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>
36 #include "dp-packet.h"
38 #include "dpif-provider.h"
40 #include "dynamic-string.h"
41 #include "fat-rwlock.h"
47 #include "meta-flow.h"
49 #include "netdev-dpdk.h"
50 #include "netdev-vport.h"
52 #include "odp-execute.h"
54 #include "ofp-print.h"
59 #include "poll-loop.h"
66 #include "tnl-arp-cache.h"
69 #include "openvswitch/vlog.h"
71 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
73 #define FLOW_DUMP_MAX_BATCH 50
74 /* Use per thread recirc_depth to prevent recirculation loop. */
75 #define MAX_RECIRC_DEPTH 5
76 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
78 /* Configuration parameters. */
79 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
81 /* Protects against changes to 'dp_netdevs'. */
82 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
84 /* Contains all 'struct dp_netdev's. */
85 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
86 = SHASH_INITIALIZER(&dp_netdevs);
88 static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600);
90 /* Stores a miniflow with inline values */
92 struct netdev_flow_key {
93 uint32_t hash; /* Hash function differs for different users. */
94 uint32_t len; /* Length of the following miniflow (incl. map). */
96 uint64_t buf[FLOW_MAX_PACKET_U64S - MINI_N_INLINE];
99 /* Exact match cache for frequently used flows
101 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
102 * search its entries for a miniflow that matches exactly the miniflow of the
103 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
105 * A cache entry holds a reference to its 'dp_netdev_flow'.
107 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
108 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
109 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
110 * value is the index of a cache entry where the miniflow could be.
116 * Each pmd_thread has its own private exact match cache.
117 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
120 #define EM_FLOW_HASH_SHIFT 10
121 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
122 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
123 #define EM_FLOW_HASH_SEGS 2
126 struct dp_netdev_flow *flow;
127 struct netdev_flow_key key; /* key.hash used for emc hash value. */
131 struct emc_entry entries[EM_FLOW_HASH_ENTRIES];
132 int sweep_idx; /* For emc_cache_slow_sweep(). */
135 /* Iterate in the exact match cache through every entry that might contain a
136 * miniflow with hash 'HASH'. */
137 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
138 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
139 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
140 i__ < EM_FLOW_HASH_SEGS; \
141 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
143 /* Simple non-wildcarding single-priority classifier. */
146 struct cmap subtables_map;
147 struct pvector subtables;
150 /* A rule to be inserted to the classifier. */
152 struct cmap_node cmap_node; /* Within struct dpcls_subtable 'rules'. */
153 struct netdev_flow_key *mask; /* Subtable's mask. */
154 struct netdev_flow_key flow; /* Matching key. */
155 /* 'flow' must be the last field, additional space is allocated here. */
158 static void dpcls_init(struct dpcls *);
159 static void dpcls_destroy(struct dpcls *);
160 static void dpcls_insert(struct dpcls *, struct dpcls_rule *,
161 const struct netdev_flow_key *mask);
162 static void dpcls_remove(struct dpcls *, struct dpcls_rule *);
163 static bool dpcls_lookup(const struct dpcls *cls,
164 const struct netdev_flow_key keys[],
165 struct dpcls_rule **rules, size_t cnt);
167 /* Datapath based on the network device interface from netdev.h.
173 * Some members, marked 'const', are immutable. Accessing other members
174 * requires synchronization, as noted in more detail below.
176 * Acquisition order is, from outermost to innermost:
178 * dp_netdev_mutex (global)
182 const struct dpif_class *const class;
183 const char *const name;
185 struct ovs_refcount ref_cnt;
186 atomic_flag destroyed;
190 * Protected by RCU. Take the mutex to add or remove ports. */
191 struct ovs_mutex port_mutex;
193 struct seq *port_seq; /* Incremented whenever a port changes. */
195 /* Protects access to ofproto-dpif-upcall interface during revalidator
196 * thread synchronization. */
197 struct fat_rwlock upcall_rwlock;
198 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
201 /* Stores all 'struct dp_netdev_pmd_thread's. */
202 struct cmap poll_threads;
204 /* Protects the access of the 'struct dp_netdev_pmd_thread'
205 * instance for non-pmd thread. */
206 struct ovs_mutex non_pmd_mutex;
208 /* Each pmd thread will store its pointer to
209 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
210 ovsthread_key_t per_pmd_key;
212 /* Number of rx queues for each dpdk interface and the cpu mask
213 * for pin of pmd threads. */
216 uint64_t last_tnl_conf_seq;
219 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
223 DP_STAT_EXACT_HIT, /* Packets that had an exact match (emc). */
224 DP_STAT_MASKED_HIT, /* Packets that matched in the flow table. */
225 DP_STAT_MISS, /* Packets that did not match. */
226 DP_STAT_LOST, /* Packets not passed up to the client. */
230 /* A port in a netdev-based datapath. */
231 struct dp_netdev_port {
232 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
234 struct netdev *netdev;
235 struct netdev_saved_flags *sf;
236 struct netdev_rxq **rxq;
237 struct ovs_refcount ref_cnt;
238 char *type; /* Port type as requested by user. */
241 /* Contained by struct dp_netdev_flow's 'stats' member. */
242 struct dp_netdev_flow_stats {
243 atomic_llong used; /* Last used time, in monotonic msecs. */
244 atomic_ullong packet_count; /* Number of packets matched. */
245 atomic_ullong byte_count; /* Number of bytes matched. */
246 atomic_uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */
249 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
255 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
256 * its pmd thread's classifier. The text below calls this classifier 'cls'.
261 * The thread safety rules described here for "struct dp_netdev_flow" are
262 * motivated by two goals:
264 * - Prevent threads that read members of "struct dp_netdev_flow" from
265 * reading bad data due to changes by some thread concurrently modifying
268 * - Prevent two threads making changes to members of a given "struct
269 * dp_netdev_flow" from interfering with each other.
275 * A flow 'flow' may be accessed without a risk of being freed during an RCU
276 * grace period. Code that needs to hold onto a flow for a while
277 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
279 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
280 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
283 * Some members, marked 'const', are immutable. Accessing other members
284 * requires synchronization, as noted in more detail below.
286 struct dp_netdev_flow {
289 /* Hash table index by unmasked flow. */
290 const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */
292 const ovs_u128 ufid; /* Unique flow identifier. */
293 const struct flow flow; /* Unmasked flow that created this entry. */
294 const int pmd_id; /* The 'core_id' of pmd thread owning this */
297 /* Number of references.
298 * The classifier owns one reference.
299 * Any thread trying to keep a rule from being freed should hold its own
301 struct ovs_refcount ref_cnt;
304 struct dp_netdev_flow_stats stats;
307 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
309 /* Packet classification. */
310 struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
311 /* 'cr' must be the last member. */
314 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
315 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
316 static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t,
319 /* A set of datapath actions within a "struct dp_netdev_flow".
325 * A struct dp_netdev_actions 'actions' is protected with RCU. */
326 struct dp_netdev_actions {
327 /* These members are immutable: they do not change during the struct's
329 struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */
330 unsigned int size; /* Size of 'actions', in bytes. */
333 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
335 struct dp_netdev_actions *dp_netdev_flow_get_actions(
336 const struct dp_netdev_flow *);
337 static void dp_netdev_actions_free(struct dp_netdev_actions *);
339 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
340 struct dp_netdev_pmd_stats {
341 /* Indexed by DP_STAT_*. */
342 atomic_ullong n[DP_N_STATS];
345 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
346 * the performance overhead of interrupt processing. Therefore netdev can
347 * not implement rx-wait for these devices. dpif-netdev needs to poll
348 * these device to check for recv buffer. pmd-thread does polling for
349 * devices assigned to itself.
351 * DPDK used PMD for accessing NIC.
353 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
354 * I/O of all non-pmd threads. There will be no actual thread created
357 * Each struct has its own flow table and classifier. Packets received
358 * from managed ports are looked up in the corresponding pmd thread's
359 * flow table, and are executed with the found actions.
361 struct dp_netdev_pmd_thread {
362 struct dp_netdev *dp;
363 struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */
364 struct cmap_node node; /* In 'dp->poll_threads'. */
366 pthread_cond_t cond; /* For synchronizing pmd thread reload. */
367 struct ovs_mutex cond_mutex; /* Mutex for condition variable. */
369 /* Per thread exact-match cache. Note, the instance for cpu core
370 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
371 * need to be protected (e.g. by 'dp_netdev_mutex'). All other
372 * instances will only be accessed by its own pmd thread. */
373 struct emc_cache flow_cache;
375 /* Classifier and Flow-Table.
377 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
378 * changes to 'cls' must be made while still holding the 'flow_mutex'.
380 struct ovs_mutex flow_mutex;
382 struct cmap flow_table OVS_GUARDED; /* Flow table. */
385 struct dp_netdev_pmd_stats stats;
387 struct latch exit_latch; /* For terminating the pmd thread. */
388 atomic_uint change_seq; /* For reloading pmd ports. */
390 int index; /* Idx of this pmd thread among pmd*/
391 /* threads on same numa node. */
392 int core_id; /* CPU core id of this pmd thread. */
393 int numa_id; /* numa node id of this pmd thread. */
396 #define PMD_INITIAL_SEQ 1
398 /* Interface to netdev-based datapath. */
401 struct dp_netdev *dp;
402 uint64_t last_port_seq;
405 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
406 struct dp_netdev_port **portp);
407 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
408 struct dp_netdev_port **portp);
409 static void dp_netdev_free(struct dp_netdev *)
410 OVS_REQUIRES(dp_netdev_mutex);
411 static int do_add_port(struct dp_netdev *dp, const char *devname,
412 const char *type, odp_port_t port_no)
413 OVS_REQUIRES(dp->port_mutex);
414 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
415 OVS_REQUIRES(dp->port_mutex);
416 static int dpif_netdev_open(const struct dpif_class *, const char *name,
417 bool create, struct dpif **);
418 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
419 struct dp_packet **, int c,
421 const struct nlattr *actions,
423 static void dp_netdev_input(struct dp_netdev_pmd_thread *,
424 struct dp_packet **, int cnt);
426 static void dp_netdev_disable_upcall(struct dp_netdev *);
427 void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd);
428 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
429 struct dp_netdev *dp, int index,
430 int core_id, int numa_id);
431 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd);
432 static void dp_netdev_set_nonpmd(struct dp_netdev *dp);
433 static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp,
435 static struct dp_netdev_pmd_thread *
436 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos);
437 static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
438 static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
439 static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id);
440 static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp);
441 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd);
442 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd);
443 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd);
445 static inline bool emc_entry_alive(struct emc_entry *ce);
446 static void emc_clear_entry(struct emc_entry *ce);
449 emc_cache_init(struct emc_cache *flow_cache)
453 BUILD_ASSERT(offsetof(struct miniflow, inline_values) == sizeof(uint64_t));
455 flow_cache->sweep_idx = 0;
456 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
457 flow_cache->entries[i].flow = NULL;
458 flow_cache->entries[i].key.hash = 0;
459 flow_cache->entries[i].key.len
460 = offsetof(struct miniflow, inline_values);
461 miniflow_initialize(&flow_cache->entries[i].key.mf,
462 flow_cache->entries[i].key.buf);
467 emc_cache_uninit(struct emc_cache *flow_cache)
471 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
472 emc_clear_entry(&flow_cache->entries[i]);
476 /* Check and clear dead flow references slowly (one entry at each
479 emc_cache_slow_sweep(struct emc_cache *flow_cache)
481 struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx];
483 if (!emc_entry_alive(entry)) {
484 emc_clear_entry(entry);
486 flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK;
489 static struct dpif_netdev *
490 dpif_netdev_cast(const struct dpif *dpif)
492 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
493 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
496 static struct dp_netdev *
497 get_dp_netdev(const struct dpif *dpif)
499 return dpif_netdev_cast(dpif)->dp;
503 dpif_netdev_enumerate(struct sset *all_dps,
504 const struct dpif_class *dpif_class)
506 struct shash_node *node;
508 ovs_mutex_lock(&dp_netdev_mutex);
509 SHASH_FOR_EACH(node, &dp_netdevs) {
510 struct dp_netdev *dp = node->data;
511 if (dpif_class != dp->class) {
512 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
513 * If the class doesn't match, skip this dpif. */
516 sset_add(all_dps, node->name);
518 ovs_mutex_unlock(&dp_netdev_mutex);
524 dpif_netdev_class_is_dummy(const struct dpif_class *class)
526 return class != &dpif_netdev_class;
530 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
532 return strcmp(type, "internal") ? type
533 : dpif_netdev_class_is_dummy(class) ? "dummy"
538 create_dpif_netdev(struct dp_netdev *dp)
540 uint16_t netflow_id = hash_string(dp->name, 0);
541 struct dpif_netdev *dpif;
543 ovs_refcount_ref(&dp->ref_cnt);
545 dpif = xmalloc(sizeof *dpif);
546 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
548 dpif->last_port_seq = seq_read(dp->port_seq);
553 /* Choose an unused, non-zero port number and return it on success.
554 * Return ODPP_NONE on failure. */
556 choose_port(struct dp_netdev *dp, const char *name)
557 OVS_REQUIRES(dp->port_mutex)
561 if (dp->class != &dpif_netdev_class) {
565 /* If the port name begins with "br", start the number search at
566 * 100 to make writing tests easier. */
567 if (!strncmp(name, "br", 2)) {
571 /* If the port name contains a number, try to assign that port number.
572 * This can make writing unit tests easier because port numbers are
574 for (p = name; *p != '\0'; p++) {
575 if (isdigit((unsigned char) *p)) {
576 port_no = start_no + strtol(p, NULL, 10);
577 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
578 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
579 return u32_to_odp(port_no);
586 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
587 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
588 return u32_to_odp(port_no);
596 create_dp_netdev(const char *name, const struct dpif_class *class,
597 struct dp_netdev **dpp)
598 OVS_REQUIRES(dp_netdev_mutex)
600 struct dp_netdev *dp;
603 dp = xzalloc(sizeof *dp);
604 shash_add(&dp_netdevs, name, dp);
606 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
607 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
608 ovs_refcount_init(&dp->ref_cnt);
609 atomic_flag_clear(&dp->destroyed);
611 ovs_mutex_init(&dp->port_mutex);
612 cmap_init(&dp->ports);
613 dp->port_seq = seq_create();
614 fat_rwlock_init(&dp->upcall_rwlock);
616 /* Disable upcalls by default. */
617 dp_netdev_disable_upcall(dp);
618 dp->upcall_aux = NULL;
619 dp->upcall_cb = NULL;
621 cmap_init(&dp->poll_threads);
622 ovs_mutex_init_recursive(&dp->non_pmd_mutex);
623 ovsthread_key_create(&dp->per_pmd_key, NULL);
625 /* Reserves the core NON_PMD_CORE_ID for all non-pmd threads. */
626 ovs_numa_try_pin_core_specific(NON_PMD_CORE_ID);
627 dp_netdev_set_nonpmd(dp);
628 dp->n_dpdk_rxqs = NR_QUEUE;
630 ovs_mutex_lock(&dp->port_mutex);
631 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
632 ovs_mutex_unlock(&dp->port_mutex);
638 dp->last_tnl_conf_seq = seq_read(tnl_conf_seq);
644 dpif_netdev_open(const struct dpif_class *class, const char *name,
645 bool create, struct dpif **dpifp)
647 struct dp_netdev *dp;
650 ovs_mutex_lock(&dp_netdev_mutex);
651 dp = shash_find_data(&dp_netdevs, name);
653 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
655 error = (dp->class != class ? EINVAL
660 *dpifp = create_dpif_netdev(dp);
663 ovs_mutex_unlock(&dp_netdev_mutex);
669 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
670 OVS_NO_THREAD_SAFETY_ANALYSIS
672 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
673 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
675 /* Before freeing a lock we should release it */
676 fat_rwlock_unlock(&dp->upcall_rwlock);
677 fat_rwlock_destroy(&dp->upcall_rwlock);
680 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
681 * through the 'dp_netdevs' shash while freeing 'dp'. */
683 dp_netdev_free(struct dp_netdev *dp)
684 OVS_REQUIRES(dp_netdev_mutex)
686 struct dp_netdev_port *port;
688 shash_find_and_delete(&dp_netdevs, dp->name);
690 dp_netdev_destroy_all_pmds(dp);
691 cmap_destroy(&dp->poll_threads);
692 ovs_mutex_destroy(&dp->non_pmd_mutex);
693 ovsthread_key_delete(dp->per_pmd_key);
695 ovs_mutex_lock(&dp->port_mutex);
696 CMAP_FOR_EACH (port, node, &dp->ports) {
697 do_del_port(dp, port);
699 ovs_mutex_unlock(&dp->port_mutex);
701 seq_destroy(dp->port_seq);
702 cmap_destroy(&dp->ports);
704 /* Upcalls must be disabled at this point */
705 dp_netdev_destroy_upcall_lock(dp);
708 free(CONST_CAST(char *, dp->name));
713 dp_netdev_unref(struct dp_netdev *dp)
716 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
717 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
718 ovs_mutex_lock(&dp_netdev_mutex);
719 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
722 ovs_mutex_unlock(&dp_netdev_mutex);
727 dpif_netdev_close(struct dpif *dpif)
729 struct dp_netdev *dp = get_dp_netdev(dpif);
736 dpif_netdev_destroy(struct dpif *dpif)
738 struct dp_netdev *dp = get_dp_netdev(dpif);
740 if (!atomic_flag_test_and_set(&dp->destroyed)) {
741 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
742 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
750 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
751 * load/store semantics. While the increment is not atomic, the load and
752 * store operations are, making it impossible to read inconsistent values.
754 * This is used to update thread local stats counters. */
756 non_atomic_ullong_add(atomic_ullong *var, unsigned long long n)
758 unsigned long long tmp;
760 atomic_read_relaxed(var, &tmp);
762 atomic_store_relaxed(var, tmp);
766 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
768 struct dp_netdev *dp = get_dp_netdev(dpif);
769 struct dp_netdev_pmd_thread *pmd;
771 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
772 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
773 unsigned long long n;
774 stats->n_flows += cmap_count(&pmd->flow_table);
776 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MASKED_HIT], &n);
778 atomic_read_relaxed(&pmd->stats.n[DP_STAT_EXACT_HIT], &n);
780 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n);
781 stats->n_missed += n;
782 atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n);
785 stats->n_masks = UINT32_MAX;
786 stats->n_mask_hit = UINT64_MAX;
792 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
796 if (pmd->core_id == NON_PMD_CORE_ID) {
800 ovs_mutex_lock(&pmd->cond_mutex);
801 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
802 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
803 ovs_mutex_unlock(&pmd->cond_mutex);
806 /* Causes all pmd threads to reload its tx/rx devices.
807 * Must be called after adding/removing ports. */
809 dp_netdev_reload_pmds(struct dp_netdev *dp)
811 struct dp_netdev_pmd_thread *pmd;
813 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
814 dp_netdev_reload_pmd__(pmd);
819 hash_port_no(odp_port_t port_no)
821 return hash_int(odp_to_u32(port_no), 0);
825 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
827 OVS_REQUIRES(dp->port_mutex)
829 struct netdev_saved_flags *sf;
830 struct dp_netdev_port *port;
831 struct netdev *netdev;
832 enum netdev_flags flags;
833 const char *open_type;
837 /* XXX reject devices already in some dp_netdev. */
839 /* Open and validate network device. */
840 open_type = dpif_netdev_port_open_type(dp->class, type);
841 error = netdev_open(devname, open_type, &netdev);
845 /* XXX reject non-Ethernet devices */
847 netdev_get_flags(netdev, &flags);
848 if (flags & NETDEV_LOOPBACK) {
849 VLOG_ERR("%s: cannot add a loopback device", devname);
850 netdev_close(netdev);
854 if (netdev_is_pmd(netdev)) {
855 int n_cores = ovs_numa_get_n_cores();
857 if (n_cores == OVS_CORE_UNSPEC) {
858 VLOG_ERR("%s, cannot get cpu core info", devname);
861 /* There can only be ovs_numa_get_n_cores() pmd threads,
862 * so creates a txq for each. */
863 error = netdev_set_multiq(netdev, n_cores, dp->n_dpdk_rxqs);
864 if (error && (error != EOPNOTSUPP)) {
865 VLOG_ERR("%s, cannot set multiq", devname);
869 port = xzalloc(sizeof *port);
870 port->port_no = port_no;
871 port->netdev = netdev;
872 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
873 port->type = xstrdup(type);
874 for (i = 0; i < netdev_n_rxq(netdev); i++) {
875 error = netdev_rxq_open(netdev, &port->rxq[i], i);
877 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
878 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
879 devname, ovs_strerror(errno));
880 netdev_close(netdev);
888 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
890 for (i = 0; i < netdev_n_rxq(netdev); i++) {
891 netdev_rxq_close(port->rxq[i]);
893 netdev_close(netdev);
901 ovs_refcount_init(&port->ref_cnt);
902 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
904 if (netdev_is_pmd(netdev)) {
905 dp_netdev_set_pmds_on_numa(dp, netdev_get_numa_id(netdev));
906 dp_netdev_reload_pmds(dp);
908 seq_change(dp->port_seq);
914 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
915 odp_port_t *port_nop)
917 struct dp_netdev *dp = get_dp_netdev(dpif);
918 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
919 const char *dpif_port;
923 ovs_mutex_lock(&dp->port_mutex);
924 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
925 if (*port_nop != ODPP_NONE) {
927 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
929 port_no = choose_port(dp, dpif_port);
930 error = port_no == ODPP_NONE ? EFBIG : 0;
934 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
936 ovs_mutex_unlock(&dp->port_mutex);
942 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
944 struct dp_netdev *dp = get_dp_netdev(dpif);
947 ovs_mutex_lock(&dp->port_mutex);
948 if (port_no == ODPP_LOCAL) {
951 struct dp_netdev_port *port;
953 error = get_port_by_number(dp, port_no, &port);
955 do_del_port(dp, port);
958 ovs_mutex_unlock(&dp->port_mutex);
964 is_valid_port_number(odp_port_t port_no)
966 return port_no != ODPP_NONE;
969 static struct dp_netdev_port *
970 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
972 struct dp_netdev_port *port;
974 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
975 if (port->port_no == port_no) {
983 get_port_by_number(struct dp_netdev *dp,
984 odp_port_t port_no, struct dp_netdev_port **portp)
986 if (!is_valid_port_number(port_no)) {
990 *portp = dp_netdev_lookup_port(dp, port_no);
991 return *portp ? 0 : ENOENT;
996 port_ref(struct dp_netdev_port *port)
999 ovs_refcount_ref(&port->ref_cnt);
1004 port_try_ref(struct dp_netdev_port *port)
1007 return ovs_refcount_try_ref_rcu(&port->ref_cnt);
1014 port_unref(struct dp_netdev_port *port)
1016 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
1017 int n_rxq = netdev_n_rxq(port->netdev);
1020 netdev_close(port->netdev);
1021 netdev_restore_flags(port->sf);
1023 for (i = 0; i < n_rxq; i++) {
1024 netdev_rxq_close(port->rxq[i]);
1033 get_port_by_name(struct dp_netdev *dp,
1034 const char *devname, struct dp_netdev_port **portp)
1035 OVS_REQUIRES(dp->port_mutex)
1037 struct dp_netdev_port *port;
1039 CMAP_FOR_EACH (port, node, &dp->ports) {
1040 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1049 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1051 struct dp_netdev_pmd_thread *pmd;
1054 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1055 if (pmd->numa_id == numa_id) {
1063 /* Returns 'true' if there is a port with pmd netdev and the netdev
1064 * is on numa node 'numa_id'. */
1066 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1068 struct dp_netdev_port *port;
1070 CMAP_FOR_EACH (port, node, &dp->ports) {
1071 if (netdev_is_pmd(port->netdev)
1072 && netdev_get_numa_id(port->netdev) == numa_id) {
1082 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1083 OVS_REQUIRES(dp->port_mutex)
1085 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
1086 seq_change(dp->port_seq);
1087 if (netdev_is_pmd(port->netdev)) {
1088 int numa_id = netdev_get_numa_id(port->netdev);
1090 /* If there is no netdev on the numa node, deletes the pmd threads
1091 * for that numa. Else, just reloads the queues. */
1092 if (!has_pmd_port_for_numa(dp, numa_id)) {
1093 dp_netdev_del_pmds_on_numa(dp, numa_id);
1095 dp_netdev_reload_pmds(dp);
1102 answer_port_query(const struct dp_netdev_port *port,
1103 struct dpif_port *dpif_port)
1105 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1106 dpif_port->type = xstrdup(port->type);
1107 dpif_port->port_no = port->port_no;
1111 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1112 struct dpif_port *dpif_port)
1114 struct dp_netdev *dp = get_dp_netdev(dpif);
1115 struct dp_netdev_port *port;
1118 error = get_port_by_number(dp, port_no, &port);
1119 if (!error && dpif_port) {
1120 answer_port_query(port, dpif_port);
1127 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1128 struct dpif_port *dpif_port)
1130 struct dp_netdev *dp = get_dp_netdev(dpif);
1131 struct dp_netdev_port *port;
1134 ovs_mutex_lock(&dp->port_mutex);
1135 error = get_port_by_name(dp, devname, &port);
1136 if (!error && dpif_port) {
1137 answer_port_query(port, dpif_port);
1139 ovs_mutex_unlock(&dp->port_mutex);
1145 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1147 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1151 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1153 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1154 ovsrcu_postpone(dp_netdev_flow_free, flow);
1159 dp_netdev_flow_hash(const ovs_u128 *ufid)
1161 return ufid->u32[0];
1165 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1166 struct dp_netdev_flow *flow)
1167 OVS_REQUIRES(pmd->flow_mutex)
1169 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1171 dpcls_remove(&pmd->cls, &flow->cr);
1172 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1175 dp_netdev_flow_unref(flow);
1179 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1181 struct dp_netdev_flow *netdev_flow;
1183 ovs_mutex_lock(&pmd->flow_mutex);
1184 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1185 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1187 ovs_mutex_unlock(&pmd->flow_mutex);
1191 dpif_netdev_flow_flush(struct dpif *dpif)
1193 struct dp_netdev *dp = get_dp_netdev(dpif);
1194 struct dp_netdev_pmd_thread *pmd;
1196 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1197 dp_netdev_pmd_flow_flush(pmd);
1203 struct dp_netdev_port_state {
1204 struct cmap_position position;
1209 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1211 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1216 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1217 struct dpif_port *dpif_port)
1219 struct dp_netdev_port_state *state = state_;
1220 struct dp_netdev *dp = get_dp_netdev(dpif);
1221 struct cmap_node *node;
1224 node = cmap_next_position(&dp->ports, &state->position);
1226 struct dp_netdev_port *port;
1228 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1231 state->name = xstrdup(netdev_get_name(port->netdev));
1232 dpif_port->name = state->name;
1233 dpif_port->type = port->type;
1234 dpif_port->port_no = port->port_no;
1245 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1247 struct dp_netdev_port_state *state = state_;
1254 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1256 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1257 uint64_t new_port_seq;
1260 new_port_seq = seq_read(dpif->dp->port_seq);
1261 if (dpif->last_port_seq != new_port_seq) {
1262 dpif->last_port_seq = new_port_seq;
1272 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1274 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1276 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1279 static struct dp_netdev_flow *
1280 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1282 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1285 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1287 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1290 /* netdev_flow_key utilities.
1292 * netdev_flow_key is basically a miniflow. We use these functions
1293 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1294 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1296 * - Since we are dealing exclusively with miniflows created by
1297 * miniflow_extract(), if the map is different the miniflow is different.
1298 * Therefore we can be faster by comparing the map and the miniflow in a
1300 * _ netdev_flow_key's miniflow has always inline values.
1301 * - These functions can be inlined by the compiler.
1303 * The following assertions make sure that what we're doing with miniflow is
1306 BUILD_ASSERT_DECL(offsetof(struct miniflow, inline_values)
1307 == sizeof(uint64_t));
1309 /* Given the number of bits set in the miniflow map, returns the size of the
1310 * 'netdev_flow_key.mf' */
1311 static inline uint32_t
1312 netdev_flow_key_size(uint32_t flow_u32s)
1314 return offsetof(struct miniflow, inline_values) +
1315 MINIFLOW_VALUES_SIZE(flow_u32s);
1319 netdev_flow_key_equal(const struct netdev_flow_key *a,
1320 const struct netdev_flow_key *b)
1322 /* 'b->len' may be not set yet. */
1323 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1326 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1327 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1328 * generated by miniflow_extract. */
1330 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1331 const struct miniflow *mf)
1333 return !memcmp(&key->mf, mf, key->len);
1337 netdev_flow_key_clone(struct netdev_flow_key *dst,
1338 const struct netdev_flow_key *src)
1341 offsetof(struct netdev_flow_key, mf) + src->len);
1346 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1347 const struct flow *src)
1349 struct dp_packet packet;
1350 uint64_t buf_stub[512 / 8];
1352 miniflow_initialize(&dst->mf, dst->buf);
1354 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1355 pkt_metadata_from_flow(&packet.md, src);
1356 flow_compose(&packet, src);
1357 miniflow_extract(&packet, &dst->mf);
1358 dp_packet_uninit(&packet);
1360 dst->len = netdev_flow_key_size(count_1bits(dst->mf.map));
1361 dst->hash = 0; /* Not computed yet. */
1364 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1366 netdev_flow_mask_init(struct netdev_flow_key *mask,
1367 const struct match *match)
1369 const uint64_t *mask_u64 = (const uint64_t *) &match->wc.masks;
1370 uint64_t *dst = mask->mf.inline_values;
1371 uint64_t map, mask_map = 0;
1375 /* Only check masks that make sense for the flow. */
1376 map = flow_wc_map(&match->flow);
1379 uint64_t rm1bit = rightmost_1bit(map);
1380 int i = raw_ctz(map);
1384 *dst++ = mask_u64[i];
1385 hash = hash_add64(hash, mask_u64[i]);
1390 mask->mf.values_inline = true;
1391 mask->mf.map = mask_map;
1393 hash = hash_add64(hash, mask_map);
1395 n = dst - mask->mf.inline_values;
1397 mask->hash = hash_finish(hash, n * 8);
1398 mask->len = netdev_flow_key_size(n);
1401 /* Initializes 'dst' as a copy of 'src' masked with 'mask'. */
1403 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1404 const struct flow *flow,
1405 const struct netdev_flow_key *mask)
1407 uint64_t *dst_u64 = dst->mf.inline_values;
1408 const uint64_t *mask_u64 = mask->mf.inline_values;
1412 dst->len = mask->len;
1413 dst->mf.values_inline = true;
1414 dst->mf.map = mask->mf.map;
1416 FLOW_FOR_EACH_IN_MAP(value, flow, mask->mf.map) {
1417 *dst_u64 = value & *mask_u64++;
1418 hash = hash_add64(hash, *dst_u64++);
1420 dst->hash = hash_finish(hash, (dst_u64 - dst->mf.inline_values) * 8);
1423 /* Iterate through all netdev_flow_key u64 values specified by 'MAP' */
1424 #define NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(VALUE, KEY, MAP) \
1425 for (struct mf_for_each_in_map_aux aux__ \
1426 = { (KEY)->mf.inline_values, (KEY)->mf.map, MAP }; \
1427 mf_get_next_in_map(&aux__, &(VALUE)); \
1430 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1432 static inline uint32_t
1433 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1434 const struct netdev_flow_key *mask)
1436 const uint64_t *p = mask->mf.inline_values;
1440 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(key_u64, key, mask->mf.map) {
1441 hash = hash_add64(hash, key_u64 & *p++);
1444 return hash_finish(hash, (p - mask->mf.inline_values) * 8);
1448 emc_entry_alive(struct emc_entry *ce)
1450 return ce->flow && !ce->flow->dead;
1454 emc_clear_entry(struct emc_entry *ce)
1457 dp_netdev_flow_unref(ce->flow);
1463 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1464 const struct netdev_flow_key *key)
1466 if (ce->flow != flow) {
1468 dp_netdev_flow_unref(ce->flow);
1471 if (dp_netdev_flow_ref(flow)) {
1478 netdev_flow_key_clone(&ce->key, key);
1483 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1484 struct dp_netdev_flow *flow)
1486 struct emc_entry *to_be_replaced = NULL;
1487 struct emc_entry *current_entry;
1489 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1490 if (netdev_flow_key_equal(¤t_entry->key, key)) {
1491 /* We found the entry with the 'mf' miniflow */
1492 emc_change_entry(current_entry, flow, NULL);
1496 /* Replacement policy: put the flow in an empty (not alive) entry, or
1497 * in the first entry where it can be */
1499 || (emc_entry_alive(to_be_replaced)
1500 && !emc_entry_alive(current_entry))
1501 || current_entry->key.hash < to_be_replaced->key.hash) {
1502 to_be_replaced = current_entry;
1505 /* We didn't find the miniflow in the cache.
1506 * The 'to_be_replaced' entry is where the new flow will be stored */
1508 emc_change_entry(to_be_replaced, flow, key);
1511 static inline struct dp_netdev_flow *
1512 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1514 struct emc_entry *current_entry;
1516 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1517 if (current_entry->key.hash == key->hash
1518 && emc_entry_alive(current_entry)
1519 && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
1521 /* We found the entry with the 'key->mf' miniflow */
1522 return current_entry->flow;
1529 static struct dp_netdev_flow *
1530 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1531 const struct netdev_flow_key *key)
1533 struct dp_netdev_flow *netdev_flow;
1534 struct dpcls_rule *rule;
1536 dpcls_lookup(&pmd->cls, key, &rule, 1);
1537 netdev_flow = dp_netdev_flow_cast(rule);
1542 static struct dp_netdev_flow *
1543 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1544 const ovs_u128 *ufidp, const struct nlattr *key,
1547 struct dp_netdev_flow *netdev_flow;
1551 /* If a UFID is not provided, determine one based on the key. */
1552 if (!ufidp && key && key_len
1553 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1554 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1559 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1561 if (ovs_u128_equal(&netdev_flow->ufid, ufidp)) {
1571 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1572 struct dpif_flow_stats *stats)
1574 struct dp_netdev_flow *netdev_flow;
1575 unsigned long long n;
1579 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1581 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1582 stats->n_packets = n;
1583 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1585 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1587 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1588 stats->tcp_flags = flags;
1591 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1592 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1593 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1596 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1597 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1598 struct dpif_flow *flow, bool terse)
1601 memset(flow, 0, sizeof *flow);
1603 struct flow_wildcards wc;
1604 struct dp_netdev_actions *actions;
1607 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1610 offset = key_buf->size;
1611 flow->key = ofpbuf_tail(key_buf);
1612 odp_flow_key_from_flow(key_buf, &netdev_flow->flow, &wc.masks,
1613 netdev_flow->flow.in_port.odp_port, true);
1614 flow->key_len = key_buf->size - offset;
1617 offset = mask_buf->size;
1618 flow->mask = ofpbuf_tail(mask_buf);
1619 odp_flow_key_from_mask(mask_buf, &wc.masks, &netdev_flow->flow,
1620 odp_to_u32(wc.masks.in_port.odp_port),
1622 flow->mask_len = mask_buf->size - offset;
1625 actions = dp_netdev_flow_get_actions(netdev_flow);
1626 flow->actions = actions->actions;
1627 flow->actions_len = actions->size;
1630 flow->ufid = netdev_flow->ufid;
1631 flow->ufid_present = true;
1632 flow->pmd_id = netdev_flow->pmd_id;
1633 get_dpif_flow_stats(netdev_flow, &flow->stats);
1637 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1638 const struct nlattr *mask_key,
1639 uint32_t mask_key_len, const struct flow *flow,
1643 enum odp_key_fitness fitness;
1645 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1647 /* This should not happen: it indicates that
1648 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1649 * disagree on the acceptable form of a mask. Log the problem
1650 * as an error, with enough details to enable debugging. */
1651 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1653 if (!VLOG_DROP_ERR(&rl)) {
1657 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1659 VLOG_ERR("internal error parsing flow mask %s (%s)",
1660 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1667 enum mf_field_id id;
1668 /* No mask key, unwildcard everything except fields whose
1669 * prerequisities are not met. */
1670 memset(mask, 0x0, sizeof *mask);
1672 for (id = 0; id < MFF_N_IDS; ++id) {
1673 /* Skip registers and metadata. */
1674 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1675 && id != MFF_METADATA) {
1676 const struct mf_field *mf = mf_from_id(id);
1677 if (mf_are_prereqs_ok(mf, flow)) {
1678 mf_mask_field(mf, mask);
1684 /* Force unwildcard the in_port.
1686 * We need to do this even in the case where we unwildcard "everything"
1687 * above because "everything" only includes the 16-bit OpenFlow port number
1688 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1689 * port number mask->in_port.odp_port. */
1690 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1696 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1701 if (odp_flow_key_to_flow(key, key_len, flow)) {
1702 /* This should not happen: it indicates that odp_flow_key_from_flow()
1703 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1704 * flow. Log the problem as an error, with enough details to enable
1706 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1708 if (!VLOG_DROP_ERR(&rl)) {
1712 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1713 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1720 in_port = flow->in_port.odp_port;
1721 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1729 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1731 struct dp_netdev *dp = get_dp_netdev(dpif);
1732 struct dp_netdev_flow *netdev_flow;
1733 struct dp_netdev_pmd_thread *pmd;
1734 int pmd_id = get->pmd_id == PMD_ID_NULL ? NON_PMD_CORE_ID : get->pmd_id;
1737 pmd = dp_netdev_get_pmd(dp, pmd_id);
1742 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
1745 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
1750 dp_netdev_pmd_unref(pmd);
1756 static struct dp_netdev_flow *
1757 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
1758 struct match *match, const ovs_u128 *ufid,
1759 const struct nlattr *actions, size_t actions_len)
1760 OVS_REQUIRES(pmd->flow_mutex)
1762 struct dp_netdev_flow *flow;
1763 struct netdev_flow_key mask;
1765 netdev_flow_mask_init(&mask, match);
1766 /* Make sure wc does not have metadata. */
1767 ovs_assert(!(mask.mf.map & (MINIFLOW_MAP(metadata) | MINIFLOW_MAP(regs))));
1769 /* Do not allocate extra space. */
1770 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
1771 memset(&flow->stats, 0, sizeof flow->stats);
1773 *CONST_CAST(int *, &flow->pmd_id) = pmd->core_id;
1774 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
1775 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
1776 ovs_refcount_init(&flow->ref_cnt);
1777 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
1779 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
1780 dpcls_insert(&pmd->cls, &flow->cr, &mask);
1782 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
1783 dp_netdev_flow_hash(&flow->ufid));
1785 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
1787 struct ds ds = DS_EMPTY_INITIALIZER;
1789 match.flow = flow->flow;
1790 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
1792 ds_put_cstr(&ds, "flow_add: ");
1793 odp_format_ufid(ufid, &ds);
1794 ds_put_cstr(&ds, " ");
1795 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
1796 ds_put_cstr(&ds, ", actions:");
1797 format_odp_actions(&ds, actions, actions_len);
1799 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
1808 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
1810 struct dp_netdev *dp = get_dp_netdev(dpif);
1811 struct dp_netdev_flow *netdev_flow;
1812 struct netdev_flow_key key;
1813 struct dp_netdev_pmd_thread *pmd;
1816 int pmd_id = put->pmd_id == PMD_ID_NULL ? NON_PMD_CORE_ID : put->pmd_id;
1819 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
1823 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
1824 put->mask, put->mask_len,
1825 &match.flow, &match.wc.masks);
1830 pmd = dp_netdev_get_pmd(dp, pmd_id);
1835 /* Must produce a netdev_flow_key for lookup.
1836 * This interface is no longer performance critical, since it is not used
1837 * for upcall processing any more. */
1838 netdev_flow_key_from_flow(&key, &match.flow);
1843 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
1846 ovs_mutex_lock(&pmd->flow_mutex);
1847 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
1849 if (put->flags & DPIF_FP_CREATE) {
1850 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
1852 memset(put->stats, 0, sizeof *put->stats);
1854 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
1864 if (put->flags & DPIF_FP_MODIFY
1865 && flow_equal(&match.flow, &netdev_flow->flow)) {
1866 struct dp_netdev_actions *new_actions;
1867 struct dp_netdev_actions *old_actions;
1869 new_actions = dp_netdev_actions_create(put->actions,
1872 old_actions = dp_netdev_flow_get_actions(netdev_flow);
1873 ovsrcu_set(&netdev_flow->actions, new_actions);
1876 get_dpif_flow_stats(netdev_flow, put->stats);
1878 if (put->flags & DPIF_FP_ZERO_STATS) {
1879 /* XXX: The userspace datapath uses thread local statistics
1880 * (for flows), which should be updated only by the owning
1881 * thread. Since we cannot write on stats memory here,
1882 * we choose not to support this flag. Please note:
1883 * - This feature is currently used only by dpctl commands with
1885 * - Should the need arise, this operation can be implemented
1886 * by keeping a base value (to be update here) for each
1887 * counter, and subtracting it before outputting the stats */
1891 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
1892 } else if (put->flags & DPIF_FP_CREATE) {
1895 /* Overlapping flow. */
1899 ovs_mutex_unlock(&pmd->flow_mutex);
1900 dp_netdev_pmd_unref(pmd);
1906 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
1908 struct dp_netdev *dp = get_dp_netdev(dpif);
1909 struct dp_netdev_flow *netdev_flow;
1910 struct dp_netdev_pmd_thread *pmd;
1911 int pmd_id = del->pmd_id == PMD_ID_NULL ? NON_PMD_CORE_ID : del->pmd_id;
1914 pmd = dp_netdev_get_pmd(dp, pmd_id);
1919 ovs_mutex_lock(&pmd->flow_mutex);
1920 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
1924 get_dpif_flow_stats(netdev_flow, del->stats);
1926 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1930 ovs_mutex_unlock(&pmd->flow_mutex);
1931 dp_netdev_pmd_unref(pmd);
1936 struct dpif_netdev_flow_dump {
1937 struct dpif_flow_dump up;
1938 struct cmap_position poll_thread_pos;
1939 struct cmap_position flow_pos;
1940 struct dp_netdev_pmd_thread *cur_pmd;
1942 struct ovs_mutex mutex;
1945 static struct dpif_netdev_flow_dump *
1946 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
1948 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
1951 static struct dpif_flow_dump *
1952 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
1954 struct dpif_netdev_flow_dump *dump;
1956 dump = xzalloc(sizeof *dump);
1957 dpif_flow_dump_init(&dump->up, dpif_);
1958 dump->up.terse = terse;
1959 ovs_mutex_init(&dump->mutex);
1965 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
1967 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1969 ovs_mutex_destroy(&dump->mutex);
1974 struct dpif_netdev_flow_dump_thread {
1975 struct dpif_flow_dump_thread up;
1976 struct dpif_netdev_flow_dump *dump;
1977 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
1978 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
1981 static struct dpif_netdev_flow_dump_thread *
1982 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
1984 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
1987 static struct dpif_flow_dump_thread *
1988 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
1990 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
1991 struct dpif_netdev_flow_dump_thread *thread;
1993 thread = xmalloc(sizeof *thread);
1994 dpif_flow_dump_thread_init(&thread->up, &dump->up);
1995 thread->dump = dump;
2000 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
2002 struct dpif_netdev_flow_dump_thread *thread
2003 = dpif_netdev_flow_dump_thread_cast(thread_);
2009 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2010 struct dpif_flow *flows, int max_flows)
2012 struct dpif_netdev_flow_dump_thread *thread
2013 = dpif_netdev_flow_dump_thread_cast(thread_);
2014 struct dpif_netdev_flow_dump *dump = thread->dump;
2015 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2019 ovs_mutex_lock(&dump->mutex);
2020 if (!dump->status) {
2021 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2022 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2023 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2024 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2026 /* First call to dump_next(), extracts the first pmd thread.
2027 * If there is no pmd thread, returns immediately. */
2029 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2031 ovs_mutex_unlock(&dump->mutex);
2038 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2039 struct cmap_node *node;
2041 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2045 netdev_flows[n_flows] = CONTAINER_OF(node,
2046 struct dp_netdev_flow,
2049 /* When finishing dumping the current pmd thread, moves to
2051 if (n_flows < flow_limit) {
2052 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2053 dp_netdev_pmd_unref(pmd);
2054 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2060 /* Keeps the reference to next caller. */
2061 dump->cur_pmd = pmd;
2063 /* If the current dump is empty, do not exit the loop, since the
2064 * remaining pmds could have flows to be dumped. Just dumps again
2065 * on the new 'pmd'. */
2068 ovs_mutex_unlock(&dump->mutex);
2070 for (i = 0; i < n_flows; i++) {
2071 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2072 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2073 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2074 struct dpif_flow *f = &flows[i];
2075 struct ofpbuf key, mask;
2077 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2078 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2079 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2087 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2088 OVS_NO_THREAD_SAFETY_ANALYSIS
2090 struct dp_netdev *dp = get_dp_netdev(dpif);
2091 struct dp_netdev_pmd_thread *pmd;
2092 struct dp_packet *pp;
2094 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2095 dp_packet_size(execute->packet) > UINT16_MAX) {
2099 /* Tries finding the 'pmd'. If NULL is returned, that means
2100 * the current thread is a non-pmd thread and should use
2101 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2102 pmd = ovsthread_getspecific(dp->per_pmd_key);
2104 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2107 /* If the current thread is non-pmd thread, acquires
2108 * the 'non_pmd_mutex'. */
2109 if (pmd->core_id == NON_PMD_CORE_ID) {
2110 ovs_mutex_lock(&dp->non_pmd_mutex);
2111 ovs_mutex_lock(&dp->port_mutex);
2114 pp = execute->packet;
2115 dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
2116 execute->actions_len);
2117 if (pmd->core_id == NON_PMD_CORE_ID) {
2118 dp_netdev_pmd_unref(pmd);
2119 ovs_mutex_unlock(&dp->port_mutex);
2120 ovs_mutex_unlock(&dp->non_pmd_mutex);
2127 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2131 for (i = 0; i < n_ops; i++) {
2132 struct dpif_op *op = ops[i];
2135 case DPIF_OP_FLOW_PUT:
2136 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2139 case DPIF_OP_FLOW_DEL:
2140 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2143 case DPIF_OP_EXECUTE:
2144 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2147 case DPIF_OP_FLOW_GET:
2148 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2154 /* Returns true if the configuration for rx queues or cpu mask
2157 pmd_config_changed(const struct dp_netdev *dp, size_t rxqs, const char *cmask)
2159 if (dp->n_dpdk_rxqs != rxqs) {
2162 if (dp->pmd_cmask != NULL && cmask != NULL) {
2163 return strcmp(dp->pmd_cmask, cmask);
2165 return (dp->pmd_cmask != NULL || cmask != NULL);
2170 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2172 dpif_netdev_pmd_set(struct dpif *dpif, unsigned int n_rxqs, const char *cmask)
2174 struct dp_netdev *dp = get_dp_netdev(dpif);
2176 if (pmd_config_changed(dp, n_rxqs, cmask)) {
2177 struct dp_netdev_port *port;
2179 dp_netdev_destroy_all_pmds(dp);
2181 CMAP_FOR_EACH (port, node, &dp->ports) {
2182 if (netdev_is_pmd(port->netdev)) {
2185 /* Closes the existing 'rxq's. */
2186 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2187 netdev_rxq_close(port->rxq[i]);
2188 port->rxq[i] = NULL;
2191 /* Sets the new rx queue config. */
2192 err = netdev_set_multiq(port->netdev, ovs_numa_get_n_cores(),
2194 if (err && (err != EOPNOTSUPP)) {
2195 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2196 " %u", netdev_get_name(port->netdev),
2201 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2202 port->rxq = xrealloc(port->rxq, sizeof *port->rxq
2203 * netdev_n_rxq(port->netdev));
2204 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2205 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2209 dp->n_dpdk_rxqs = n_rxqs;
2211 /* Reconfigures the cpu mask. */
2212 ovs_numa_set_cpu_mask(cmask);
2213 free(dp->pmd_cmask);
2214 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2216 /* Restores the non-pmd. */
2217 dp_netdev_set_nonpmd(dp);
2218 /* Restores all pmd threads. */
2219 dp_netdev_reset_pmd_threads(dp);
2226 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2227 uint32_t queue_id, uint32_t *priority)
2229 *priority = queue_id;
2234 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
2235 * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
2237 struct dp_netdev_actions *
2238 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2240 struct dp_netdev_actions *netdev_actions;
2242 netdev_actions = xmalloc(sizeof *netdev_actions);
2243 netdev_actions->actions = xmemdup(actions, size);
2244 netdev_actions->size = size;
2246 return netdev_actions;
2249 struct dp_netdev_actions *
2250 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2252 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2256 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2258 free(actions->actions);
2264 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2265 struct dp_netdev_port *port,
2266 struct netdev_rxq *rxq)
2268 struct dp_packet *packets[NETDEV_MAX_RX_BATCH];
2271 error = netdev_rxq_recv(rxq, packets, &cnt);
2275 *recirc_depth_get() = 0;
2277 /* XXX: initialize md in netdev implementation. */
2278 for (i = 0; i < cnt; i++) {
2279 packets[i]->md = PKT_METADATA_INITIALIZER(port->port_no);
2281 dp_netdev_input(pmd, packets, cnt);
2282 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2283 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2285 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2286 netdev_get_name(port->netdev), ovs_strerror(error));
2290 /* Return true if needs to revalidate datapath flows. */
2292 dpif_netdev_run(struct dpif *dpif)
2294 struct dp_netdev_port *port;
2295 struct dp_netdev *dp = get_dp_netdev(dpif);
2296 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2298 uint64_t new_tnl_seq;
2300 ovs_mutex_lock(&dp->non_pmd_mutex);
2301 CMAP_FOR_EACH (port, node, &dp->ports) {
2302 if (!netdev_is_pmd(port->netdev)) {
2305 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2306 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2310 ovs_mutex_unlock(&dp->non_pmd_mutex);
2311 dp_netdev_pmd_unref(non_pmd);
2313 tnl_arp_cache_run();
2314 new_tnl_seq = seq_read(tnl_conf_seq);
2316 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2317 dp->last_tnl_conf_seq = new_tnl_seq;
2324 dpif_netdev_wait(struct dpif *dpif)
2326 struct dp_netdev_port *port;
2327 struct dp_netdev *dp = get_dp_netdev(dpif);
2329 ovs_mutex_lock(&dp_netdev_mutex);
2330 CMAP_FOR_EACH (port, node, &dp->ports) {
2331 if (!netdev_is_pmd(port->netdev)) {
2334 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2335 netdev_rxq_wait(port->rxq[i]);
2339 ovs_mutex_unlock(&dp_netdev_mutex);
2340 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2344 struct dp_netdev_port *port;
2345 struct netdev_rxq *rx;
2349 pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
2350 struct rxq_poll **ppoll_list, int poll_cnt)
2352 struct rxq_poll *poll_list = *ppoll_list;
2353 struct dp_netdev_port *port;
2354 int n_pmds_on_numa, index, i;
2356 /* Simple scheduler for netdev rx polling. */
2357 for (i = 0; i < poll_cnt; i++) {
2358 port_unref(poll_list[i].port);
2362 n_pmds_on_numa = get_n_pmd_threads_on_numa(pmd->dp, pmd->numa_id);
2365 CMAP_FOR_EACH (port, node, &pmd->dp->ports) {
2366 /* Calls port_try_ref() to prevent the main thread
2367 * from deleting the port. */
2368 if (port_try_ref(port)) {
2369 if (netdev_is_pmd(port->netdev)
2370 && netdev_get_numa_id(port->netdev) == pmd->numa_id) {
2373 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2374 if ((index % n_pmds_on_numa) == pmd->index) {
2375 poll_list = xrealloc(poll_list,
2376 sizeof *poll_list * (poll_cnt + 1));
2379 poll_list[poll_cnt].port = port;
2380 poll_list[poll_cnt].rx = port->rxq[i];
2386 /* Unrefs the port_try_ref(). */
2391 *ppoll_list = poll_list;
2396 pmd_thread_main(void *f_)
2398 struct dp_netdev_pmd_thread *pmd = f_;
2399 unsigned int lc = 0;
2400 struct rxq_poll *poll_list;
2401 unsigned int port_seq = PMD_INITIAL_SEQ;
2408 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2409 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2410 pmd_thread_setaffinity_cpu(pmd->core_id);
2412 emc_cache_init(&pmd->flow_cache);
2413 poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
2415 /* Signal here to make sure the pmd finishes
2416 * reloading the updated configuration. */
2417 dp_netdev_pmd_reload_done(pmd);
2422 for (i = 0; i < poll_cnt; i++) {
2423 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2431 emc_cache_slow_sweep(&pmd->flow_cache);
2434 atomic_read_relaxed(&pmd->change_seq, &seq);
2435 if (seq != port_seq) {
2442 emc_cache_uninit(&pmd->flow_cache);
2444 if (!latch_is_set(&pmd->exit_latch)){
2448 for (i = 0; i < poll_cnt; i++) {
2449 port_unref(poll_list[i].port);
2452 dp_netdev_pmd_reload_done(pmd);
2459 dp_netdev_disable_upcall(struct dp_netdev *dp)
2460 OVS_ACQUIRES(dp->upcall_rwlock)
2462 fat_rwlock_wrlock(&dp->upcall_rwlock);
2466 dpif_netdev_disable_upcall(struct dpif *dpif)
2467 OVS_NO_THREAD_SAFETY_ANALYSIS
2469 struct dp_netdev *dp = get_dp_netdev(dpif);
2470 dp_netdev_disable_upcall(dp);
2474 dp_netdev_enable_upcall(struct dp_netdev *dp)
2475 OVS_RELEASES(dp->upcall_rwlock)
2477 fat_rwlock_unlock(&dp->upcall_rwlock);
2481 dpif_netdev_enable_upcall(struct dpif *dpif)
2482 OVS_NO_THREAD_SAFETY_ANALYSIS
2484 struct dp_netdev *dp = get_dp_netdev(dpif);
2485 dp_netdev_enable_upcall(dp);
2489 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2491 ovs_mutex_lock(&pmd->cond_mutex);
2492 xpthread_cond_signal(&pmd->cond);
2493 ovs_mutex_unlock(&pmd->cond_mutex);
2496 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2497 * the pointer if succeeds, otherwise, NULL.
2499 * Caller must unrefs the returned reference. */
2500 static struct dp_netdev_pmd_thread *
2501 dp_netdev_get_pmd(struct dp_netdev *dp, int core_id)
2503 struct dp_netdev_pmd_thread *pmd;
2504 const struct cmap_node *pnode;
2506 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2510 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2512 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2515 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2517 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2519 struct dp_netdev_pmd_thread *non_pmd;
2521 non_pmd = xzalloc(sizeof *non_pmd);
2522 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2526 /* Caller must have valid pointer to 'pmd'. */
2528 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2530 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2534 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2536 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2537 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2541 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2542 * fails, keeps checking for next node until reaching the end of cmap.
2544 * Caller must unrefs the returned reference. */
2545 static struct dp_netdev_pmd_thread *
2546 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2548 struct dp_netdev_pmd_thread *next;
2551 struct cmap_node *node;
2553 node = cmap_next_position(&dp->poll_threads, pos);
2554 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2556 } while (next && !dp_netdev_pmd_try_ref(next));
2561 /* Configures the 'pmd' based on the input argument. */
2563 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2564 int index, int core_id, int numa_id)
2568 pmd->core_id = core_id;
2569 pmd->numa_id = numa_id;
2571 ovs_refcount_init(&pmd->ref_cnt);
2572 latch_init(&pmd->exit_latch);
2573 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2574 xpthread_cond_init(&pmd->cond, NULL);
2575 ovs_mutex_init(&pmd->cond_mutex);
2576 ovs_mutex_init(&pmd->flow_mutex);
2577 dpcls_init(&pmd->cls);
2578 cmap_init(&pmd->flow_table);
2579 /* init the 'flow_cache' since there is no
2580 * actual thread created for NON_PMD_CORE_ID. */
2581 if (core_id == NON_PMD_CORE_ID) {
2582 emc_cache_init(&pmd->flow_cache);
2584 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2585 hash_int(core_id, 0));
2589 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2591 dp_netdev_pmd_flow_flush(pmd);
2592 dpcls_destroy(&pmd->cls);
2593 cmap_destroy(&pmd->flow_table);
2594 ovs_mutex_destroy(&pmd->flow_mutex);
2595 latch_destroy(&pmd->exit_latch);
2596 xpthread_cond_destroy(&pmd->cond);
2597 ovs_mutex_destroy(&pmd->cond_mutex);
2601 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2602 * and unrefs the struct. */
2604 dp_netdev_del_pmd(struct dp_netdev_pmd_thread *pmd)
2606 /* Uninit the 'flow_cache' since there is
2607 * no actual thread uninit it for NON_PMD_CORE_ID. */
2608 if (pmd->core_id == NON_PMD_CORE_ID) {
2609 emc_cache_uninit(&pmd->flow_cache);
2611 latch_set(&pmd->exit_latch);
2612 dp_netdev_reload_pmd__(pmd);
2613 ovs_numa_unpin_core(pmd->core_id);
2614 xpthread_join(pmd->thread, NULL);
2616 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2617 dp_netdev_pmd_unref(pmd);
2620 /* Destroys all pmd threads. */
2622 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2624 struct dp_netdev_pmd_thread *pmd;
2626 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2627 dp_netdev_del_pmd(pmd);
2631 /* Deletes all pmd threads on numa node 'numa_id'. */
2633 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2635 struct dp_netdev_pmd_thread *pmd;
2637 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2638 if (pmd->numa_id == numa_id) {
2639 dp_netdev_del_pmd(pmd);
2644 /* Checks the numa node id of 'netdev' and starts pmd threads for
2647 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2651 if (!ovs_numa_numa_id_is_valid(numa_id)) {
2652 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
2653 "invalid", numa_id);
2657 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
2659 /* If there are already pmd threads created for the numa node
2660 * in which 'netdev' is on, do nothing. Else, creates the
2661 * pmd threads for the numa node. */
2663 int can_have, n_unpinned, i;
2665 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
2667 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
2668 "cores on numa node");
2672 /* If cpu mask is specified, uses all unpinned cores, otherwise
2673 * tries creating NR_PMD_THREADS pmd threads. */
2674 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
2675 for (i = 0; i < can_have; i++) {
2676 struct dp_netdev_pmd_thread *pmd = xzalloc(sizeof *pmd);
2677 int core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
2679 dp_netdev_configure_pmd(pmd, dp, i, core_id, numa_id);
2680 /* Each thread will distribute all devices rx-queues among
2682 pmd->thread = ovs_thread_create("pmd", pmd_thread_main, pmd);
2684 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
2689 /* Called after pmd threads config change. Restarts pmd threads with
2690 * new configuration. */
2692 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
2694 struct dp_netdev_port *port;
2696 CMAP_FOR_EACH (port, node, &dp->ports) {
2697 if (netdev_is_pmd(port->netdev)) {
2698 int numa_id = netdev_get_numa_id(port->netdev);
2700 dp_netdev_set_pmds_on_numa(dp, numa_id);
2706 dpif_netdev_get_datapath_version(void)
2708 return xstrdup("<built-in>");
2712 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
2715 long long now = time_msec();
2718 atomic_store_relaxed(&netdev_flow->stats.used, now);
2719 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
2720 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
2721 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
2723 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
2727 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
2728 enum dp_stat_type type, int cnt)
2730 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
2734 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
2735 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
2736 enum dpif_upcall_type type, const struct nlattr *userdata,
2737 struct ofpbuf *actions, struct ofpbuf *put_actions)
2739 struct dp_netdev *dp = pmd->dp;
2741 if (OVS_UNLIKELY(!dp->upcall_cb)) {
2745 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
2746 struct ds ds = DS_EMPTY_INITIALIZER;
2750 ofpbuf_init(&key, 0);
2751 odp_flow_key_from_flow(&key, flow, &wc->masks, flow->in_port.odp_port,
2753 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
2754 dp_packet_size(packet_));
2756 odp_flow_key_format(key.data, key.size, &ds);
2758 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
2759 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
2761 ofpbuf_uninit(&key);
2767 return dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
2768 actions, wc, put_actions, dp->upcall_aux);
2771 static inline uint32_t
2772 dpif_netdev_packet_get_dp_hash(struct dp_packet *packet,
2773 const struct miniflow *mf)
2777 hash = dp_packet_get_dp_hash(packet);
2778 if (OVS_UNLIKELY(!hash)) {
2779 hash = miniflow_hash_5tuple(mf, 0);
2780 dp_packet_set_dp_hash(packet, hash);
2785 struct packet_batch {
2786 unsigned int packet_count;
2787 unsigned int byte_count;
2790 struct dp_netdev_flow *flow;
2792 struct dp_packet *packets[NETDEV_MAX_RX_BATCH];
2796 packet_batch_update(struct packet_batch *batch, struct dp_packet *packet,
2797 const struct miniflow *mf)
2799 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
2800 batch->packets[batch->packet_count++] = packet;
2801 batch->byte_count += dp_packet_size(packet);
2805 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
2809 batch->packet_count = 0;
2810 batch->byte_count = 0;
2811 batch->tcp_flags = 0;
2815 packet_batch_execute(struct packet_batch *batch,
2816 struct dp_netdev_pmd_thread *pmd,
2817 enum dp_stat_type hit_type)
2819 struct dp_netdev_actions *actions;
2820 struct dp_netdev_flow *flow = batch->flow;
2822 dp_netdev_flow_used(batch->flow, batch->packet_count, batch->byte_count,
2825 actions = dp_netdev_flow_get_actions(flow);
2827 dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
2828 actions->actions, actions->size);
2830 dp_netdev_count_packet(pmd, hit_type, batch->packet_count);
2834 dp_netdev_queue_batches(struct dp_packet *pkt,
2835 struct dp_netdev_flow *flow, const struct miniflow *mf,
2836 struct packet_batch *batches, size_t *n_batches,
2839 struct packet_batch *batch = NULL;
2842 if (OVS_UNLIKELY(!flow)) {
2845 /* XXX: This O(n^2) algortihm makes sense if we're operating under the
2846 * assumption that the number of distinct flows (and therefore the
2847 * number of distinct batches) is quite small. If this turns out not
2848 * to be the case, it may make sense to pre sort based on the
2849 * netdev_flow pointer. That done we can get the appropriate batching
2850 * in O(n * log(n)) instead. */
2851 for (j = *n_batches - 1; j >= 0; j--) {
2852 if (batches[j].flow == flow) {
2853 batch = &batches[j];
2854 packet_batch_update(batch, pkt, mf);
2858 if (OVS_UNLIKELY(*n_batches >= max_batches)) {
2862 batch = &batches[(*n_batches)++];
2863 packet_batch_init(batch, flow);
2864 packet_batch_update(batch, pkt, mf);
2869 dp_packet_swap(struct dp_packet **a, struct dp_packet **b)
2871 struct dp_packet *tmp = *a;
2876 /* Try to process all ('cnt') the 'packets' using only the exact match cache
2877 * 'flow_cache'. If a flow is not found for a packet 'packets[i]', or if there
2878 * is no matching batch for a packet's flow, the miniflow is copied into 'keys'
2879 * and the packet pointer is moved at the beginning of the 'packets' array.
2881 * The function returns the number of packets that needs to be processed in the
2882 * 'packets' array (they have been moved to the beginning of the vector).
2884 static inline size_t
2885 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet **packets,
2886 size_t cnt, struct netdev_flow_key *keys)
2888 struct netdev_flow_key key;
2889 struct packet_batch batches[4];
2890 struct emc_cache *flow_cache = &pmd->flow_cache;
2891 size_t n_batches, i;
2892 size_t notfound_cnt = 0;
2895 miniflow_initialize(&key.mf, key.buf);
2896 for (i = 0; i < cnt; i++) {
2897 struct dp_netdev_flow *flow;
2899 if (OVS_UNLIKELY(dp_packet_size(packets[i]) < ETH_HEADER_LEN)) {
2900 dp_packet_delete(packets[i]);
2904 miniflow_extract(packets[i], &key.mf);
2905 key.len = 0; /* Not computed yet. */
2906 key.hash = dpif_netdev_packet_get_dp_hash(packets[i], &key.mf);
2908 flow = emc_lookup(flow_cache, &key);
2909 if (OVS_UNLIKELY(!dp_netdev_queue_batches(packets[i], flow, &key.mf,
2910 batches, &n_batches,
2911 ARRAY_SIZE(batches)))) {
2912 if (i != notfound_cnt) {
2913 dp_packet_swap(&packets[i], &packets[notfound_cnt]);
2916 keys[notfound_cnt++] = key;
2920 for (i = 0; i < n_batches; i++) {
2921 packet_batch_execute(&batches[i], pmd, DP_STAT_EXACT_HIT);
2924 return notfound_cnt;
2928 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
2929 struct dp_packet **packets, size_t cnt,
2930 struct netdev_flow_key *keys)
2932 #if !defined(__CHECKER__) && !defined(_WIN32)
2933 const size_t PKT_ARRAY_SIZE = cnt;
2935 /* Sparse or MSVC doesn't like variable length array. */
2936 enum { PKT_ARRAY_SIZE = NETDEV_MAX_RX_BATCH };
2938 struct packet_batch batches[PKT_ARRAY_SIZE];
2939 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
2940 struct dp_netdev *dp = pmd->dp;
2941 struct emc_cache *flow_cache = &pmd->flow_cache;
2942 size_t n_batches, i;
2945 for (i = 0; i < cnt; i++) {
2946 /* Key length is needed in all the cases, hash computed on demand. */
2947 keys[i].len = netdev_flow_key_size(count_1bits(keys[i].mf.map));
2949 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
2950 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
2951 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
2952 struct ofpbuf actions, put_actions;
2953 int miss_cnt = 0, lost_cnt = 0;
2956 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
2957 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
2959 for (i = 0; i < cnt; i++) {
2960 struct dp_netdev_flow *netdev_flow;
2961 struct ofpbuf *add_actions;
2965 if (OVS_LIKELY(rules[i])) {
2969 /* It's possible that an earlier slow path execution installed
2970 * a rule covering this flow. In this case, it's a lot cheaper
2971 * to catch it here than execute a miss. */
2972 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
2974 rules[i] = &netdev_flow->cr;
2980 miniflow_expand(&keys[i].mf, &match.flow);
2982 ofpbuf_clear(&actions);
2983 ofpbuf_clear(&put_actions);
2985 dpif_flow_hash(dp->dpif, &match.flow, sizeof match.flow, &ufid);
2986 error = dp_netdev_upcall(pmd, packets[i], &match.flow, &match.wc,
2987 &ufid, DPIF_UC_MISS, NULL, &actions,
2989 if (OVS_UNLIKELY(error && error != ENOSPC)) {
2990 dp_packet_delete(packets[i]);
2995 /* We can't allow the packet batching in the next loop to execute
2996 * the actions. Otherwise, if there are any slow path actions,
2997 * we'll send the packet up twice. */
2998 dp_netdev_execute_actions(pmd, &packets[i], 1, true,
2999 actions.data, actions.size);
3001 add_actions = put_actions.size ? &put_actions : &actions;
3002 if (OVS_LIKELY(error != ENOSPC)) {
3003 /* XXX: There's a race window where a flow covering this packet
3004 * could have already been installed since we last did the flow
3005 * lookup before upcall. This could be solved by moving the
3006 * mutex lock outside the loop, but that's an awful long time
3007 * to be locking everyone out of making flow installs. If we
3008 * move to a per-core classifier, it would be reasonable. */
3009 ovs_mutex_lock(&pmd->flow_mutex);
3010 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3011 if (OVS_LIKELY(!netdev_flow)) {
3012 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3016 ovs_mutex_unlock(&pmd->flow_mutex);
3018 emc_insert(flow_cache, &keys[i], netdev_flow);
3022 ofpbuf_uninit(&actions);
3023 ofpbuf_uninit(&put_actions);
3024 fat_rwlock_unlock(&dp->upcall_rwlock);
3025 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3026 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3027 } else if (OVS_UNLIKELY(any_miss)) {
3028 int dropped_cnt = 0;
3030 for (i = 0; i < cnt; i++) {
3031 if (OVS_UNLIKELY(!rules[i])) {
3032 dp_packet_delete(packets[i]);
3037 dp_netdev_count_packet(pmd, DP_STAT_MISS, dropped_cnt);
3038 dp_netdev_count_packet(pmd, DP_STAT_LOST, dropped_cnt);
3042 for (i = 0; i < cnt; i++) {
3043 struct dp_packet *packet = packets[i];
3044 struct dp_netdev_flow *flow;
3046 if (OVS_UNLIKELY(!rules[i])) {
3050 flow = dp_netdev_flow_cast(rules[i]);
3052 emc_insert(flow_cache, &keys[i], flow);
3053 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches,
3054 &n_batches, ARRAY_SIZE(batches));
3057 for (i = 0; i < n_batches; i++) {
3058 packet_batch_execute(&batches[i], pmd, DP_STAT_MASKED_HIT);
3063 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3064 struct dp_packet **packets, int cnt)
3066 #if !defined(__CHECKER__) && !defined(_WIN32)
3067 const size_t PKT_ARRAY_SIZE = cnt;
3069 /* Sparse or MSVC doesn't like variable length array. */
3070 enum { PKT_ARRAY_SIZE = NETDEV_MAX_RX_BATCH };
3072 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3075 newcnt = emc_processing(pmd, packets, cnt, keys);
3076 if (OVS_UNLIKELY(newcnt)) {
3077 fast_path_processing(pmd, packets, newcnt, keys);
3081 struct dp_netdev_execute_aux {
3082 struct dp_netdev_pmd_thread *pmd;
3086 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3089 struct dp_netdev *dp = get_dp_netdev(dpif);
3090 dp->upcall_aux = aux;
3095 dp_netdev_drop_packets(struct dp_packet ** packets, int cnt, bool may_steal)
3100 for (i = 0; i < cnt; i++) {
3101 dp_packet_delete(packets[i]);
3107 push_tnl_action(const struct dp_netdev *dp,
3108 const struct nlattr *attr,
3109 struct dp_packet **packets, int cnt)
3111 struct dp_netdev_port *tun_port;
3112 const struct ovs_action_push_tnl *data;
3114 data = nl_attr_get(attr);
3116 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3120 netdev_push_header(tun_port->netdev, packets, cnt, data);
3126 dp_netdev_clone_pkt_batch(struct dp_packet **tnl_pkt,
3127 struct dp_packet **packets, int cnt)
3131 for (i = 0; i < cnt; i++) {
3132 tnl_pkt[i] = dp_packet_clone(packets[i]);
3137 dp_execute_cb(void *aux_, struct dp_packet **packets, int cnt,
3138 const struct nlattr *a, bool may_steal)
3139 OVS_NO_THREAD_SAFETY_ANALYSIS
3141 struct dp_netdev_execute_aux *aux = aux_;
3142 uint32_t *depth = recirc_depth_get();
3143 struct dp_netdev_pmd_thread *pmd= aux->pmd;
3144 struct dp_netdev *dp= pmd->dp;
3145 int type = nl_attr_type(a);
3146 struct dp_netdev_port *p;
3149 switch ((enum ovs_action_attr)type) {
3150 case OVS_ACTION_ATTR_OUTPUT:
3151 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3152 if (OVS_LIKELY(p)) {
3153 netdev_send(p->netdev, pmd->core_id, packets, cnt, may_steal);
3158 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3159 if (*depth < MAX_RECIRC_DEPTH) {
3160 struct dp_packet *tnl_pkt[NETDEV_MAX_RX_BATCH];
3164 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3168 err = push_tnl_action(dp, a, packets, cnt);
3171 dp_netdev_input(pmd, packets, cnt);
3174 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3180 case OVS_ACTION_ATTR_TUNNEL_POP:
3181 if (*depth < MAX_RECIRC_DEPTH) {
3182 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3184 p = dp_netdev_lookup_port(dp, portno);
3186 struct dp_packet *tnl_pkt[NETDEV_MAX_RX_BATCH];
3190 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3194 err = netdev_pop_header(p->netdev, packets, cnt);
3197 for (i = 0; i < cnt; i++) {
3198 packets[i]->md.in_port.odp_port = portno;
3202 dp_netdev_input(pmd, packets, cnt);
3205 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3212 case OVS_ACTION_ATTR_USERSPACE:
3213 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3214 const struct nlattr *userdata;
3215 struct ofpbuf actions;
3219 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3220 ofpbuf_init(&actions, 0);
3222 for (i = 0; i < cnt; i++) {
3225 ofpbuf_clear(&actions);
3227 flow_extract(packets[i], &flow);
3228 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3229 error = dp_netdev_upcall(pmd, packets[i], &flow, NULL, &ufid,
3230 DPIF_UC_ACTION, userdata,&actions,
3232 if (!error || error == ENOSPC) {
3233 dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
3234 actions.data, actions.size);
3235 } else if (may_steal) {
3236 dp_packet_delete(packets[i]);
3239 ofpbuf_uninit(&actions);
3240 fat_rwlock_unlock(&dp->upcall_rwlock);
3246 case OVS_ACTION_ATTR_RECIRC:
3247 if (*depth < MAX_RECIRC_DEPTH) {
3250 for (i = 0; i < cnt; i++) {
3251 struct dp_packet *recirc_pkt;
3253 recirc_pkt = (may_steal) ? packets[i]
3254 : dp_packet_clone(packets[i]);
3256 recirc_pkt->md.recirc_id = nl_attr_get_u32(a);
3258 /* Hash is private to each packet */
3259 recirc_pkt->md.dp_hash = dp_packet_get_dp_hash(packets[i]);
3261 dp_netdev_input(pmd, &recirc_pkt, 1);
3268 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3271 case OVS_ACTION_ATTR_PUSH_VLAN:
3272 case OVS_ACTION_ATTR_POP_VLAN:
3273 case OVS_ACTION_ATTR_PUSH_MPLS:
3274 case OVS_ACTION_ATTR_POP_MPLS:
3275 case OVS_ACTION_ATTR_SET:
3276 case OVS_ACTION_ATTR_SET_MASKED:
3277 case OVS_ACTION_ATTR_SAMPLE:
3278 case OVS_ACTION_ATTR_HASH:
3279 case OVS_ACTION_ATTR_UNSPEC:
3280 case __OVS_ACTION_ATTR_MAX:
3284 dp_netdev_drop_packets(packets, cnt, may_steal);
3288 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3289 struct dp_packet **packets, int cnt,
3291 const struct nlattr *actions, size_t actions_len)
3293 struct dp_netdev_execute_aux aux = { pmd };
3295 odp_execute_actions(&aux, packets, cnt, may_steal, actions,
3296 actions_len, dp_execute_cb);
3299 const struct dpif_class dpif_netdev_class = {
3301 dpif_netdev_enumerate,
3302 dpif_netdev_port_open_type,
3305 dpif_netdev_destroy,
3308 dpif_netdev_get_stats,
3309 dpif_netdev_port_add,
3310 dpif_netdev_port_del,
3311 dpif_netdev_port_query_by_number,
3312 dpif_netdev_port_query_by_name,
3313 NULL, /* port_get_pid */
3314 dpif_netdev_port_dump_start,
3315 dpif_netdev_port_dump_next,
3316 dpif_netdev_port_dump_done,
3317 dpif_netdev_port_poll,
3318 dpif_netdev_port_poll_wait,
3319 dpif_netdev_flow_flush,
3320 dpif_netdev_flow_dump_create,
3321 dpif_netdev_flow_dump_destroy,
3322 dpif_netdev_flow_dump_thread_create,
3323 dpif_netdev_flow_dump_thread_destroy,
3324 dpif_netdev_flow_dump_next,
3325 dpif_netdev_operate,
3326 NULL, /* recv_set */
3327 NULL, /* handlers_set */
3328 dpif_netdev_pmd_set,
3329 dpif_netdev_queue_to_priority,
3331 NULL, /* recv_wait */
3332 NULL, /* recv_purge */
3333 dpif_netdev_register_upcall_cb,
3334 dpif_netdev_enable_upcall,
3335 dpif_netdev_disable_upcall,
3336 dpif_netdev_get_datapath_version,
3340 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3341 const char *argv[], void *aux OVS_UNUSED)
3343 struct dp_netdev_port *old_port;
3344 struct dp_netdev_port *new_port;
3345 struct dp_netdev *dp;
3348 ovs_mutex_lock(&dp_netdev_mutex);
3349 dp = shash_find_data(&dp_netdevs, argv[1]);
3350 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3351 ovs_mutex_unlock(&dp_netdev_mutex);
3352 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3355 ovs_refcount_ref(&dp->ref_cnt);
3356 ovs_mutex_unlock(&dp_netdev_mutex);
3358 ovs_mutex_lock(&dp->port_mutex);
3359 if (get_port_by_name(dp, argv[2], &old_port)) {
3360 unixctl_command_reply_error(conn, "unknown port");
3364 port_no = u32_to_odp(atoi(argv[3]));
3365 if (!port_no || port_no == ODPP_NONE) {
3366 unixctl_command_reply_error(conn, "bad port number");
3369 if (dp_netdev_lookup_port(dp, port_no)) {
3370 unixctl_command_reply_error(conn, "port number already in use");
3374 /* Remove old port. */
3375 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
3376 ovsrcu_postpone(free, old_port);
3378 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3379 new_port = xmemdup(old_port, sizeof *old_port);
3380 new_port->port_no = port_no;
3381 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3383 seq_change(dp->port_seq);
3384 unixctl_command_reply(conn, NULL);
3387 ovs_mutex_unlock(&dp->port_mutex);
3388 dp_netdev_unref(dp);
3392 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
3393 const char *argv[], void *aux OVS_UNUSED)
3395 struct dp_netdev_port *port;
3396 struct dp_netdev *dp;
3398 ovs_mutex_lock(&dp_netdev_mutex);
3399 dp = shash_find_data(&dp_netdevs, argv[1]);
3400 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3401 ovs_mutex_unlock(&dp_netdev_mutex);
3402 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3405 ovs_refcount_ref(&dp->ref_cnt);
3406 ovs_mutex_unlock(&dp_netdev_mutex);
3408 ovs_mutex_lock(&dp->port_mutex);
3409 if (get_port_by_name(dp, argv[2], &port)) {
3410 unixctl_command_reply_error(conn, "unknown port");
3411 } else if (port->port_no == ODPP_LOCAL) {
3412 unixctl_command_reply_error(conn, "can't delete local port");
3414 do_del_port(dp, port);
3415 unixctl_command_reply(conn, NULL);
3417 ovs_mutex_unlock(&dp->port_mutex);
3419 dp_netdev_unref(dp);
3423 dpif_dummy_register__(const char *type)
3425 struct dpif_class *class;
3427 class = xmalloc(sizeof *class);
3428 *class = dpif_netdev_class;
3429 class->type = xstrdup(type);
3430 dp_register_provider(class);
3434 dpif_dummy_register(bool override)
3441 dp_enumerate_types(&types);
3442 SSET_FOR_EACH (type, &types) {
3443 if (!dp_unregister_provider(type)) {
3444 dpif_dummy_register__(type);
3447 sset_destroy(&types);
3450 dpif_dummy_register__("dummy");
3452 unixctl_command_register("dpif-dummy/change-port-number",
3453 "dp port new-number",
3454 3, 3, dpif_dummy_change_port_number, NULL);
3455 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3456 2, 2, dpif_dummy_delete_port, NULL);
3459 /* Datapath Classifier. */
3461 /* A set of rules that all have the same fields wildcarded. */
3462 struct dpcls_subtable {
3463 /* The fields are only used by writers. */
3464 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
3466 /* These fields are accessed by readers. */
3467 struct cmap rules; /* Contains "struct dpcls_rule"s. */
3468 struct netdev_flow_key mask; /* Wildcards for fields (const). */
3469 /* 'mask' must be the last field, additional space is allocated here. */
3472 /* Initializes 'cls' as a classifier that initially contains no classification
3475 dpcls_init(struct dpcls *cls)
3477 cmap_init(&cls->subtables_map);
3478 pvector_init(&cls->subtables);
3482 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
3484 pvector_remove(&cls->subtables, subtable);
3485 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
3486 subtable->mask.hash);
3487 cmap_destroy(&subtable->rules);
3488 ovsrcu_postpone(free, subtable);
3491 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
3492 * caller's responsibility.
3493 * May only be called after all the readers have been terminated. */
3495 dpcls_destroy(struct dpcls *cls)
3498 struct dpcls_subtable *subtable;
3500 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
3501 dpcls_destroy_subtable(cls, subtable);
3503 cmap_destroy(&cls->subtables_map);
3504 pvector_destroy(&cls->subtables);
3508 static struct dpcls_subtable *
3509 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3511 struct dpcls_subtable *subtable;
3513 /* Need to add one. */
3514 subtable = xmalloc(sizeof *subtable
3515 - sizeof subtable->mask.mf + mask->len);
3516 cmap_init(&subtable->rules);
3517 netdev_flow_key_clone(&subtable->mask, mask);
3518 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
3519 pvector_insert(&cls->subtables, subtable, 0);
3520 pvector_publish(&cls->subtables);
3525 static inline struct dpcls_subtable *
3526 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3528 struct dpcls_subtable *subtable;
3530 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
3531 &cls->subtables_map) {
3532 if (netdev_flow_key_equal(&subtable->mask, mask)) {
3536 return dpcls_create_subtable(cls, mask);
3539 /* Insert 'rule' into 'cls'. */
3541 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
3542 const struct netdev_flow_key *mask)
3544 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
3546 rule->mask = &subtable->mask;
3547 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
3550 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
3552 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
3554 struct dpcls_subtable *subtable;
3556 ovs_assert(rule->mask);
3558 INIT_CONTAINER(subtable, rule->mask, mask);
3560 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
3562 dpcls_destroy_subtable(cls, subtable);
3563 pvector_publish(&cls->subtables);
3567 /* Returns true if 'target' satisifies 'key' in 'mask', that is, if each 1-bit
3568 * in 'mask' the values in 'key' and 'target' are the same.
3570 * Note: 'key' and 'mask' have the same mask, and 'key' is already masked. */
3572 dpcls_rule_matches_key(const struct dpcls_rule *rule,
3573 const struct netdev_flow_key *target)
3575 const uint64_t *keyp = rule->flow.mf.inline_values;
3576 const uint64_t *maskp = rule->mask->mf.inline_values;
3577 uint64_t target_u64;
3579 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(target_u64, target, rule->flow.mf.map) {
3580 if (OVS_UNLIKELY((target_u64 & *maskp++) != *keyp++)) {
3587 /* For each miniflow in 'flows' performs a classifier lookup writing the result
3588 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
3589 * NULL it is skipped.
3591 * This function is optimized for use in the userspace datapath and therefore
3592 * does not implement a lot of features available in the standard
3593 * classifier_lookup() function. Specifically, it does not implement
3594 * priorities, instead returning any rule which matches the flow.
3596 * Returns true if all flows found a corresponding rule. */
3598 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
3599 struct dpcls_rule **rules, const size_t cnt)
3601 /* The batch size 16 was experimentally found faster than 8 or 32. */
3602 typedef uint16_t map_type;
3603 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
3605 #if !defined(__CHECKER__) && !defined(_WIN32)
3606 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
3608 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_RX_BATCH, MAP_BITS) };
3610 map_type maps[N_MAPS];
3611 struct dpcls_subtable *subtable;
3613 memset(maps, 0xff, sizeof maps);
3614 if (cnt % MAP_BITS) {
3615 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
3617 memset(rules, 0, cnt * sizeof *rules);
3619 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
3620 const struct netdev_flow_key *mkeys = keys;
3621 struct dpcls_rule **mrules = rules;
3622 map_type remains = 0;
3625 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
3627 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
3628 uint32_t hashes[MAP_BITS];
3629 const struct cmap_node *nodes[MAP_BITS];
3630 unsigned long map = maps[m];
3634 continue; /* Skip empty maps. */
3637 /* Compute hashes for the remaining keys. */
3638 ULONG_FOR_EACH_1(i, map) {
3639 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
3643 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
3644 /* Check results. */
3645 ULONG_FOR_EACH_1(i, map) {
3646 struct dpcls_rule *rule;
3648 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
3649 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
3654 ULONG_SET0(map, i); /* Did not match. */
3656 ; /* Keep Sparse happy. */
3658 maps[m] &= ~map; /* Clear the found rules. */
3662 return true; /* All found. */
3665 return false; /* Some misses. */