lib/mac-learning.c

   1 /*
   2  * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013, 2014, 2015 Nicira, Inc.
   3  *
   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:
   7  *
   8  *     http://www.apache.org/licenses/LICENSE-2.0
   9  *
  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.
  15  */
  16
  17 #include <config.h>
  18 #include "mac-learning.h"
  19
  20 #include <inttypes.h>
  21 #include <stdlib.h>
  22
  23 #include "bitmap.h"
  24 #include "coverage.h"
  25 #include "hash.h"
  26 #include "openvswitch/list.h"
  27 #include "openvswitch/poll-loop.h"
  28 #include "timeval.h"
  29 #include "unaligned.h"
  30 #include "util.h"
  31 #include "vlan-bitmap.h"
  32
  33 COVERAGE_DEFINE(mac_learning_learned);
  34 COVERAGE_DEFINE(mac_learning_expired);
  35 COVERAGE_DEFINE(mac_learning_evicted);
  36 COVERAGE_DEFINE(mac_learning_moved);
  37
  38 /* Returns the number of seconds since 'e' (within 'ml') was last learned. */
  39 int
  40 mac_entry_age(const struct mac_learning *ml, const struct mac_entry *e)
  41 {
  42     time_t remaining = e->expires - time_now();
  43     return ml->idle_time - remaining;
  44 }
  45
  46 static uint32_t
  47 mac_table_hash(const struct mac_learning *ml, const struct eth_addr mac,
  48                uint16_t vlan)
  49 {
  50     return hash_mac(mac, vlan, ml->secret);
  51 }
  52
  53 static struct mac_entry *
  54 mac_entry_from_lru_node(struct ovs_list *list)
  55 {
  56     return CONTAINER_OF(list, struct mac_entry, lru_node);
  57 }
  58
  59 static struct mac_entry *
  60 mac_entry_lookup(const struct mac_learning *ml,
  61                  const struct eth_addr mac, uint16_t vlan)
  62 {
  63     struct mac_entry *e;
  64
  65     HMAP_FOR_EACH_WITH_HASH (e, hmap_node, mac_table_hash(ml, mac, vlan),
  66                              &ml->table) {
  67         if (e->vlan == vlan && eth_addr_equals(e->mac, mac)) {
  68             return e;
  69         }
  70     }
  71     return NULL;
  72 }
  73
  74 static struct mac_learning_port *
  75 mac_learning_port_lookup(struct mac_learning *ml, void *port)
  76 {
  77     struct mac_learning_port *mlport;
  78
  79     HMAP_FOR_EACH_IN_BUCKET (mlport, hmap_node, hash_pointer(port, ml->secret),
  80                              &ml->ports_by_ptr) {
  81         if (mlport->port == port) {
  82             return mlport;
  83         }
  84     }
  85     return NULL;
  86 }
  87
  88 /* Changes the client-owned pointer for entry 'e' in 'ml' to 'port'.  The
  89  * pointer can be retrieved with mac_entry_get_port().
  90  *
  91  * The MAC-learning implementation treats the data that 'port' points to as
  92  * opaque and never tries to dereference it.  However, when a MAC learning
  93  * table becomes overfull, so that eviction is required, the implementation
  94  * does first evict MAC entries for the most common 'port's values in 'ml', so
  95  * that there is a degree of fairness, that is, each port is entitled to its
  96  * fair share of MAC entries. */
  97 void
  98 mac_entry_set_port(struct mac_learning *ml, struct mac_entry *e, void *port)
  99     OVS_REQ_WRLOCK(ml->rwlock)
 100 {
 101     if (mac_entry_get_port(ml, e) != port) {
 102         ml->need_revalidate = true;
 103
 104         if (e->mlport) {
 105             struct mac_learning_port *mlport = e->mlport;
 106             ovs_list_remove(&e->port_lru_node);
 107
 108             if (ovs_list_is_empty(&mlport->port_lrus)) {
 109                 ovs_assert(mlport->heap_node.priority == 1);
 110                 hmap_remove(&ml->ports_by_ptr, &mlport->hmap_node);
 111                 heap_remove(&ml->ports_by_usage, &mlport->heap_node);
 112                 free(mlport);
 113             } else {
 114                 ovs_assert(mlport->heap_node.priority > 1);
 115                 heap_change(&ml->ports_by_usage, &mlport->heap_node,
 116                             mlport->heap_node.priority - 1);
 117             }
 118             e->mlport = NULL;
 119         }
 120
 121         if (port) {
 122             struct mac_learning_port *mlport;
 123
 124             mlport = mac_learning_port_lookup(ml, port);
 125             if (!mlport) {
 126                 mlport = xzalloc(sizeof *mlport);
 127                 hmap_insert(&ml->ports_by_ptr, &mlport->hmap_node,
 128                             hash_pointer(port, ml->secret));
 129                 heap_insert(&ml->ports_by_usage, &mlport->heap_node, 1);
 130                 mlport->port = port;
 131                 ovs_list_init(&mlport->port_lrus);
 132             } else {
 133                 heap_change(&ml->ports_by_usage, &mlport->heap_node,
 134                             mlport->heap_node.priority + 1);
 135             }
 136             ovs_list_push_back(&mlport->port_lrus, &e->port_lru_node);
 137             e->mlport = mlport;
 138         }
 139     }
 140 }
 141
 142 /* Finds one of the ports with the most MAC entries and evicts its least
 143  * recently used entry. */
 144 static void
 145 evict_mac_entry_fairly(struct mac_learning *ml)
 146     OVS_REQ_WRLOCK(ml->rwlock)
 147 {
 148     struct mac_learning_port *mlport;
 149     struct mac_entry *e;
 150
 151     mlport = CONTAINER_OF(heap_max(&ml->ports_by_usage),
 152                           struct mac_learning_port, heap_node);
 153     e = CONTAINER_OF(ovs_list_front(&mlport->port_lrus),
 154                      struct mac_entry, port_lru_node);
 155     COVERAGE_INC(mac_learning_evicted);
 156     mac_learning_expire(ml, e);
 157 }
 158
 159 /* If the LRU list is not empty, stores the least-recently-used entry in '*e'
 160  * and returns true.  Otherwise, if the LRU list is empty, stores NULL in '*e'
 161  * and return false. */
 162 static bool
 163 get_lru(struct mac_learning *ml, struct mac_entry **e)
 164     OVS_REQ_RDLOCK(ml->rwlock)
 165 {
 166     if (!ovs_list_is_empty(&ml->lrus)) {
 167         *e = mac_entry_from_lru_node(ml->lrus.next);
 168         return true;
 169     } else {
 170         *e = NULL;
 171         return false;
 172     }
 173 }
 174
 175 static unsigned int
 176 normalize_idle_time(unsigned int idle_time)
 177 {
 178     return (idle_time < 15 ? 15
 179             : idle_time > 3600 ? 3600
 180             : idle_time);
 181 }
 182
 183 /* Creates and returns a new MAC learning table with an initial MAC aging
 184  * timeout of 'idle_time' seconds and an initial maximum of MAC_DEFAULT_MAX
 185  * entries. */
 186 struct mac_learning *
 187 mac_learning_create(unsigned int idle_time)
 188 {
 189     struct mac_learning *ml;
 190
 191     ml = xmalloc(sizeof *ml);
 192     ovs_list_init(&ml->lrus);
 193     hmap_init(&ml->table);
 194     ml->secret = random_uint32();
 195     ml->flood_vlans = NULL;
 196     ml->idle_time = normalize_idle_time(idle_time);
 197     ml->max_entries = MAC_DEFAULT_MAX;
 198     ml->need_revalidate = false;
 199     hmap_init(&ml->ports_by_ptr);
 200     heap_init(&ml->ports_by_usage);
 201     ovs_refcount_init(&ml->ref_cnt);
 202     ovs_rwlock_init(&ml->rwlock);
 203     return ml;
 204 }
 205
 206 struct mac_learning *
 207 mac_learning_ref(const struct mac_learning *ml_)
 208 {
 209     struct mac_learning *ml = CONST_CAST(struct mac_learning *, ml_);
 210     if (ml) {
 211         ovs_refcount_ref(&ml->ref_cnt);
 212     }
 213     return ml;
 214 }
 215
 216 /* Unreferences (and possibly destroys) MAC learning table 'ml'. */
 217 void
 218 mac_learning_unref(struct mac_learning *ml)
 219 {
 220     if (ml && ovs_refcount_unref(&ml->ref_cnt) == 1) {
 221         struct mac_entry *e, *next;
 222
 223         ovs_rwlock_wrlock(&ml->rwlock);
 224         HMAP_FOR_EACH_SAFE (e, next, hmap_node, &ml->table) {
 225             mac_learning_expire(ml, e);
 226         }
 227         hmap_destroy(&ml->table);
 228         hmap_destroy(&ml->ports_by_ptr);
 229         heap_destroy(&ml->ports_by_usage);
 230
 231         bitmap_free(ml->flood_vlans);
 232         ovs_rwlock_unlock(&ml->rwlock);
 233         ovs_rwlock_destroy(&ml->rwlock);
 234         free(ml);
 235     }
 236 }
 237
 238 /* Provides a bitmap of VLANs which have learning disabled, that is, VLANs on
 239  * which all packets are flooded.  Returns true if the set has changed from the
 240  * previous value. */
 241 bool
 242 mac_learning_set_flood_vlans(struct mac_learning *ml,
 243                              const unsigned long *bitmap)
 244 {
 245     if (vlan_bitmap_equal(ml->flood_vlans, bitmap)) {
 246         return false;
 247     } else {
 248         bitmap_free(ml->flood_vlans);
 249         ml->flood_vlans = vlan_bitmap_clone(bitmap);
 250         return true;
 251     }
 252 }
 253
 254 /* Changes the MAC aging timeout of 'ml' to 'idle_time' seconds. */
 255 void
 256 mac_learning_set_idle_time(struct mac_learning *ml, unsigned int idle_time)
 257 {
 258     idle_time = normalize_idle_time(idle_time);
 259     if (idle_time != ml->idle_time) {
 260         struct mac_entry *e;
 261         int delta;
 262
 263         delta = (int) idle_time - (int) ml->idle_time;
 264         LIST_FOR_EACH (e, lru_node, &ml->lrus) {
 265             e->expires += delta;
 266         }
 267         ml->idle_time = idle_time;
 268     }
 269 }
 270
 271 /* Sets the maximum number of entries in 'ml' to 'max_entries', adjusting it
 272  * to be within a reasonable range. */
 273 void
 274 mac_learning_set_max_entries(struct mac_learning *ml, size_t max_entries)
 275 {
 276     ml->max_entries = (max_entries < 10 ? 10
 277                        : max_entries > 1000 * 1000 ? 1000 * 1000
 278                        : max_entries);
 279 }
 280
 281 static bool
 282 is_learning_vlan(const struct mac_learning *ml, uint16_t vlan)
 283 {
 284     return !ml->flood_vlans || !bitmap_is_set(ml->flood_vlans, vlan);
 285 }
 286
 287 /* Returns true if 'src_mac' may be learned on 'vlan' for 'ml'.
 288  * Returns false if 'ml' is NULL, if src_mac is not valid for learning, or if
 289  * 'vlan' is configured on 'ml' to flood all packets. */
 290 bool
 291 mac_learning_may_learn(const struct mac_learning *ml,
 292                        const struct eth_addr src_mac, uint16_t vlan)
 293 {
 294     return ml && is_learning_vlan(ml, vlan) && !eth_addr_is_multicast(src_mac);
 295 }
 296
 297 /* Searches 'ml' for and returns a MAC learning entry for 'src_mac' in 'vlan',
 298  * inserting a new entry if necessary.  The caller must have already verified,
 299  * by calling mac_learning_may_learn(), that 'src_mac' and 'vlan' are
 300  * learnable.
 301  *
 302  * If the returned MAC entry is new (that is, if it has a NULL client-provided
 303  * port, as returned by mac_entry_get_port()), then the caller must initialize
 304  * the new entry's port to a nonnull value with mac_entry_set_port(). */
 305 struct mac_entry *
 306 mac_learning_insert(struct mac_learning *ml,
 307                     const struct eth_addr src_mac, uint16_t vlan)
 308 {
 309     struct mac_entry *e;
 310
 311     e = mac_entry_lookup(ml, src_mac, vlan);
 312     if (!e) {
 313         uint32_t hash = mac_table_hash(ml, src_mac, vlan);
 314
 315         if (hmap_count(&ml->table) >= ml->max_entries) {
 316             evict_mac_entry_fairly(ml);
 317         }
 318
 319         e = xmalloc(sizeof *e);
 320         hmap_insert(&ml->table, &e->hmap_node, hash);
 321         e->mac = src_mac;
 322         e->vlan = vlan;
 323         e->grat_arp_lock = TIME_MIN;
 324         e->mlport = NULL;
 325         COVERAGE_INC(mac_learning_learned);
 326     } else {
 327         ovs_list_remove(&e->lru_node);
 328     }
 329
 330     /* Mark 'e' as recently used. */
 331     ovs_list_push_back(&ml->lrus, &e->lru_node);
 332     if (e->mlport) {
 333         ovs_list_remove(&e->port_lru_node);
 334         ovs_list_push_back(&e->mlport->port_lrus, &e->port_lru_node);
 335     }
 336     e->expires = time_now() + ml->idle_time;
 337
 338     return e;
 339 }
 340
 341 /* Checks whether a MAC learning update is necessary for MAC learning table
 342  * 'ml' given that a packet matching 'src' was received on 'in_port' in 'vlan',
 343  * and given that the packet was gratuitous ARP if 'is_gratuitous_arp' is
 344  * 'true' and 'in_port' is a bond port if 'is_bond' is 'true'.
 345  *
 346  * Most packets processed through the MAC learning table do not actually
 347  * change it in any way.  This function requires only a read lock on the MAC
 348  * learning table, so it is much cheaper in this common case.
 349  *
 350  * Keep the code here synchronized with that in update_learning_table__()
 351  * below. */
 352 static bool
 353 is_mac_learning_update_needed(const struct mac_learning *ml,
 354                               struct eth_addr src, int vlan,
 355                               bool is_gratuitous_arp, bool is_bond,
 356                               void *in_port)
 357     OVS_REQ_RDLOCK(ml->rwlock)
 358 {
 359     struct mac_entry *mac;
 360
 361     if (!mac_learning_may_learn(ml, src, vlan)) {
 362         return false;
 363     }
 364
 365     mac = mac_learning_lookup(ml, src, vlan);
 366     if (!mac || mac_entry_age(ml, mac)) {
 367         return true;
 368     }
 369
 370     if (is_gratuitous_arp) {
 371         /* We don't want to learn from gratuitous ARP packets that are
 372          * reflected back over bond slaves so we lock the learning table.  For
 373          * more detail, see the bigger comment in update_learning_table__(). */
 374         if (!is_bond) {
 375             return true;   /* Need to set the gratuitous ARP lock. */
 376         } else if (mac_entry_is_grat_arp_locked(mac)) {
 377             return false;
 378         }
 379     }
 380
 381     return mac_entry_get_port(ml, mac) != in_port /* ofbundle */;
 382 }
 383
 384 /* Updates MAC learning table 'ml' given that a packet matching 'src' was
 385  * received on 'in_port' in 'vlan', and given that the packet was gratuitous
 386  * ARP if 'is_gratuitous_arp' is 'true' and 'in_port' is a bond port if
 387  * 'is_bond' is 'true'.
 388  *
 389  * This code repeats all the checks in is_mac_learning_update_needed() because
 390  * the lock was released between there and here and thus the MAC learning state
 391  * could have changed.
 392  *
 393  * Returns 'true' if 'ml' was updated, 'false' otherwise.
 394  *
 395  * Keep the code here synchronized with that in is_mac_learning_update_needed()
 396  * above. */
 397 static bool
 398 update_learning_table__(struct mac_learning *ml, struct eth_addr src,
 399                         int vlan, bool is_gratuitous_arp, bool is_bond,
 400                         void *in_port)
 401     OVS_REQ_WRLOCK(ml->rwlock)
 402 {
 403     struct mac_entry *mac;
 404
 405     if (!mac_learning_may_learn(ml, src, vlan)) {
 406         return false;
 407     }
 408
 409     mac = mac_learning_insert(ml, src, vlan);
 410     if (is_gratuitous_arp) {
 411         /* Gratuitous ARP packets received over non-bond interfaces could be
 412          * reflected back over bond slaves.  We don't want to learn from these
 413          * reflected packets, so we lock each entry for which a gratuitous ARP
 414          * packet was received over a non-bond interface and refrain from
 415          * learning from gratuitous ARP packets that arrive over bond
 416          * interfaces for this entry while the lock is in effect. Refer to the
 417          * 'ovs-vswitch Internals' document for more in-depth discussion on
 418          * this topic. */
 419         if (!is_bond) {
 420             mac_entry_set_grat_arp_lock(mac);
 421         } else if (mac_entry_is_grat_arp_locked(mac)) {
 422             return false;
 423         }
 424     }
 425
 426     if (mac_entry_get_port(ml, mac) != in_port) {
 427         if (mac_entry_get_port(ml, mac) != NULL) {
 428             COVERAGE_INC(mac_learning_moved);
 429         }
 430         mac_entry_set_port(ml, mac, in_port);
 431         return true;
 432     }
 433     return false;
 434 }
 435
 436 /* Updates MAC learning table 'ml' given that a packet matching 'src' was
 437  * received on 'in_port' in 'vlan', and given that the packet was gratuitous
 438  * ARP if 'is_gratuitous_arp' is 'true' and 'in_port' is a bond port if
 439  * 'is_bond' is 'true'.
 440  *
 441  * Returns 'true' if 'ml' was updated, 'false' otherwise. */
 442 bool
 443 mac_learning_update(struct mac_learning *ml, struct eth_addr src,
 444                     int vlan, bool is_gratuitous_arp, bool is_bond,
 445                     void *in_port)
 446     OVS_EXCLUDED(ml->rwlock)
 447 {
 448     bool need_update;
 449     bool updated = false;
 450
 451     /* Don't learn the OFPP_NONE port. */
 452     if (in_port != NULL) {
 453         /* First try the common case: no change to MAC learning table. */
 454         ovs_rwlock_rdlock(&ml->rwlock);
 455         need_update = is_mac_learning_update_needed(ml, src, vlan,
 456                                                     is_gratuitous_arp, is_bond,
 457                                                     in_port);
 458         ovs_rwlock_unlock(&ml->rwlock);
 459
 460         if (need_update) {
 461             /* Slow path: MAC learning table might need an update. */
 462             ovs_rwlock_wrlock(&ml->rwlock);
 463             updated = update_learning_table__(ml, src, vlan, is_gratuitous_arp,
 464                                               is_bond, in_port);
 465             ovs_rwlock_unlock(&ml->rwlock);
 466         }
 467     }
 468     return updated;
 469 }
 470
 471 /* Looks up MAC 'dst' for VLAN 'vlan' in 'ml' and returns the associated MAC
 472  * learning entry, if any. */
 473 struct mac_entry *
 474 mac_learning_lookup(const struct mac_learning *ml,
 475                     const struct eth_addr dst, uint16_t vlan)
 476 {
 477     if (eth_addr_is_multicast(dst)) {
 478         /* No tag because the treatment of multicast destinations never
 479          * changes. */
 480         return NULL;
 481     } else if (!is_learning_vlan(ml, vlan)) {
 482         /* We don't tag this property.  The set of learning VLANs changes so
 483          * rarely that we revalidate every flow when it changes. */
 484         return NULL;
 485     } else {
 486         struct mac_entry *e = mac_entry_lookup(ml, dst, vlan);
 487
 488         ovs_assert(e == NULL || mac_entry_get_port(ml, e) != NULL);
 489         return e;
 490     }
 491 }
 492
 493 /* Expires 'e' from the 'ml' hash table. */
 494 void
 495 mac_learning_expire(struct mac_learning *ml, struct mac_entry *e)
 496 {
 497     ml->need_revalidate = true;
 498     mac_entry_set_port(ml, e, NULL);
 499     hmap_remove(&ml->table, &e->hmap_node);
 500     ovs_list_remove(&e->lru_node);
 501     free(e);
 502 }
 503
 504 /* Expires all the mac-learning entries in 'ml'. */
 505 void
 506 mac_learning_flush(struct mac_learning *ml)
 507 {
 508     struct mac_entry *e;
 509     while (get_lru(ml, &e)){
 510         mac_learning_expire(ml, e);
 511     }
 512     hmap_shrink(&ml->table);
 513 }
 514
 515 /* Does periodic work required by 'ml'.  Returns true if something changed that
 516  * may require flow revalidation. */
 517 bool
 518 mac_learning_run(struct mac_learning *ml)
 519 {
 520     bool need_revalidate;
 521     struct mac_entry *e;
 522
 523     while (get_lru(ml, &e)
 524            && (hmap_count(&ml->table) > ml->max_entries
 525                || time_now() >= e->expires)) {
 526         COVERAGE_INC(mac_learning_expired);
 527         mac_learning_expire(ml, e);
 528     }
 529
 530     need_revalidate = ml->need_revalidate;
 531     ml->need_revalidate = false;
 532     return need_revalidate;
 533 }
 534
 535 void
 536 mac_learning_wait(struct mac_learning *ml)
 537 {
 538     if (hmap_count(&ml->table) > ml->max_entries
 539         || ml->need_revalidate) {
 540         poll_immediate_wake();
 541     } else if (!ovs_list_is_empty(&ml->lrus)) {
 542         struct mac_entry *e = mac_entry_from_lru_node(ml->lrus.next);
 543         poll_timer_wait_until(e->expires * 1000LL);
 544     }
 545 }