ofproto/ofproto-dpif-rid.c

   1 /*
   2  * Copyright (c) 2014, 2015, 2016 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
  19 #include "ofpbuf.h"
  20 #include "ofproto-dpif.h"
  21 #include "ofproto-dpif-rid.h"
  22 #include "ofproto-provider.h"
  23 #include "openvswitch/vlog.h"
  24
  25 VLOG_DEFINE_THIS_MODULE(ofproto_dpif_rid);
  26
  27 static struct ovs_mutex mutex;
  28
  29 static struct cmap id_map;
  30 static struct cmap metadata_map;
  31
  32 static struct ovs_list expiring OVS_GUARDED_BY(mutex);
  33 static struct ovs_list expired OVS_GUARDED_BY(mutex);
  34
  35 static uint32_t next_id OVS_GUARDED_BY(mutex); /* Possible next free id. */
  36
  37 #define RECIRC_POOL_STATIC_IDS 1024
  38
  39 static void recirc_id_node_free(struct recirc_id_node *);
  40
  41 void
  42 recirc_init(void)
  43 {
  44     static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
  45
  46     if (ovsthread_once_start(&once)) {
  47         ovs_mutex_init(&mutex);
  48         ovs_mutex_lock(&mutex);
  49         next_id = 1; /* 0 is not a valid ID. */
  50         cmap_init(&id_map);
  51         cmap_init(&metadata_map);
  52         list_init(&expiring);
  53         list_init(&expired);
  54         ovs_mutex_unlock(&mutex);
  55
  56         ovsthread_once_done(&once);
  57     }
  58
  59 }
  60
  61 /* This should be called by the revalidator once at each round (every 500ms or
  62  * more). */
  63 void
  64 recirc_run(void)
  65 {
  66     static long long int last = 0;
  67     long long int now = time_msec();
  68
  69     /* Do maintenance at most 4 times / sec. */
  70     ovs_mutex_lock(&mutex);
  71     if (now - last > 250) {
  72         struct recirc_id_node *node;
  73
  74         last = now;
  75
  76         /* Nodes in 'expiring' and 'expired' lists have the refcount of zero,
  77          * which means that while they can still be found (by id), no new
  78          * references can be taken on them.  We have removed the entry from the
  79          * 'metadata_map', at the time when refcount reached zero, causing any
  80          * new translations to allocate a new ID.  This allows the expiring
  81          * entry to be safely deleted while any sudden new use of the similar
  82          * recirculation will safely start using a new recirculation ID.  When
  83          * the refcount gets to zero, the node is also added to the 'expiring'
  84          * list.  At any time after that the nodes in the 'expiring' list can
  85          * be moved to the 'expired' list, from which they are deleted at least
  86          * 250ms afterwards. */
  87
  88         /* Delete the expired.  These have been lingering for at least 250 ms,
  89          * which should be enough for any ongoing recirculations to be
  90          * finished. */
  91         LIST_FOR_EACH_POP (node, exp_node, &expired) {
  92             cmap_remove(&id_map, &node->id_node, node->id);
  93             ovsrcu_postpone(recirc_id_node_free, node);
  94         }
  95
  96         if (!list_is_empty(&expiring)) {
  97             /* 'expired' is now empty, move nodes in 'expiring' to it. */
  98             list_splice(&expired, list_front(&expiring), &expiring);
  99         }
 100     }
 101     ovs_mutex_unlock(&mutex);
 102 }
 103
 104 /* We use the id as the hash value, which works due to cmap internal rehashing.
 105  * We also only insert nodes with unique IDs, so all possible hash collisions
 106  * remain internal to the cmap. */
 107 static struct recirc_id_node *
 108 recirc_find__(uint32_t id)
 109     OVS_REQUIRES(mutex)
 110 {
 111     struct cmap_node *node = cmap_find_protected(&id_map, id);
 112
 113     return node ? CONTAINER_OF(node, struct recirc_id_node, id_node) : NULL;
 114 }
 115
 116 /* Lockless RCU protected lookup.  If node is needed accross RCU quiescent
 117  * state, caller should copy the contents. */
 118 const struct recirc_id_node *
 119 recirc_id_node_find(uint32_t id)
 120 {
 121     const struct cmap_node *node = cmap_find(&id_map, id);
 122
 123     return node
 124         ? CONTAINER_OF(node, const struct recirc_id_node, id_node)
 125         : NULL;
 126 }
 127
 128 static uint32_t
 129 recirc_metadata_hash(const struct recirc_state *state)
 130 {
 131     uint32_t hash;
 132
 133     hash = hash_pointer(state->ofproto, 0);
 134     hash = hash_int(state->table_id, hash);
 135     if (flow_tnl_dst_is_set(state->metadata.tunnel)) {
 136         /* We may leave remainder bytes unhashed, but that is unlikely as
 137          * the tunnel is not in the datapath format. */
 138         hash = hash_words64((const uint64_t *) state->metadata.tunnel,
 139                             flow_tnl_size(state->metadata.tunnel)
 140                             / sizeof(uint64_t), hash);
 141     }
 142     hash = hash_boolean(state->conntracked, hash);
 143     hash = hash_words64((const uint64_t *) &state->metadata.metadata,
 144                         (sizeof state->metadata - sizeof state->metadata.tunnel)
 145                         / sizeof(uint64_t),
 146                         hash);
 147     if (state->stack && state->stack->size != 0) {
 148         hash = hash_words64((const uint64_t *) state->stack->data,
 149                             state->stack->size / sizeof(uint64_t), hash);
 150     }
 151     hash = hash_int(state->mirrors, hash);
 152     hash = hash_int(state->action_set_len, hash);
 153     if (state->ofpacts_len) {
 154         hash = hash_words64(ALIGNED_CAST(const uint64_t *, state->ofpacts),
 155                             state->ofpacts_len / sizeof(uint64_t),
 156                             hash);
 157     }
 158     return hash;
 159 }
 160
 161 static bool
 162 recirc_metadata_equal(const struct recirc_state *a,
 163                       const struct recirc_state *b)
 164 {
 165     return (a->table_id == b->table_id
 166             && a->ofproto == b->ofproto
 167             && flow_tnl_equal(a->metadata.tunnel, b->metadata.tunnel)
 168             && !memcmp(&a->metadata.metadata, &b->metadata.metadata,
 169                        sizeof a->metadata - sizeof a->metadata.tunnel)
 170             && (((!a->stack || !a->stack->size) &&
 171                  (!b->stack || !b->stack->size))
 172                 || (a->stack && b->stack && ofpbuf_equal(a->stack, b->stack)))
 173             && a->mirrors == b->mirrors
 174             && a->conntracked == b->conntracked
 175             && a->action_set_len == b->action_set_len
 176             && ofpacts_equal(a->ofpacts, a->ofpacts_len,
 177                              b->ofpacts, b->ofpacts_len));
 178 }
 179
 180 /* Lockless RCU protected lookup.  If node is needed accross RCU quiescent
 181  * state, caller should take a reference. */
 182 static struct recirc_id_node *
 183 recirc_find_equal(const struct recirc_state *target, uint32_t hash)
 184 {
 185     struct recirc_id_node *node;
 186
 187     CMAP_FOR_EACH_WITH_HASH (node, metadata_node, hash, &metadata_map) {
 188         if (recirc_metadata_equal(&node->state, target)) {
 189             return node;
 190         }
 191     }
 192     return NULL;
 193 }
 194
 195 static struct recirc_id_node *
 196 recirc_ref_equal(const struct recirc_state *target, uint32_t hash)
 197 {
 198     struct recirc_id_node *node;
 199
 200     do {
 201         node = recirc_find_equal(target, hash);
 202
 203         /* Try again if the node was released before we get the reference. */
 204     } while (node && !ovs_refcount_try_ref_rcu(&node->refcount));
 205
 206     return node;
 207 }
 208
 209 static void
 210 recirc_state_clone(struct recirc_state *new, const struct recirc_state *old,
 211                    struct flow_tnl *tunnel)
 212 {
 213     *new = *old;
 214     flow_tnl_copy__(tunnel, old->metadata.tunnel);
 215     new->metadata.tunnel = tunnel;
 216
 217     if (new->stack) {
 218         new->stack = new->stack->size ? ofpbuf_clone(new->stack) : NULL;
 219     }
 220     if (new->ofpacts) {
 221         new->ofpacts = (new->ofpacts_len
 222                         ? xmemdup(new->ofpacts, new->ofpacts_len)
 223                         : NULL);
 224     }
 225 }
 226
 227 static void
 228 recirc_state_free(struct recirc_state *state)
 229 {
 230     ofpbuf_delete(state->stack);
 231     free(state->ofpacts);
 232 }
 233
 234 /* Allocate a unique recirculation id for the given set of flow metadata.
 235  * The ID space is 2^^32, so there should never be a situation in which all
 236  * the IDs are used up.  We loop until we find a free one.
 237  * hash is recomputed if it is passed in as 0. */
 238 static struct recirc_id_node *
 239 recirc_alloc_id__(const struct recirc_state *state, uint32_t hash)
 240 {
 241     ovs_assert(state->action_set_len <= state->ofpacts_len);
 242
 243     struct recirc_id_node *node = xzalloc(sizeof *node);
 244
 245     node->hash = hash;
 246     ovs_refcount_init(&node->refcount);
 247     recirc_state_clone(CONST_CAST(struct recirc_state *, &node->state), state,
 248                        &node->state_metadata_tunnel);
 249
 250     ovs_mutex_lock(&mutex);
 251     for (;;) {
 252         /* Claim the next ID.  The ID space should be sparse enough for the
 253            allocation to succeed at the first try.  We do skip the first
 254            RECIRC_POOL_STATIC_IDS IDs on the later rounds, though, as some of
 255            the initial allocations may be for long term uses (like bonds). */
 256         node->id = next_id++;
 257         if (OVS_UNLIKELY(!node->id)) {
 258             next_id = RECIRC_POOL_STATIC_IDS + 1;
 259             node->id = next_id++;
 260         }
 261         /* Find if the id is free. */
 262         if (OVS_LIKELY(!recirc_find__(node->id))) {
 263             break;
 264         }
 265     }
 266     cmap_insert(&id_map, &node->id_node, node->id);
 267     cmap_insert(&metadata_map, &node->metadata_node, node->hash);
 268     ovs_mutex_unlock(&mutex);
 269     return node;
 270 }
 271
 272 /* Look up an existing ID for the given flow's metadata and optional actions.
 273  */
 274 uint32_t
 275 recirc_find_id(const struct recirc_state *target)
 276 {
 277     uint32_t hash = recirc_metadata_hash(target);
 278     struct recirc_id_node *node = recirc_find_equal(target, hash);
 279     return node ? node->id : 0;
 280 }
 281
 282 /* Allocate a unique recirculation id for the given set of flow metadata and
 283    optional actions. */
 284 uint32_t
 285 recirc_alloc_id_ctx(const struct recirc_state *state)
 286 {
 287     uint32_t hash = recirc_metadata_hash(state);
 288     struct recirc_id_node *node = recirc_ref_equal(state, hash);
 289     if (!node) {
 290         node = recirc_alloc_id__(state, hash);
 291     }
 292     return node->id;
 293 }
 294
 295 /* Allocate a unique recirculation id. */
 296 uint32_t
 297 recirc_alloc_id(struct ofproto_dpif *ofproto)
 298 {
 299     struct flow_tnl tunnel;
 300     tunnel.ip_dst = htonl(0);
 301     tunnel.ipv6_dst = in6addr_any;
 302     struct recirc_state state = {
 303         .table_id = TBL_INTERNAL,
 304         .ofproto = ofproto,
 305         .metadata = { .tunnel = &tunnel, .in_port = OFPP_NONE },
 306     };
 307     return recirc_alloc_id__(&state, recirc_metadata_hash(&state))->id;
 308 }
 309
 310 static void
 311 recirc_id_node_free(struct recirc_id_node *node)
 312 {
 313     recirc_state_free(CONST_CAST(struct recirc_state *, &node->state));
 314     free(node);
 315 }
 316
 317 void
 318 recirc_id_node_unref(const struct recirc_id_node *node_)
 319     OVS_EXCLUDED(mutex)
 320 {
 321     struct recirc_id_node *node = CONST_CAST(struct recirc_id_node *, node_);
 322
 323     if (node && ovs_refcount_unref(&node->refcount) == 1) {
 324         ovs_mutex_lock(&mutex);
 325         /* Prevent re-use of this node by removing the node from 'metadata_map'
 326          */
 327         cmap_remove(&metadata_map, &node->metadata_node, node->hash);
 328         /* We keep the node in the 'id_map' so that it can be found as long
 329          * as it lingers, and add it to the 'expiring' list. */
 330         list_insert(&expiring, &node->exp_node);
 331         ovs_mutex_unlock(&mutex);
 332     }
 333 }
 334
 335 void
 336 recirc_free_id(uint32_t id)
 337 {
 338     const struct recirc_id_node *node;
 339
 340     node = recirc_id_node_find(id);
 341     if (node) {
 342         recirc_id_node_unref(node);
 343     } else {
 344         VLOG_ERR("Freeing nonexistent recirculation ID: %"PRIu32, id);
 345     }
 346 }
 347
 348 /* Called when 'ofproto' is destructed.  Checks for and clears any
 349  * recirc_id leak.
 350  * No other thread may have access to the 'ofproto' being destructed.
 351  * All related datapath flows must be deleted before calling this. */
 352 void
 353 recirc_free_ofproto(struct ofproto_dpif *ofproto, const char *ofproto_name)
 354 {
 355     struct recirc_id_node *n;
 356
 357     CMAP_FOR_EACH (n, metadata_node, &metadata_map) {
 358         if (n->state.ofproto == ofproto) {
 359             VLOG_ERR("recirc_id %"PRIu32
 360                      " left allocated when ofproto (%s)"
 361                      " is destructed", n->id, ofproto_name);
 362         }
 363     }
 364 }