdev_port *port)
960 ovs_refcount_ref(&port->ref_cnt);
965 port_try_ref(struct dp_netdev_port *port)
968 return ovs_refcount_try_ref_rcu(&port->ref_cnt);
975 port_destroy__(struct dp_netdev_port *port)
977 int n_rxq = netdev_n_rxq(port->netdev);
980 netdev_close(port->netdev);
981 netdev_restore_flags(port->sf);
983 for (i = 0; i < n_rxq; i++) {
984 netdev_rxq_close(port->rxq[i]);
992 port_unref(struct dp_netdev_port *port)
994 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
995 ovsrcu_postpone(port_destroy__, port);
1000 get_port_by_name(struct dp_netdev *dp,
1001 const char *devname, struct dp_netdev_port **portp)
1002 OVS_REQUIRES(dp->port_mutex)
1004 struct dp_netdev_port *port;
1006 CMAP_FOR_EACH (port, node, &dp->ports) {
1007 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1016 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1018 struct dp_netdev_pmd_thread *pmd;
1021 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1022 if (pmd->numa_id == numa_id) {
1030 /* Returns 'true' if there is a port with pmd netdev and the netdev
1031 * is on numa node 'numa_id'. */
1033 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1035 struct dp_netdev_port *port;
1037 CMAP_FOR_EACH (port, node, &dp->ports) {
1038 if (netdev_is_pmd(port->netdev)
1039 && netdev_get_numa_id(port->netdev) == numa_id) {
1049 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1050 OVS_REQUIRES(dp->port_mutex)
1052 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
1053 seq_change(dp->port_seq);
1054 if (netdev_is_pmd(port->netdev)) {
1055 int numa_id = netdev_get_numa_id(port->netdev);
1057 /* If there is no netdev on the numa node, deletes the pmd threads
1058 * for that numa. Else, just reloads the queues. */
1059 if (!has_pmd_port_for_numa(dp, numa_id)) {
1060 dp_netdev_del_pmds_on_numa(dp, numa_id);
1062 dp_netdev_reload_pmds(dp);
1069 answer_port_query(const struct dp_netdev_port *port,
1070 struct dpif_port *dpif_port)
1072 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1073 dpif_port->type = xstrdup(port->type);
1074 dpif_port->port_no = port->port_no;
1078 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1079 struct dpif_port *dpif_port)
1081 struct dp_netdev *dp = get_dp_netdev(dpif);
1082 struct dp_netdev_port *port;
1085 error = get_port_by_number(dp, port_no, &port);
1086 if (!error && dpif_port) {
1087 answer_port_query(port, dpif_port);
1094 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1095 struct dpif_port *dpif_port)
1097 struct dp_netdev *dp = get_dp_netdev(dpif);
1098 struct dp_netdev_port *port;
1101 ovs_mutex_lock(&dp->port_mutex);
1102 error = get_port_by_name(dp, devname, &port);
1103 if (!error && dpif_port) {
1104 answer_port_query(port, dpif_port);
1106 ovs_mutex_unlock(&dp->port_mutex);
1112 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1114 struct dp_netdev_flow_stats *bucket;
1117 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
1118 ovs_mutex_destroy(&bucket->mutex);
1119 free_cacheline(bucket);
1121 ovsthread_stats_destroy(&flow->stats);
1123 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1127 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1129 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1130 ovsrcu_postpone(dp_netdev_flow_free, flow);
1135 dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
1136 OVS_REQUIRES(dp->flow_mutex)
1138 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1140 dpcls_remove(&dp->cls, &flow->cr);
1141 cmap_remove(&dp->flow_table, node, flow_hash(&flow->flow, 0));
1144 dp_netdev_flow_unref(flow);
1148 dp_netdev_flow_flush(struct dp_netdev *dp)
1150 struct dp_netdev_flow *netdev_flow;
1152 ovs_mutex_lock(&dp->flow_mutex);
1153 CMAP_FOR_EACH (netdev_flow, node, &dp->flow_table) {
1154 dp_netdev_remove_flow(dp, netdev_flow);
1156 ovs_mutex_unlock(&dp->flow_mutex);
1160 dpif_netdev_flow_flush(struct dpif *dpif)
1162 struct dp_netdev *dp = get_dp_netdev(dpif);
1164 dp_netdev_flow_flush(dp);
1168 struct dp_netdev_port_state {
1169 struct cmap_position position;
1174 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1176 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1181 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1182 struct dpif_port *dpif_port)
1184 struct dp_netdev_port_state *state = state_;
1185 struct dp_netdev *dp = get_dp_netdev(dpif);
1186 struct cmap_node *node;
1189 node = cmap_next_position(&dp->ports, &state->position);
1191 struct dp_netdev_port *port;
1193 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1196 state->name = xstrdup(netdev_get_name(port->netdev));
1197 dpif_port->name = state->name;
1198 dpif_port->type = port->type;
1199 dpif_port->port_no = port->port_no;
1210 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1212 struct dp_netdev_port_state *state = state_;
1219 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1221 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1222 uint64_t new_port_seq;
1225 new_port_seq = seq_read(dpif->dp->port_seq);
1226 if (dpif->last_port_seq != new_port_seq) {
1227 dpif->last_port_seq = new_port_seq;
1237 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1239 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1241 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1244 static struct dp_netdev_flow *
1245 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1247 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1250 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1252 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1255 /* netdev_flow_key utilities.
1257 * netdev_flow_key is basically a miniflow. We use these functions
1258 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1259 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1261 * - Since we are dealing exclusively with miniflows created by
1262 * miniflow_extract(), if the map is different the miniflow is different.
1263 * Therefore we can be faster by comparing the map and the miniflow in a
1265 * _ netdev_flow_key's miniflow has always inline values.
1266 * - These functions can be inlined by the compiler.
1268 * The following assertions make sure that what we're doing with miniflow is
1271 BUILD_ASSERT_DECL(offsetof(struct miniflow, inline_values)
1272 == sizeof(uint64_t));
1274 /* Given the number of bits set in the miniflow map, returns the size of the
1275 * 'netdev_flow_key.mf' */
1276 static inline uint32_t
1277 netdev_flow_key_size(uint32_t flow_u32s)
1279 return offsetof(struct miniflow, inline_values) +
1280 MINIFLOW_VALUES_SIZE(flow_u32s);
1284 netdev_flow_key_equal(const struct netdev_flow_key *a,
1285 const struct netdev_flow_key *b)
1287 /* 'b->len' may be not set yet. */
1288 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1291 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1292 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1293 * generated by miniflow_extract. */
1295 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1296 const struct miniflow *mf)
1298 return !memcmp(&key->mf, mf, key->len);
1302 netdev_flow_key_clone(struct netdev_flow_key *dst,
1303 const struct netdev_flow_key *src)
1306 offsetof(struct netdev_flow_key, mf) + src->len);
1311 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1312 const struct flow *src)
1314 struct ofpbuf packet;
1315 uint64_t buf_stub[512 / 8];
1316 struct pkt_metadata md = pkt_metadata_from_flow(src);
1318 miniflow_initialize(&dst->mf, dst->buf);
1320 ofpbuf_use_stub(&packet, buf_stub, sizeof buf_stub);
1321 flow_compose(&packet, src);
1322 miniflow_extract(&packet, &md, &dst->mf);
1323 ofpbuf_uninit(&packet);
1325 dst->len = netdev_flow_key_size(count_1bits(dst->mf.map));
1326 dst->hash = 0; /* Not computed yet. */
1329 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1331 netdev_flow_mask_init(struct netdev_flow_key *mask,
1332 const struct match *match)
1334 const uint32_t *mask_u32 = (const uint32_t *) &match->wc.masks;
1335 uint32_t *dst = mask->mf.inline_values;
1336 uint64_t map, mask_map = 0;
1340 /* Only check masks that make sense for the flow. */
1341 map = flow_wc_map(&match->flow);
1344 uint64_t rm1bit = rightmost_1bit(map);
1345 int i = raw_ctz(map);
1349 *dst++ = mask_u32[i];
1350 hash = hash_add(hash, mask_u32[i]);
1355 mask->mf.values_inline = true;
1356 mask->mf.map = mask_map;
1358 hash = hash_add(hash, mask_map);
1359 hash = hash_add(hash, mask_map >> 32);
1361 n = dst - mask->mf.inline_values;
1363 mask->hash = hash_finish(hash, n * 4);
1364 mask->len = netdev_flow_key_size(n);
1367 /* Initializes 'dst' as a copy of 'src' masked with 'mask'. */
1369 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1370 const struct flow *flow,
1371 const struct netdev_flow_key *mask)
1373 uint32_t *dst_u32 = dst->mf.inline_values;
1374 const uint32_t *mask_u32 = mask->mf.inline_values;
1378 dst->len = mask->len;
1379 dst->mf.values_inline = true;
1380 dst->mf.map = mask->mf.map;
1382 FLOW_FOR_EACH_IN_MAP(value, flow, mask->mf.map) {
1383 *dst_u32 = value & *mask_u32++;
1384 hash = hash_add(hash, *dst_u32++);
1386 dst->hash = hash_finish(hash, (dst_u32 - dst->mf.inline_values) * 4);
1389 /* Iterate through all netdev_flow_key u32 values specified by 'MAP' */
1390 #define NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(VALUE, KEY, MAP) \
1391 for (struct mf_for_each_in_map_aux aux__ \
1392 = { (KEY)->mf.inline_values, (KEY)->mf.map, MAP }; \
1393 mf_get_next_in_map(&aux__, &(VALUE)); \
1396 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1398 static inline uint32_t
1399 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1400 const struct netdev_flow_key *mask)
1402 const uint32_t *p = mask->mf.inline_values;
1406 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(key_u32, key, mask->mf.map) {
1407 hash = hash_add(hash, key_u32 & *p++);
1410 return hash_finish(hash, (p - mask->mf.inline_values) * 4);
1414 emc_entry_alive(struct emc_entry *ce)
1416 return ce->flow && !ce->flow->dead;
1420 emc_clear_entry(struct emc_entry *ce)
1423 dp_netdev_flow_unref(ce->flow);
1429 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1430 const struct netdev_flow_key *key)
1432 if (ce->flow != flow) {
1434 dp_netdev_flow_unref(ce->flow);
1437 if (dp_netdev_flow_ref(flow)) {
1444 netdev_flow_key_clone(&ce->key, key);
1449 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1450 struct dp_netdev_flow *flow)
1452 struct emc_entry *to_be_replaced = NULL;
1453 struct emc_entry *current_entry;
1455 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1456 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1457 /* We found the entry with the 'mf' miniflow */
1458 emc_change_entry(current_entry, flow, NULL);
1462 /* Replacement policy: put the flow in an empty (not alive) entry, or
1463 * in the first entry where it can be */
1465 || (emc_entry_alive(to_be_replaced)
1466 && !emc_entry_alive(current_entry))
1467 || current_entry->key.hash < to_be_replaced->key.hash) {
1468 to_be_replaced = current_entry;
1471 /* We didn't find the miniflow in the cache.
1472 * The 'to_be_replaced' entry is where the new flow will be stored */
1474 emc_change_entry(to_be_replaced, flow, key);
1477 static inline struct dp_netdev_flow *
1478 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1480 struct emc_entry *current_entry;
1482 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1483 if (current_entry->key.hash == key->hash
1484 && emc_entry_alive(current_entry)
1485 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1487 /* We found the entry with the 'key->mf' miniflow */
1488 return current_entry->flow;
1495 static struct dp_netdev_flow *
1496 dp_netdev_lookup_flow(const struct dp_netdev *dp,
1497 const struct netdev_flow_key *key)
1499 struct dp_netdev_flow *netdev_flow;
1500 struct dpcls_rule *rule;
1502 dpcls_lookup(&dp->cls, key, &rule, 1);
1503 netdev_flow = dp_netdev_flow_cast(rule);
1508 static struct dp_netdev_flow *
1509 dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
1511 struct dp_netdev_flow *netdev_flow;
1513 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
1515 if (flow_equal(&netdev_flow->flow, flow)) {
1524 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow,
1525 struct dpif_flow_stats *stats)
1527 struct dp_netdev_flow_stats *bucket;
1530 memset(stats, 0, sizeof *stats);
1531 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1532 ovs_mutex_lock(&bucket->mutex);
1533 stats->n_packets += bucket->packet_count;
1534 stats->n_bytes += bucket->byte_count;
1535 stats->used = MAX(stats->used, bucket->used);
1536 stats->tcp_flags |= bucket->tcp_flags;
1537 ovs_mutex_unlock(&bucket->mutex);
1542 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1543 struct ofpbuf *buffer, struct dpif_flow *flow)
1545 struct flow_wildcards wc;
1546 struct dp_netdev_actions *actions;
1548 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1549 odp_flow_key_from_mask(buffer, &wc.masks, &netdev_flow->flow,
1550 odp_to_u32(wc.masks.in_port.odp_port),
1552 flow->mask = ofpbuf_data(buffer);
1553 flow->mask_len = ofpbuf_size(buffer);
1555 actions = dp_netdev_flow_get_actions(netdev_flow);
1556 flow->actions = actions->actions;
1557 flow->actions_len = actions->size;
1559 get_dpif_flow_stats(netdev_flow, &flow->stats);
1563 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1564 const struct nlattr *mask_key,
1565 uint32_t mask_key_len, const struct flow *flow,
1569 enum odp_key_fitness fitness;
1571 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1573 /* This should not happen: it indicates that
1574 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1575 * disagree on the acceptable form of a mask. Log the problem
1576 * as an error, with enough details to enable debugging. */
1577 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1579 if (!VLOG_DROP_ERR(&rl)) {
1583 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1585 VLOG_ERR("internal error parsing flow mask %s (%s)",
1586 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1593 enum mf_field_id id;
1594 /* No mask key, unwildcard everything except fields whose
1595 * prerequisities are not met. */
1596 memset(mask, 0x0, sizeof *mask);
1598 for (id = 0; id < MFF_N_IDS; ++id) {
1599 /* Skip registers and metadata. */
1600 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1601 && id != MFF_METADATA) {
1602 const struct mf_field *mf = mf_from_id(id);
1603 if (mf_are_prereqs_ok(mf, flow)) {
1604 mf_mask_field(mf, mask);
1610 /* Force unwildcard the in_port.
1612 * We need to do this even in the case where we unwildcard "everything"
1613 * above because "everything" only includes the 16-bit OpenFlow port number
1614 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1615 * port number mask->in_port.odp_port. */
1616 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1622 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1627 if (odp_flow_key_to_flow(key, key_len, flow)) {
1628 /* This should not happen: it indicates that odp_flow_key_from_flow()
1629 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1630 * flow. Log the problem as an error, with enough details to enable
1632 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1634 if (!VLOG_DROP_ERR(&rl)) {
1638 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1639 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1646 in_port = flow->in_port.odp_port;
1647 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1655 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1657 struct dp_netdev *dp = get_dp_netdev(dpif);
1658 struct dp_netdev_flow *netdev_flow;
1662 error = dpif_netdev_flow_from_nlattrs(get->key, get->key_len, &key);
1667 netdev_flow = dp_netdev_find_flow(dp, &key);
1670 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->flow);
1678 static struct dp_netdev_flow *
1679 dp_netdev_flow_add(struct dp_netdev *dp, struct match *match,
1680 const struct nlattr *actions, size_t actions_len)
1681 OVS_REQUIRES(dp->flow_mutex)
1683 struct dp_netdev_flow *flow;
1684 struct netdev_flow_key mask;
1686 netdev_flow_mask_init(&mask, match);
1687 /* Make sure wc does not have metadata. */
1688 ovs_assert(!(mask.mf.map & (MINIFLOW_MAP(metadata) | MINIFLOW_MAP(regs))));
1690 /* Do not allocate extra space. */
1691 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
1693 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
1694 ovs_refcount_init(&flow->ref_cnt);
1695 ovsthread_stats_init(&flow->stats);
1696 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
1698 cmap_insert(&dp->flow_table,
1699 CONST_CAST(struct cmap_node *, &flow->node),
1700 flow_hash(&flow->flow, 0));
1701 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
1702 dpcls_insert(&dp->cls, &flow->cr, &mask);
1704 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
1706 struct ds ds = DS_EMPTY_INITIALIZER;
1708 match.flow = flow->flow;
1709 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
1711 ds_put_cstr(&ds, "flow_add: ");
1712 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
1713 ds_put_cstr(&ds, ", actions:");
1714 format_odp_actions(&ds, actions, actions_len);
1716 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
1725 clear_stats(struct dp_netdev_flow *netdev_flow)
1727 struct dp_netdev_flow_stats *bucket;
1730 OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
1731 ovs_mutex_lock(&bucket->mutex);
1733 bucket->packet_count = 0;
1734 bucket->byte_count = 0;
1735 bucket->tcp_flags = 0;
1736 ovs_mutex_unlock(&bucket->mutex);
1741 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1743 struct dp_netdev *dp = get_dp_netdev(dpif);
1744 struct dp_netdev_flow *netdev_flow;
1745 struct netdev_flow_key key;
1749 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
1753 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1754 put->mask, put->mask_len,
1755 &match.flow, &match.wc.masks);
1760 /* Must produce a netdev_flow_key for lookup.
1761 * This interface is no longer performance critical, since it is not used
1762 * for upcall processing any more. */
1763 netdev_flow_key_from_flow(&key, &match.flow);
1765 ovs_mutex_lock(&dp->flow_mutex);
1766 netdev_flow = dp_netdev_lookup_flow(dp, &key);
1768 if (put->flags & DPIF_FP_CREATE) {
1769 if (cmap_count(&dp->flow_table) < MAX_FLOWS) {
1771 memset(put->stats, 0, sizeof *put->stats);
1773 dp_netdev_flow_add(dp, &match, put->actions, put->actions_len);
1782 if (put->flags & DPIF_FP_MODIFY
1783 && flow_equal(&match.flow, &netdev_flow->flow)) {
1784 struct dp_netdev_actions *new_actions;
1785 struct dp_netdev_actions *old_actions;
1787 new_actions = dp_netdev_actions_create(put->actions,
1790 old_actions = dp_netdev_flow_get_actions(netdev_flow);
1791 ovsrcu_set(&netdev_flow->actions, new_actions);
1794 get_dpif_flow_stats(netdev_flow, put->stats);
1796 if (put->flags & DPIF_FP_ZERO_STATS) {
1797 clear_stats(netdev_flow);
1800 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
1801 } else if (put->flags & DPIF_FP_CREATE) {
1804 /* Overlapping flow. */
1808 ovs_mutex_unlock(&dp->flow_mutex);
1814 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
1816 struct dp_netdev *dp = get_dp_netdev(dpif);
1817 struct dp_netdev_flow *netdev_flow;
1821 error = dpif_netdev_flow_from_nlattrs(del->key, del->key_len, &key);
1826 ovs_mutex_lock(&dp->flow_mutex);
1827 netdev_flow = dp_netdev_find_flow(dp, &key);
1830 get_dpif_flow_stats(netdev_flow, del->stats);
1832 dp_netdev_remove_flow(dp, netdev_flow);
1836 ovs_mutex_unlock(&dp->flow_mutex);
1841 struct dpif_netdev_flow_dump {
1842 struct dpif_flow_dump up;
1843 struct cmap_position pos;
1845 struct ovs_mutex mutex;
1848 static struct dpif_netdev_flow_dump *
1849 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
1851 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
1854 static struct dpif_flow_dump *
1855 dpif_netdev_flow_dump_create(const struct dpif *dpif_)
1857 struct dpif_netdev_flow_dump *dump;
1859 dump = xmalloc(sizeof *dump);
1860 dpif_flow_dump_init(&dump->up, dpif_);
1861 memset(&dump->pos, 0, sizeof dump->pos);
1863 ovs_mutex_init(&dump->mutex);
1869 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
1871 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1873 ovs_mutex_destroy(&dump->mutex);
1878 struct dpif_netdev_flow_dump_thread {
1879 struct dpif_flow_dump_thread up;
1880 struct dpif_netdev_flow_dump *dump;
1881 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
1882 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
1885 static struct dpif_netdev_flow_dump_thread *
1886 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
1888 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
1891 static struct dpif_flow_dump_thread *
1892 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
1894 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1895 struct dpif_netdev_flow_dump_thread *thread;
1897 thread = xmalloc(sizeof *thread);
1898 dpif_flow_dump_thread_init(&thread->up, &dump->up);
1899 thread->dump = dump;
1904 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
1906 struct dpif_netdev_flow_dump_thread *thread
1907 = dpif_netdev_flow_dump_thread_cast(thread_);
1913 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
1914 struct dpif_flow *flows, int max_flows)
1916 struct dpif_netdev_flow_dump_thread *thread
1917 = dpif_netdev_flow_dump_thread_cast(thread_);
1918 struct dpif_netdev_flow_dump *dump = thread->dump;
1919 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
1920 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
1921 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
1925 ovs_mutex_lock(&dump->mutex);
1926 if (!dump->status) {
1927 for (n_flows = 0; n_flows < MIN(max_flows, FLOW_DUMP_MAX_BATCH);
1929 struct cmap_node *node;
1931 node = cmap_next_position(&dp->flow_table, &dump->pos);
1936 netdev_flows[n_flows] = CONTAINER_OF(node, struct dp_netdev_flow,
1940 ovs_mutex_unlock(&dump->mutex);
1942 for (i = 0; i < n_flows; i++) {
1943 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
1944 struct odputil_keybuf *keybuf = &thread->keybuf[i];
1945 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
1946 struct dpif_flow *f = &flows[i];
1947 struct dp_netdev_actions *dp_actions;
1948 struct flow_wildcards wc;
1951 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1954 ofpbuf_use_stack(&buf, keybuf, sizeof *keybuf);
1955 odp_flow_key_from_flow(&buf, &netdev_flow->flow, &wc.masks,
1956 netdev_flow->flow.in_port.odp_port, true);
1957 f->key = ofpbuf_data(&buf);
1958 f->key_len = ofpbuf_size(&buf);
1961 ofpbuf_use_stack(&buf, maskbuf, sizeof *maskbuf);
1962 odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
1963 odp_to_u32(wc.masks.in_port.odp_port),
1965 f->mask = ofpbuf_data(&buf);
1966 f->mask_len = ofpbuf_size(&buf);
1969 dp_actions = dp_netdev_flow_get_actions(netdev_flow);
1970 f->actions = dp_actions->actions;
1971 f->actions_len = dp_actions->size;
1974 get_dpif_flow_stats(netdev_flow, &f->stats);
1981 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
1982 OVS_NO_THREAD_SAFETY_ANALYSIS
1984 struct dp_netdev *dp = get_dp_netdev(dpif);
1985 struct dp_netdev_pmd_thread *pmd;
1986 struct dpif_packet packet, *pp;
1988 if (ofpbuf_size(execute->packet) < ETH_HEADER_LEN ||
1989 ofpbuf_size(execute->packet) > UINT16_MAX) {
1993 packet.ofpbuf = *execute->packet;
1994 packet.md = execute->md;
1997 /* Tries finding the 'pmd'. If NULL is returned, that means
1998 * the current thread is a non-pmd thread and should use
1999 * dp_netdev_get_nonpmd(). */
2000 pmd = ovsthread_getspecific(dp->per_pmd_key);
2002 pmd = dp_netdev_get_nonpmd(dp);
2005 /* If the current thread is non-pmd thread, acquires
2006 * the 'non_pmd_mutex'. */
2007 if (pmd->core_id == NON_PMD_CORE_ID) {
2008 ovs_mutex_lock(&dp->non_pmd_mutex);
2010 dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
2011 execute->actions_len);
2012 if (pmd->core_id == NON_PMD_CORE_ID) {
2013 ovs_mutex_unlock(&dp->non_pmd_mutex);
2016 /* Even though may_steal is set to false, some actions could modify or
2017 * reallocate the ofpbuf memory. We need to pass those changes to the
2019 *execute->packet = packet.ofpbuf;
2020 execute->md = packet.md;
2025 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2029 for (i = 0; i < n_ops; i++) {
2030 struct dpif_op *op = ops[i];
2033 case DPIF_OP_FLOW_PUT:
2034 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2037 case DPIF_OP_FLOW_DEL:
2038 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2041 case DPIF_OP_EXECUTE:
2042 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2045 case DPIF_OP_FLOW_GET:
2046 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2052 /* Returns true if the configuration for rx queues or cpu mask
2055 pmd_config_changed(const struct dp_netdev *dp, size_t rxqs, const char *cmask)
2057 if (dp->n_dpdk_rxqs != rxqs) {
2060 if (dp->pmd_cmask != NULL && cmask != NULL) {
2061 return strcmp(dp->pmd_cmask, cmask);
2063 return (dp->pmd_cmask != NULL || cmask != NULL);
2068 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2070 dpif_netdev_pmd_set(struct dpif *dpif, unsigned int n_rxqs, const char *cmask)
2072 struct dp_netdev *dp = get_dp_netdev(dpif);
2074 if (pmd_config_changed(dp, n_rxqs, cmask)) {
2075 struct dp_netdev_port *port;
2077 dp_netdev_destroy_all_pmds(dp);
2079 CMAP_FOR_EACH (port, node, &dp->ports) {
2080 if (netdev_is_pmd(port->netdev)) {
2083 /* Closes the existing 'rxq's. */
2084 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2085 netdev_rxq_close(port->rxq[i]);
2086 port->rxq[i] = NULL;
2089 /* Sets the new rx queue config. */
2090 err = netdev_set_multiq(port->netdev, ovs_numa_get_n_cores(),
2092 if (err && (err != EOPNOTSUPP)) {
2093 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2094 " %u", netdev_get_name(port->netdev),
2099 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2100 port->rxq = xrealloc(port->rxq, sizeof *port->rxq
2101 * netdev_n_rxq(port->netdev));
2102 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2103 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2107 dp->n_dpdk_rxqs = n_rxqs;
2109 /* Reconfigures the cpu mask. */
2110 ovs_numa_set_cpu_mask(cmask);
2111 free(dp->pmd_cmask);
2112 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2114 /* Restores the non-pmd. */
2115 dp_netdev_set_nonpmd(dp);
2116 /* Restores all pmd threads. */
2117 dp_netdev_reset_pmd_threads(dp);
2124 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2125 uint32_t queue_id, uint32_t *priority)
2127 *priority = queue_id;
2132 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
2133 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
2135 struct dp_netdev_actions *
2136 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2138 struct dp_netdev_actions *netdev_actions;
2140 netdev_actions = xmalloc(sizeof *netdev_actions);
2141 netdev_actions->actions = xmemdup(actions, size);
2142 netdev_actions->size = size;
2144 return netdev_actions;
2147 struct dp_netdev_actions *
2148 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2150 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2154 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2156 free(actions->actions);
2162 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2163 struct dp_netdev_port *port,
2164 struct netdev_rxq *rxq)
2166 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
2169 error = netdev_rxq_recv(rxq, packets, &cnt);
2173 *recirc_depth_get() = 0;
2175 /* XXX: initialize md in netdev implementation. */
2176 for (i = 0; i < cnt; i++) {
2177 packets[i]->md = PKT_METADATA_INITIALIZER(port->port_no);
2179 dp_netdev_input(pmd, packets, cnt);
2180 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2181 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2183 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2184 netdev_get_name(port->netdev), ovs_strerror(error));
2189 dpif_netdev_run(struct dpif *dpif)
2191 struct dp_netdev_port *port;
2192 struct dp_netdev *dp = get_dp_netdev(dpif);
2193 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_nonpmd(dp);
2195 ovs_mutex_lock(&dp->non_pmd_mutex);
2196 CMAP_FOR_EACH (port, node, &dp->ports) {
2197 if (!netdev_is_pmd(port->netdev)) {
2200 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2201 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2205 ovs_mutex_unlock(&dp->non_pmd_mutex);
2209 dpif_netdev_wait(struct dpif *dpif)
2211 struct dp_netdev_port *port;
2212 struct dp_netdev *dp = get_dp_netdev(dpif);
2214 ovs_mutex_lock(&dp_netdev_mutex);
2215 CMAP_FOR_EACH (port, node, &dp->ports) {
2216 if (!netdev_is_pmd(port->netdev)) {
2219 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2220 netdev_rxq_wait(port->rxq[i]);
2224 ovs_mutex_unlock(&dp_netdev_mutex);
2228 struct dp_netdev_port *port;
2229 struct netdev_rxq *rx;
2233 pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
2234 struct rxq_poll **ppoll_list, int poll_cnt)
2236 struct rxq_poll *poll_list = *ppoll_list;
2237 struct dp_netdev_port *port;
2238 int n_pmds_on_numa, index, i;
2240 /* Simple scheduler for netdev rx polling. */
2241 for (i = 0; i < poll_cnt; i++) {
2242 port_unref(poll_list[i].port);
2246 n_pmds_on_numa = get_n_pmd_threads_on_numa(pmd->dp, pmd->numa_id);
2249 CMAP_FOR_EACH (port, node, &pmd->dp->ports) {
2250 /* Calls port_try_ref() to prevent the main thread
2251 * from deleting the port. */
2252 if (port_try_ref(port)) {
2253 if (netdev_is_pmd(port->netdev)
2254 && netdev_get_numa_id(port->netdev) == pmd->numa_id) {
2257 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2258 if ((index % n_pmds_on_numa) == pmd->index) {
2259 poll_list = xrealloc(poll_list,
2260 sizeof *poll_list * (poll_cnt + 1));
2263 poll_list[poll_cnt].port = port;
2264 poll_list[poll_cnt].rx = port->rxq[i];
2270 /* Unrefs the port_try_ref(). */
2275 *ppoll_list = poll_list;
2280 pmd_thread_main(void *f_)
2282 struct dp_netdev_pmd_thread *pmd = f_;
2283 unsigned int lc = 0;
2284 struct rxq_poll *poll_list;
2285 unsigned int port_seq = PMD_INITIAL_SEQ;
2292 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2293 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2294 pmd_thread_setaffinity_cpu(pmd->core_id);
2296 emc_cache_init(&pmd->flow_cache);
2297 poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
2302 for (i = 0; i < poll_cnt; i++) {
2303 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2313 atomic_read_relaxed(&pmd->change_seq, &seq);
2314 if (seq != port_seq) {
2321 emc_cache_uninit(&pmd->flow_cache);
2323 if (!latch_is_set(&pmd->exit_latch)){
2327 for (i = 0; i < poll_cnt; i++) {
2328 port_unref(poll_list[i].port);
2336 dp_netdev_disable_upcall(struct dp_netdev *dp)
2337 OVS_ACQUIRES(dp->upcall_rwlock)
2339 fat_rwlock_wrlock(&dp->upcall_rwlock);
2343 dpif_netdev_disable_upcall(struct dpif *dpif)
2344 OVS_NO_THREAD_SAFETY_ANALYSIS
2346 struct dp_netdev *dp = get_dp_netdev(dpif);
2347 dp_netdev_disable_upcall(dp);
2351 dp_netdev_enable_upcall(struct dp_netdev *dp)
2352 OVS_RELEASES(dp->upcall_rwlock)
2354 fat_rwlock_unlock(&dp->upcall_rwlock);
2358 dpif_netdev_enable_upcall(struct dpif *dpif)
2359 OVS_NO_THREAD_SAFETY_ANALYSIS
2361 struct dp_netdev *dp = get_dp_netdev(dpif);
2362 dp_netdev_enable_upcall(dp);
2365 /* Returns the pointer to the dp_netdev_pmd_thread for non-pmd threads. */
2366 static struct dp_netdev_pmd_thread *
2367 dp_netdev_get_nonpmd(struct dp_netdev *dp)
2369 struct dp_netdev_pmd_thread *pmd;
2370 const struct cmap_node *pnode;
2372 pnode = cmap_find(&dp->poll_threads, hash_int(NON_PMD_CORE_ID, 0));
2374 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2379 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2381 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2383 struct dp_netdev_pmd_thread *non_pmd;
2385 non_pmd = xzalloc(sizeof *non_pmd);
2386 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2390 /* Configures the 'pmd' based on the input argument. */
2392 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2393 int index, int core_id, int numa_id)
2397 pmd->core_id = core_id;
2398 pmd->numa_id = numa_id;
2399 latch_init(&pmd->exit_latch);
2400 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2401 /* init the 'flow_cache' since there is no
2402 * actual thread created for NON_PMD_CORE_ID. */
2403 if (core_id == NON_PMD_CORE_ID) {
2404 emc_cache_init(&pmd->flow_cache);
2406 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2407 hash_int(core_id, 0));
2410 /* Stops the pmd thread, removes it from the 'dp->poll_threads'
2411 * and destroys the struct. */
2413 dp_netdev_del_pmd(struct dp_netdev_pmd_thread *pmd)
2415 /* Uninit the 'flow_cache' since there is
2416 * no actual thread uninit it. */
2417 if (pmd->core_id == NON_PMD_CORE_ID) {
2418 emc_cache_uninit(&pmd->flow_cache);
2420 latch_set(&pmd->exit_latch);
2421 dp_netdev_reload_pmd__(pmd);
2422 ovs_numa_unpin_core(pmd->core_id);
2423 xpthread_join(pmd->thread, NULL);
2425 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2426 latch_destroy(&pmd->exit_latch);
2430 /* Destroys all pmd threads. */
2432 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2434 struct dp_netdev_pmd_thread *pmd;
2436 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2437 dp_netdev_del_pmd(pmd);
2441 /* Deletes all pmd threads on numa node 'numa_id'. */
2443 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2445 struct dp_netdev_pmd_thread *pmd;
2447 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2448 if (pmd->numa_id == numa_id) {
2449 dp_netdev_del_pmd(pmd);
2454 /* Checks the numa node id of 'netdev' and starts pmd threads for
2457 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2461 if (!ovs_numa_numa_id_is_valid(numa_id)) {
2462 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
2463 "invalid", numa_id);
2467 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
2469 /* If there are already pmd threads created for the numa node
2470 * in which 'netdev' is on, do nothing. Else, creates the
2471 * pmd threads for the numa node. */
2473 int can_have, n_unpinned, i;
2475 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
2477 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
2478 "cores on numa node");
2482 /* If cpu mask is specified, uses all unpinned cores, otherwise
2483 * tries creating NR_PMD_THREADS pmd threads. */
2484 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
2485 for (i = 0; i < can_have; i++) {
2486 struct dp_netdev_pmd_thread *pmd = xzalloc(sizeof *pmd);
2487 int core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
2489 dp_netdev_configure_pmd(pmd, dp, i, core_id, numa_id);
2490 /* Each thread will distribute all devices rx-queues among
2492 pmd->thread = ovs_thread_create("pmd", pmd_thread_main, pmd);
2494 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
2500 dp_netdev_flow_stats_new_cb(void)
2502 struct dp_netdev_flow_stats *bucket = xzalloc_cacheline(sizeof *bucket);
2503 ovs_mutex_init(&bucket->mutex);
2507 /* Called after pmd threads config change. Restarts pmd threads with
2508 * new configuration. */
2510 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
2512 struct dp_netdev_port *port;
2514 CMAP_FOR_EACH (port, node, &dp->ports) {
2515 if (netdev_is_pmd(port->netdev)) {
2516 int numa_id = netdev_get_numa_id(port->netdev);
2518 dp_netdev_set_pmds_on_numa(dp, numa_id);
2524 dpif_netdev_get_datapath_version(void)
2526 return xstrdup("<built-in>");
2530 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow,
2534 long long int now = time_msec();
2535 struct dp_netdev_flow_stats *bucket;
2537 bucket = ovsthread_stats_bucket_get(&netdev_flow->stats,
2538 dp_netdev_flow_stats_new_cb);
2540 ovs_mutex_lock(&bucket->mutex);
2541 bucket->used = MAX(now, bucket->used);
2542 bucket->packet_count += cnt;
2543 bucket->byte_count += size;
2544 bucket->tcp_flags |= tcp_flags;
2545 ovs_mutex_unlock(&bucket->mutex);
2549 dp_netdev_stats_new_cb(void)
2551 struct dp_netdev_stats *bucket = xzalloc_cacheline(sizeof *bucket);
2552 ovs_mutex_init(&bucket->mutex);
2557 dp_netdev_count_packet(struct dp_netdev *dp, enum dp_stat_type type, int cnt)
2559 struct dp_netdev_stats *bucket;
2561 bucket = ovsthread_stats_bucket_get(&dp->stats, dp_netdev_stats_new_cb);
2562 ovs_mutex_lock(&bucket->mutex);
2563 bucket->n[type] += cnt;
2564 ovs_mutex_unlock(&bucket->mutex);
2568 dp_netdev_upcall(struct dp_netdev *dp, struct dpif_packet *packet_,
2569 struct flow *flow, struct flow_wildcards *wc,
2570 enum dpif_upcall_type type, const struct nlattr *userdata,
2571 struct ofpbuf *actions, struct ofpbuf *put_actions)
2573 struct ofpbuf *packet = &packet_->ofpbuf;
2575 if (type == DPIF_UC_MISS) {
2576 dp_netdev_count_packet(dp, DP_STAT_MISS, 1);
2579 if (OVS_UNLIKELY(!dp->upcall_cb)) {
2583 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
2584 struct ds ds = DS_EMPTY_INITIALIZER;
2588 ofpbuf_init(&key, 0);
2589 odp_flow_key_from_flow(&key, flow, &wc->masks, flow->in_port.odp_port,
2592 packet_str = ofp_packet_to_string(ofpbuf_data(packet),
2593 ofpbuf_size(packet));
2595 odp_flow_key_format(ofpbuf_data(&key), ofpbuf_size(&key), &ds);
2597 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
2598 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
2600 ofpbuf_uninit(&key);
2605 return dp->upcall_cb(packet, flow, type, userdata, actions, wc,
2606 put_actions, dp->upcall_aux);
2609 static inline uint32_t
2610 dpif_netdev_packet_get_dp_hash(struct dpif_packet *packet,
2611 const struct miniflow *mf)
2615 hash = dpif_packet_get_dp_hash(packet);
2616 if (OVS_UNLIKELY(!hash)) {
2617 hash = miniflow_hash_5tuple(mf, 0);
2618 dpif_packet_set_dp_hash(packet, hash);
2623 struct packet_batch {
2624 unsigned int packet_count;
2625 unsigned int byte_count;
2628 struct dp_netdev_flow *flow;
2630 struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
2634 packet_batch_update(struct packet_batch *batch, struct dpif_packet *packet,
2635 const struct miniflow *mf)
2637 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
2638 batch->packets[batch->packet_count++] = packet;
2639 batch->byte_count += ofpbuf_size(&packet->ofpbuf);
2643 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
2647 batch->packet_count = 0;
2648 batch->byte_count = 0;
2649 batch->tcp_flags = 0;
2653 packet_batch_execute(struct packet_batch *batch,
2654 struct dp_netdev_pmd_thread *pmd)
2656 struct dp_netdev_actions *actions;
2657 struct dp_netdev_flow *flow = batch->flow;
2659 dp_netdev_flow_used(batch->flow, batch->packet_count, batch->byte_count,
2662 actions = dp_netdev_flow_get_actions(flow);
2664 dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
2665 actions->actions, actions->size);
2667 dp_netdev_count_packet(pmd->dp, DP_STAT_HIT, batch->packet_count);
2671 dp_netdev_queue_batches(struct dpif_packet *pkt,
2672 struct dp_netdev_flow *flow, const struct miniflow *mf,
2673 struct packet_batch *batches, size_t *n_batches,
2676 struct packet_batch *batch = NULL;
2679 if (OVS_UNLIKELY(!flow)) {
2682 /* XXX: This O(n^2) algortihm makes sense if we're operating under the
2683 * assumption that the number of distinct flows (and therefore the
2684 * number of distinct batches) is quite small. If this turns out not
2685 * to be the case, it may make sense to pre sort based on the
2686 * netdev_flow pointer. That done we can get the appropriate batching
2687 * in O(n * log(n)) instead. */
2688 for (j = *n_batches - 1; j >= 0; j--) {
2689 if (batches[j].flow == flow) {
2690 batch = &batches[j];
2691 packet_batch_update(batch, pkt, mf);
2695 if (OVS_UNLIKELY(*n_batches >= max_batches)) {
2699 batch = &batches[(*n_batches)++];
2700 packet_batch_init(batch, flow);
2701 packet_batch_update(batch, pkt, mf);
2706 dpif_packet_swap(struct dpif_packet **a, struct dpif_packet **b)
2708 struct dpif_packet *tmp = *a;
2713 /* Try to process all ('cnt') the 'packets' using only the exact match cache
2714 * 'flow_cache'. If a flow is not found for a packet 'packets[i]', or if there
2715 * is no matching batch for a packet's flow, the miniflow is copied into 'keys'
2716 * and the packet pointer is moved at the beginning of the 'packets' array.
2718 * The function returns the number of packets that needs to be processed in the
2719 * 'packets' array (they have been moved to the beginning of the vector).
2721 static inline size_t
2722 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dpif_packet **packets,
2723 size_t cnt, struct netdev_flow_key *keys)
2725 struct netdev_flow_key key;
2726 struct packet_batch batches[4];
2727 struct emc_cache *flow_cache = &pmd->flow_cache;
2728 size_t n_batches, i;
2729 size_t notfound_cnt = 0;
2732 miniflow_initialize(&key.mf, key.buf);
2733 for (i = 0; i < cnt; i++) {
2734 struct dp_netdev_flow *flow;
2736 if (OVS_UNLIKELY(ofpbuf_size(&packets[i]->ofpbuf) < ETH_HEADER_LEN)) {
2737 dpif_packet_delete(packets[i]);
2741 miniflow_extract(&packets[i]->ofpbuf, &packets[i]->md, &key.mf);
2742 key.len = 0; /* Not computed yet. */
2743 key.hash = dpif_netdev_packet_get_dp_hash(packets[i], &key.mf);
2745 flow = emc_lookup(flow_cache, &key);
2746 if (OVS_UNLIKELY(!dp_netdev_queue_batches(packets[i], flow, &key.mf,
2747 batches, &n_batches,
2748 ARRAY_SIZE(batches)))) {
2749 if (i != notfound_cnt) {
2750 dpif_packet_swap(&packets[i], &packets[notfound_cnt]);
2753 keys[notfound_cnt++] = key;
2757 for (i = 0; i < n_batches; i++) {
2758 packet_batch_execute(&batches[i], pmd);
2761 return notfound_cnt;
2765 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
2766 struct dpif_packet **packets, size_t cnt,
2767 struct netdev_flow_key *keys)
2769 #if !defined(__CHECKER__) && !defined(_WIN32)
2770 const size_t PKT_ARRAY_SIZE = cnt;
2772 /* Sparse or MSVC doesn't like variable length array. */
2773 enum { PKT_ARRAY_SIZE = NETDEV_MAX_RX_BATCH };
2775 struct packet_batch batches[PKT_ARRAY_SIZE];
2776 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
2777 struct dp_netdev *dp = pmd->dp;
2778 struct emc_cache *flow_cache = &pmd->flow_cache;
2779 size_t n_batches, i;
2782 for (i = 0; i < cnt; i++) {
2783 /* Key length is needed in all the cases, hash computed on demand. */
2784 keys[i].len = netdev_flow_key_size(count_1bits(keys[i].mf.map));
2786 any_miss = !dpcls_lookup(&dp->cls, keys, rules, cnt);
2787 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
2788 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
2789 struct ofpbuf actions, put_actions;
2791 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
2792 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
2794 for (i = 0; i < cnt; i++) {
2795 struct dp_netdev_flow *netdev_flow;
2796 struct ofpbuf *add_actions;
2800 if (OVS_LIKELY(rules[i])) {
2804 /* It's possible that an earlier slow path execution installed
2805 * a rule covering this flow. In this case, it's a lot cheaper
2806 * to catch it here than execute a miss. */
2807 netdev_flow = dp_netdev_lookup_flow(dp, &keys[i]);
2809 rules[i] = &netdev_flow->cr;
2813 miniflow_expand(&keys[i].mf, &match.flow);
2815 ofpbuf_clear(&actions);
2816 ofpbuf_clear(&put_actions);
2818 error = dp_netdev_upcall(dp, packets[i], &match.flow, &match.wc,
2819 DPIF_UC_MISS, NULL, &actions,
2821 if (OVS_UNLIKELY(error && error != ENOSPC)) {
2825 /* We can't allow the packet batching in the next loop to execute
2826 * the actions. Otherwise, if there are any slow path actions,
2827 * we'll send the packet up twice. */
2828 dp_netdev_execute_actions(pmd, &packets[i], 1, true,
2829 ofpbuf_data(&actions),
2830 ofpbuf_size(&actions));
2832 add_actions = ofpbuf_size(&put_actions)
2836 if (OVS_LIKELY(error != ENOSPC)) {
2837 /* XXX: There's a race window where a flow covering this packet
2838 * could have already been installed since we last did the flow
2839 * lookup before upcall. This could be solved by moving the
2840 * mutex lock outside the loop, but that's an awful long time
2841 * to be locking everyone out of making flow installs. If we
2842 * move to a per-core classifier, it would be reasonable. */
2843 ovs_mutex_lock(&dp->flow_mutex);
2844 netdev_flow = dp_netdev_lookup_flow(dp, &keys[i]);
2845 if (OVS_LIKELY(!netdev_flow)) {
2846 netdev_flow = dp_netdev_flow_add(dp, &match,
2847 ofpbuf_data(add_actions),
2848 ofpbuf_size(add_actions));
2850 ovs_mutex_unlock(&dp->flow_mutex);
2852 emc_insert(flow_cache, &keys[i], netdev_flow);
2856 ofpbuf_uninit(&actions);
2857 ofpbuf_uninit(&put_actions);
2858 fat_rwlock_unlock(&dp->upcall_rwlock);
2859 } else if (OVS_UNLIKELY(any_miss)) {
2860 int dropped_cnt = 0;
2862 for (i = 0; i < cnt; i++) {
2863 if (OVS_UNLIKELY(!rules[i])) {
2864 dpif_packet_delete(packets[i]);
2869 dp_netdev_count_packet(dp, DP_STAT_LOST, dropped_cnt);
2873 for (i = 0; i < cnt; i++) {
2874 struct dpif_packet *packet = packets[i];
2875 struct dp_netdev_flow *flow;
2877 if (OVS_UNLIKELY(!rules[i])) {
2881 flow = dp_netdev_flow_cast(rules[i]);
2883 emc_insert(flow_cache, &keys[i], flow);
2884 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches,
2885 &n_batches, ARRAY_SIZE(batches));
2888 for (i = 0; i < n_batches; i++) {
2889 packet_batch_execute(&batches[i], pmd);
2894 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
2895 struct dpif_packet **packets, int cnt)
2897 #if !defined(__CHECKER__) && !defined(_WIN32)
2898 const size_t PKT_ARRAY_SIZE = cnt;
2900 /* Sparse or MSVC doesn't like variable length array. */
2901 enum { PKT_ARRAY_SIZE = NETDEV_MAX_RX_BATCH };
2903 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
2906 newcnt = emc_processing(pmd, packets, cnt, keys);
2907 if (OVS_UNLIKELY(newcnt)) {
2908 fast_path_processing(pmd, packets, newcnt, keys);
2912 struct dp_netdev_execute_aux {
2913 struct dp_netdev_pmd_thread *pmd;
2917 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
2920 struct dp_netdev *dp = get_dp_netdev(dpif);
2921 dp->upcall_aux = aux;
2926 dp_netdev_drop_packets(struct dpif_packet ** packets, int cnt, bool may_steal)
2931 for (i = 0; i < cnt; i++) {
2932 dpif_packet_delete(packets[i]);
2938 dp_execute_cb(void *aux_, struct dpif_packet **packets, int cnt,
2939 const struct nlattr *a, bool may_steal)
2940 OVS_NO_THREAD_SAFETY_ANALYSIS
2942 struct dp_netdev_execute_aux *aux = aux_;
2943 uint32_t *depth = recirc_depth_get();
2944 struct dp_netdev_pmd_thread *pmd= aux->pmd;
2945 struct dp_netdev *dp= pmd->dp;
2946 int type = nl_attr_type(a);
2947 struct dp_netdev_port *p;
2950 switch ((enum ovs_action_attr)type) {
2951 case OVS_ACTION_ATTR_OUTPUT:
2952 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
2953 if (OVS_LIKELY(p)) {
2954 netdev_send(p->netdev, pmd->core_id, packets, cnt, may_steal);
2959 case OVS_ACTION_ATTR_USERSPACE:
2960 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
2961 const struct nlattr *userdata;
2962 struct ofpbuf actions;
2965 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
2966 ofpbuf_init(&actions, 0);
2968 for (i = 0; i < cnt; i++) {
2971 ofpbuf_clear(&actions);
2973 flow_extract(&packets[i]->ofpbuf, &packets[i]->md, &flow);
2974 error = dp_netdev_upcall(dp, packets[i], &flow, NULL,
2975 DPIF_UC_ACTION, userdata, &actions,
2977 if (!error || error == ENOSPC) {
2978 dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
2979 ofpbuf_data(&actions),
2980 ofpbuf_size(&actions));
2981 } else if (may_steal) {
2982 dpif_packet_delete(packets[i]);
2985 ofpbuf_uninit(&actions);
2986 fat_rwlock_unlock(&dp->upcall_rwlock);
2992 case OVS_ACTION_ATTR_HASH: {
2993 const struct ovs_action_hash *hash_act;
2996 hash_act = nl_attr_get(a);
2998 for (i = 0; i < cnt; i++) {
3000 if (hash_act->hash_alg == OVS_HASH_ALG_L4) {
3001 /* Hash need not be symmetric, nor does it need to include
3003 hash = hash_2words(dpif_packet_get_dp_hash(packets[i]),
3004 hash_act->hash_basis);
3006 VLOG_WARN("Unknown hash algorithm specified "
3007 "for the hash action.");
3012 hash = 1; /* 0 is not valid */
3015 dpif_packet_set_dp_hash(packets[i], hash);
3020 case OVS_ACTION_ATTR_RECIRC:
3021 if (*depth < MAX_RECIRC_DEPTH) {
3024 for (i = 0; i < cnt; i++) {
3025 struct dpif_packet *recirc_pkt;
3027 recirc_pkt = (may_steal) ? packets[i]
3028 : dpif_packet_clone(packets[i]);
3030 recirc_pkt->md.recirc_id = nl_attr_get_u32(a);
3032 /* Hash is private to each packet */
3033 recirc_pkt->md.dp_hash = dpif_packet_get_dp_hash(packets[i]);
3035 dp_netdev_input(pmd, &recirc_pkt, 1);
3042 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3045 case OVS_ACTION_ATTR_PUSH_VLAN:
3046 case OVS_ACTION_ATTR_POP_VLAN:
3047 case OVS_ACTION_ATTR_PUSH_MPLS:
3048 case OVS_ACTION_ATTR_POP_MPLS:
3049 case OVS_ACTION_ATTR_SET:
3050 case OVS_ACTION_ATTR_SET_MASKED:
3051 case OVS_ACTION_ATTR_SAMPLE:
3052 case OVS_ACTION_ATTR_UNSPEC:
3053 case __OVS_ACTION_ATTR_MAX:
3057 dp_netdev_drop_packets(packets, cnt, may_steal);
3061 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3062 struct dpif_packet **packets, int cnt,
3064 const struct nlattr *actions, size_t actions_len)
3066 struct dp_netdev_execute_aux aux = { pmd };
3068 odp_execute_actions(&aux, packets, cnt, may_steal, actions,
3069 actions_len, dp_execute_cb);
3072 const struct dpif_class dpif_netdev_class = {
3074 dpif_netdev_enumerate,
3075 dpif_netdev_port_open_type,
3078 dpif_netdev_destroy,
3081 dpif_netdev_get_stats,
3082 dpif_netdev_port_add,
3083 dpif_netdev_port_del,
3084 dpif_netdev_port_query_by_number,
3085 dpif_netdev_port_query_by_name,
3086 NULL, /* port_get_pid */
3087 dpif_netdev_port_dump_start,
3088 dpif_netdev_port_dump_next,
3089 dpif_netdev_port_dump_done,
3090 dpif_netdev_port_poll,
3091 dpif_netdev_port_poll_wait,
3092 dpif_netdev_flow_flush,
3093 dpif_netdev_flow_dump_create,
3094 dpif_netdev_flow_dump_destroy,
3095 dpif_netdev_flow_dump_thread_create,
3096 dpif_netdev_flow_dump_thread_destroy,
3097 dpif_netdev_flow_dump_next,
3098 dpif_netdev_operate,
3099 NULL, /* recv_set */
3100 NULL, /* handlers_set */
3101 dpif_netdev_pmd_set,
3102 dpif_netdev_queue_to_priority,
3104 NULL, /* recv_wait */
3105 NULL, /* recv_purge */
3106 dpif_netdev_register_upcall_cb,
3107 dpif_netdev_enable_upcall,
3108 dpif_netdev_disable_upcall,
3109 dpif_netdev_get_datapath_version,
3113 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3114 const char *argv[], void *aux OVS_UNUSED)
3116 struct dp_netdev_port *old_port;
3117 struct dp_netdev_port *new_port;
3118 struct dp_netdev *dp;
3121 ovs_mutex_lock(&dp_netdev_mutex);
3122 dp = shash_find_data(&dp_netdevs, argv[1]);
3123 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3124 ovs_mutex_unlock(&dp_netdev_mutex);
3125 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3128 ovs_refcount_ref(&dp->ref_cnt);
3129 ovs_mutex_unlock(&dp_netdev_mutex);
3131 ovs_mutex_lock(&dp->port_mutex);
3132 if (get_port_by_name(dp, argv[2], &old_port)) {
3133 unixctl_command_reply_error(conn, "unknown port");
3137 port_no = u32_to_odp(atoi(argv[3]));
3138 if (!port_no || port_no == ODPP_NONE) {
3139 unixctl_command_reply_error(conn, "bad port number");
3142 if (dp_netdev_lookup_port(dp, port_no)) {
3143 unixctl_command_reply_error(conn, "port number already in use");
3147 /* Remove old port. */
3148 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
3149 ovsrcu_postpone(free, old_port);
3151 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3152 new_port = xmemdup(old_port, sizeof *old_port);
3153 new_port->port_no = port_no;
3154 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3156 seq_change(dp->port_seq);
3157 unixctl_command_reply(conn, NULL);
3160 ovs_mutex_unlock(&dp->port_mutex);
3161 dp_netdev_unref(dp);
3165 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
3166 const char *argv[], void *aux OVS_UNUSED)
3168 struct dp_netdev_port *port;
3169 struct dp_netdev *dp;
3171 ovs_mutex_lock(&dp_netdev_mutex);
3172 dp = shash_find_data(&dp_netdevs, argv[1]);
3173 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3174 ovs_mutex_unlock(&dp_netdev_mutex);
3175 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3178 ovs_refcount_ref(&dp->ref_cnt);
3179 ovs_mutex_unlock(&dp_netdev_mutex);
3181 ovs_mutex_lock(&dp->port_mutex);
3182 if (get_port_by_name(dp, argv[2], &port)) {
3183 unixctl_command_reply_error(conn, "unknown port");
3184 } else if (port->port_no == ODPP_LOCAL) {
3185 unixctl_command_reply_error(conn, "can't delete local port");
3187 do_del_port(dp, port);
3188 unixctl_command_reply(conn, NULL);
3190 ovs_mutex_unlock(&dp->port_mutex);
3192 dp_netdev_unref(dp);
3196 dpif_dummy_register__(const char *type)
3198 struct dpif_class *class;
3200 class = xmalloc(sizeof *class);
3201 *class = dpif_netdev_class;
3202 class->type = xstrdup(type);
3203 dp_register_provider(class);
3207 dpif_dummy_register(bool override)
3214 dp_enumerate_types(&types);
3215 SSET_FOR_EACH (type, &types) {
3216 if (!dp_unregister_provider(type)) {
3217 dpif_dummy_register__(type);
3220 sset_destroy(&types);
3223 dpif_dummy_register__("dummy");
3225 unixctl_command_register("dpif-dummy/change-port-number",
3226 "dp port new-number",
3227 3, 3, dpif_dummy_change_port_number, NULL);
3228 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3229 2, 2, dpif_dummy_delete_port, NULL);
3232 /* Datapath Classifier. */
3234 /* A set of rules that all have the same fields wildcarded. */
3235 struct dpcls_subtable {
3236 /* The fields are only used by writers. */
3237 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
3239 /* These fields are accessed by readers. */
3240 struct cmap rules; /* Contains "struct dpcls_rule"s. */
3241 struct netdev_flow_key mask; /* Wildcards for fields (const). */
3242 /* 'mask' must be the last field, additional space is allocated here. */
3245 /* Initializes 'cls' as a classifier that initially contains no classification
3248 dpcls_init(struct dpcls *cls)
3250 cmap_init(&cls->subtables_map);
3251 pvector_init(&cls->subtables);
3255 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
3257 pvector_remove(&cls->subtables, subtable);
3258 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
3259 subtable->mask.hash);
3260 cmap_destroy(&subtable->rules);
3261 ovsrcu_postpone(free, subtable);
3264 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
3265 * caller's responsibility.
3266 * May only be called after all the readers have been terminated. */
3268 dpcls_destroy(struct dpcls *cls)
3271 struct dpcls_subtable *subtable;
3273 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
3274 dpcls_destroy_subtable(cls, subtable);
3276 cmap_destroy(&cls->subtables_map);
3277 pvector_destroy(&cls->subtables);
3281 static struct dpcls_subtable *
3282 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3284 struct dpcls_subtable *subtable;
3286 /* Need to add one. */
3287 subtable = xmalloc(sizeof *subtable
3288 - sizeof subtable->mask.mf + mask->len);
3289 cmap_init(&subtable->rules);
3290 netdev_flow_key_clone(&subtable->mask, mask);
3291 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
3292 pvector_insert(&cls->subtables, subtable, 0);
3297 static inline struct dpcls_subtable *
3298 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3300 struct dpcls_subtable *subtable;
3302 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
3303 &cls->subtables_map) {
3304 if (netdev_flow_key_equal(&subtable->mask, mask)) {
3308 return dpcls_create_subtable(cls, mask);
3311 /* Insert 'rule' into 'cls'. */
3313 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
3314 const struct netdev_flow_key *mask)
3316 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
3318 rule->mask = &subtable->mask;
3319 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
3322 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
3324 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
3326 struct dpcls_subtable *subtable;
3328 ovs_assert(rule->mask);
3330 INIT_CONTAINER(subtable, rule->mask, mask);
3332 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
3334 dpcls_destroy_subtable(cls, subtable);
3338 /* Returns true if 'target' satisifies 'key' in 'mask', that is, if each 1-bit
3339 * in 'mask' the values in 'key' and 'target' are the same.
3341 * Note: 'key' and 'mask' have the same mask, and 'key' is already masked. */
3343 dpcls_rule_matches_key(const struct dpcls_rule *rule,
3344 const struct netdev_flow_key *target)
3346 const uint32_t *keyp = rule->flow.mf.inline_values;
3347 const uint32_t *maskp = rule->mask->mf.inline_values;
3348 uint32_t target_u32;
3350 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(target_u32, target, rule->flow.mf.map) {
3351 if (OVS_UNLIKELY((target_u32 & *maskp++) != *keyp++)) {
3358 /* For each miniflow in 'flows' performs a classifier lookup writing the result
3359 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
3360 * NULL it is skipped.
3362 * This function is optimized for use in the userspace datapath and therefore
3363 * does not implement a lot of features available in the standard
3364 * classifier_lookup() function. Specifically, it does not implement
3365 * priorities, instead returning any rule which matches the flow.
3367 * Returns true if all flows found a corresponding rule. */
3369 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
3370 struct dpcls_rule **rules, const size_t cnt)
3372 /* The batch size 16 was experimentally found faster than 8 or 32. */
3373 typedef uint16_t map_type;
3374 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
3376 #if !defined(__CHECKER__) && !defined(_WIN32)
3377 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
3379 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_RX_BATCH, MAP_BITS) };
3381 map_type maps[N_MAPS];
3382 struct dpcls_subtable *subtable;
3384 memset(maps, 0xff, sizeof maps);
3385 if (cnt % MAP_BITS) {
3386 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
3388 memset(rules, 0, cnt * sizeof *rules);
3390 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
3391 const struct netdev_flow_key *mkeys = keys;
3392 struct dpcls_rule **mrules = rules;
3393 map_type remains = 0;
3396 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
3398 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
3399 uint32_t hashes[MAP_BITS];
3400 const struct cmap_node *nodes[MAP_BITS];
3401 unsigned long map = maps[m];
3405 continue; /* Skip empty maps. */
3408 /* Compute hashes for the remaining keys. */
3409 ULONG_FOR_EACH_1(i, map) {
3410 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
3414 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
3415 /* Check results. */
3416 ULONG_FOR_EACH_1(i, map) {
3417 struct dpcls_rule *rule;
3419 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
3420 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
3425 ULONG_SET0(map, i); /* Did not match. */
3427 ; /* Keep Sparse happy. */
3429 maps[m] &= ~map; /* Clear the found rules. */
3433 return true; /* All found. */
3436 return false; /* Some misses. */