ofproto: Fully construct rules before putting them in the classifier.
[sliver-openvswitch.git] / ofproto / ofproto.c
index bead3c4..e5ad442 100644 (file)
@@ -76,7 +76,8 @@ enum ofproto_state {
 enum ofoperation_type {
     OFOPERATION_ADD,
     OFOPERATION_DELETE,
-    OFOPERATION_MODIFY
+    OFOPERATION_MODIFY,
+    OFOPERATION_REPLACE
 };
 
 /* A single OpenFlow request can execute any number of operations.  The
@@ -121,10 +122,8 @@ struct ofoperation {
     struct rule *rule;          /* Rule being operated upon. */
     enum ofoperation_type type; /* Type of operation. */
 
-    /* OFOPERATION_ADD. */
-    struct rule *victim;        /* Rule being replaced, if any.. */
-
-    /* OFOPERATION_MODIFY: The old actions, if the actions are changing. */
+    /* OFOPERATION_MODIFY, OFOPERATION_REPLACE: The old actions, if the actions
+     * are changing. */
     struct ofpact *ofpacts;
     size_t ofpacts_len;
     uint32_t meter_id;
@@ -133,6 +132,9 @@ struct ofoperation {
     enum ofp_flow_removed_reason reason; /* Reason flow was removed. */
 
     ovs_be64 flow_cookie;       /* Rule's old flow cookie. */
+    uint16_t idle_timeout;      /* Rule's old idle timeout. */
+    uint16_t hard_timeout;      /* Rule's old hard timeout. */
+    bool send_flow_removed;     /* Rule's old 'send_flow_removed'. */
     enum ofperr error;          /* 0 if no error. */
 };
 
@@ -157,8 +159,7 @@ static void oftable_remove_rule(struct rule *rule) OVS_RELEASES(rule->evict);
 static void oftable_remove_rule__(struct ofproto *ofproto,
                                   struct classifier *cls, struct rule *rule)
     OVS_REQ_WRLOCK(cls->rwlock) OVS_RELEASES(rule->evict);
-static struct rule *oftable_replace_rule(struct rule *);
-static void oftable_substitute_rule(struct rule *old, struct rule *new);
+static void oftable_insert_rule(struct rule *);
 
 /* A set of rules within a single OpenFlow table (oftable) that have the same
  * values for the oftable's eviction_fields.  A rule to be evicted, when one is
@@ -186,6 +187,8 @@ static bool choose_rule_to_evict(struct oftable *table, struct rule **rulep)
     OVS_TRY_WRLOCK(true, (*rulep)->evict);
 static void ofproto_evict(struct ofproto *);
 static uint32_t rule_eviction_priority(struct rule *);
+static void eviction_group_add_rule(struct rule *);
+static void eviction_group_remove_rule(struct rule *);
 
 /* ofport. */
 static void ofport_destroy__(struct ofport *);
@@ -204,6 +207,9 @@ static bool rule_is_modifiable(const struct rule *);
 static enum ofperr add_flow(struct ofproto *, struct ofconn *,
                             struct ofputil_flow_mod *,
                             const struct ofp_header *);
+static enum ofperr modify_flows__(struct ofproto *, struct ofconn *,
+                                  struct ofputil_flow_mod *,
+                                  const struct ofp_header *, struct list *);
 static void delete_flow__(struct rule *rule, struct ofopgroup *,
                           enum ofp_flow_removed_reason)
     OVS_RELEASES(rule->evict);
@@ -432,12 +438,12 @@ ofproto_create(const char *datapath_name, const char *datapath_type,
     hmap_init(&ofproto->ports);
     shash_init(&ofproto->port_by_name);
     simap_init(&ofproto->ofp_requests);
-    ofproto->max_ports = OFPP_MAX;
+    ofproto->max_ports = ofp_to_u16(OFPP_MAX);
     ofproto->tables = NULL;
     ofproto->n_tables = 0;
     hindex_init(&ofproto->cookies);
     list_init(&ofproto->expirable);
-    ovs_mutex_init(&ofproto->expirable_mutex, PTHREAD_MUTEX_RECURSIVE);
+    ovs_mutex_init_recursive(&ofproto->expirable_mutex);
     ofproto->connmgr = connmgr_create(ofproto, datapath_name, datapath_name);
     ofproto->state = S_OPENFLOW;
     list_init(&ofproto->pending);
@@ -461,7 +467,7 @@ ofproto_create(const char *datapath_name, const char *datapath_type,
 
     /* The "max_ports" member should have been set by ->construct(ofproto).
      * Port 0 is not a valid OpenFlow port, so mark that as unavailable. */
-    ofproto->ofp_port_ids = bitmap_allocate(ofp_to_u16(ofproto->max_ports));
+    ofproto->ofp_port_ids = bitmap_allocate(ofproto->max_ports);
     bitmap_set1(ofproto->ofp_port_ids, 0);
 
     /* Check that hidden tables, if any, are at the end. */
@@ -520,9 +526,9 @@ ofproto_init_tables(struct ofproto *ofproto, int n_tables)
  * Reserved ports numbered OFPP_MAX and higher are special and not subject to
  * the 'max_ports' restriction. */
 void
-ofproto_init_max_ports(struct ofproto *ofproto, ofp_port_t max_ports)
+ofproto_init_max_ports(struct ofproto *ofproto, uint16_t max_ports)
 {
-    ovs_assert(ofp_to_u16(max_ports) <= ofp_to_u16(OFPP_MAX));
+    ovs_assert(max_ports <= ofp_to_u16(OFPP_MAX));
     ofproto->max_ports = max_ports;
 }
 
@@ -1068,17 +1074,37 @@ ofproto_get_snoops(const struct ofproto *ofproto, struct sset *snoops)
     connmgr_get_snoops(ofproto->connmgr, snoops);
 }
 
+/* Deletes 'rule' from 'cls' within 'ofproto'.
+ *
+ * The 'cls' argument is redundant (it is &ofproto->tables[rule->table_id].cls)
+ * but it allows Clang to do better checking. */
 static void
-ofproto_flush__(struct ofproto *ofproto)
+ofproto_delete_rule(struct ofproto *ofproto, struct classifier *cls,
+                    struct rule *rule)
+    OVS_REQ_WRLOCK(cls->rwlock)
 {
     struct ofopgroup *group;
+
+    ovs_assert(!rule->pending);
+    ovs_assert(cls == &ofproto->tables[rule->table_id].cls);
+
+    group = ofopgroup_create_unattached(ofproto);
+    ofoperation_create(group, rule, OFOPERATION_DELETE, OFPRR_DELETE);
+    ovs_rwlock_wrlock(&rule->evict);
+    oftable_remove_rule__(ofproto, cls, rule);
+    ofproto->ofproto_class->rule_delete(rule);
+    ofopgroup_submit(group);
+}
+
+static void
+ofproto_flush__(struct ofproto *ofproto)
+{
     struct oftable *table;
 
     if (ofproto->ofproto_class->flush) {
         ofproto->ofproto_class->flush(ofproto);
     }
 
-    group = ofopgroup_create_unattached(ofproto);
     OFPROTO_FOR_EACH_TABLE (table, ofproto) {
         struct rule *rule, *next_rule;
         struct cls_cursor cursor;
@@ -1091,16 +1117,11 @@ ofproto_flush__(struct ofproto *ofproto)
         cls_cursor_init(&cursor, &table->cls, NULL);
         CLS_CURSOR_FOR_EACH_SAFE (rule, next_rule, cr, &cursor) {
             if (!rule->pending) {
-                ofoperation_create(group, rule, OFOPERATION_DELETE,
-                                   OFPRR_DELETE);
-                ovs_rwlock_wrlock(&rule->evict);
-                oftable_remove_rule__(ofproto, &table->cls, rule);
-                ofproto->ofproto_class->rule_destruct(rule);
+                ofproto_delete_rule(ofproto, &table->cls, rule);
             }
         }
         ovs_rwlock_unlock(&table->cls.rwlock);
     }
-    ofopgroup_submit(group);
 }
 
 static void
@@ -1678,12 +1699,13 @@ bool
 ofproto_delete_flow(struct ofproto *ofproto,
                     const struct match *target, unsigned int priority)
 {
+    struct classifier *cls = &ofproto->tables[0].cls;
     struct rule *rule;
 
-    ovs_rwlock_rdlock(&ofproto->tables[0].cls.rwlock);
-    rule = rule_from_cls_rule(classifier_find_match_exactly(
-                                  &ofproto->tables[0].cls, target, priority));
-    ovs_rwlock_unlock(&ofproto->tables[0].cls.rwlock);
+    ovs_rwlock_rdlock(&cls->rwlock);
+    rule = rule_from_cls_rule(classifier_find_match_exactly(cls, target,
+                                                            priority));
+    ovs_rwlock_unlock(&cls->rwlock);
     if (!rule) {
         /* No such rule -> success. */
         return true;
@@ -1693,12 +1715,10 @@ ofproto_delete_flow(struct ofproto *ofproto,
         return false;
     } else {
         /* Initiate deletion -> success. */
-        struct ofopgroup *group = ofopgroup_create_unattached(ofproto);
-        ofoperation_create(group, rule, OFOPERATION_DELETE, OFPRR_DELETE);
-        ovs_rwlock_wrlock(&rule->evict);
-        oftable_remove_rule(rule);
-        ofproto->ofproto_class->rule_destruct(rule);
-        ofopgroup_submit(group);
+        ovs_rwlock_wrlock(&cls->rwlock);
+        ofproto_delete_rule(ofproto, cls, rule);
+        ovs_rwlock_unlock(&cls->rwlock);
+
         return true;
     }
 
@@ -1742,32 +1762,28 @@ reinit_ports(struct ofproto *p)
 static ofp_port_t
 alloc_ofp_port(struct ofproto *ofproto, const char *netdev_name)
 {
-    uint16_t max_ports = ofp_to_u16(ofproto->max_ports);
     uint16_t port_idx;
 
     port_idx = simap_get(&ofproto->ofp_requests, netdev_name);
-    if (!port_idx) {
-        port_idx = UINT16_MAX;
-    }
+    port_idx = port_idx ? port_idx : UINT16_MAX;
 
-    if (port_idx >= max_ports
+    if (port_idx >= ofproto->max_ports
         || bitmap_is_set(ofproto->ofp_port_ids, port_idx)) {
-        uint16_t end_port_no = ofp_to_u16(ofproto->alloc_port_no);
-        uint16_t alloc_port_no = end_port_no;
+        uint16_t end_port_no = ofproto->alloc_port_no;
 
         /* Search for a free OpenFlow port number.  We try not to
          * immediately reuse them to prevent problems due to old
          * flows. */
         for (;;) {
-            if (++alloc_port_no >= max_ports) {
-                alloc_port_no = 0;
+            if (++ofproto->alloc_port_no >= ofproto->max_ports) {
+                ofproto->alloc_port_no = 0;
             }
-            if (!bitmap_is_set(ofproto->ofp_port_ids, alloc_port_no)) {
-                port_idx = alloc_port_no;
-                ofproto->alloc_port_no = u16_to_ofp(alloc_port_no);
+            if (!bitmap_is_set(ofproto->ofp_port_ids,
+                               ofproto->alloc_port_no)) {
+                port_idx = ofproto->alloc_port_no;
                 break;
             }
-            if (alloc_port_no == end_port_no) {
+            if (ofproto->alloc_port_no == end_port_no) {
                 return OFPP_NONE;
             }
         }
@@ -1779,7 +1795,7 @@ alloc_ofp_port(struct ofproto *ofproto, const char *netdev_name)
 static void
 dealloc_ofp_port(const struct ofproto *ofproto, ofp_port_t ofp_port)
 {
-    if (ofp_to_u16(ofp_port) < ofp_to_u16(ofproto->max_ports)) {
+    if (ofp_to_u16(ofp_port) < ofproto->max_ports) {
         bitmap_set0(ofproto->ofp_port_ids, ofp_to_u16(ofp_port));
     }
 }
@@ -2103,6 +2119,10 @@ init_ports(struct ofproto *p)
             netdev = ofport_open(p, &ofproto_port, &pp);
             if (netdev) {
                 ofport_install(p, netdev, &pp);
+                if (ofp_to_u16(ofproto_port.ofp_port) < p->max_ports) {
+                    p->alloc_port_no = MAX(p->alloc_port_no,
+                                           ofp_to_u16(ofproto_port.ofp_port));
+                }
             }
         }
     }
@@ -2196,6 +2216,7 @@ static void
 ofproto_rule_destroy__(struct rule *rule)
 {
     if (rule) {
+        rule->ofproto->ofproto_class->rule_destruct(rule);
         cls_rule_destroy(&rule->cr);
         free(rule->ofpacts);
         ovs_mutex_destroy(&rule->timeout_mutex);
@@ -2205,24 +2226,18 @@ ofproto_rule_destroy__(struct rule *rule)
 }
 
 /* This function allows an ofproto implementation to destroy any rules that
- * remain when its ->destruct() function is called.  The caller must have
- * already uninitialized any derived members of 'rule' (step 5 described in the
- * large comment in ofproto/ofproto-provider.h titled "Life Cycle").
- * This function implements steps 6 and 7.
+ * remain when its ->destruct() function is called..  This function implements
+ * steps 4.4 and 4.5 in the section titled "Rule Life Cycle" in
+ * ofproto-provider.h.
  *
  * This function should only be called from an ofproto implementation's
  * ->destruct() function.  It is not suitable elsewhere. */
 void
-ofproto_rule_destroy(struct ofproto *ofproto, struct classifier *cls,
-                     struct rule *rule) OVS_REQ_WRLOCK(cls->rwlock)
+ofproto_rule_delete(struct ofproto *ofproto, struct classifier *cls,
+                    struct rule *rule)
+    OVS_REQ_WRLOCK(cls->rwlock)
 {
-    ovs_assert(!rule->pending);
-    if (!ovs_rwlock_trywrlock(&rule->evict)) {
-        oftable_remove_rule__(ofproto, cls, rule);
-    } else {
-        NOT_REACHED();
-    }
-    ofproto_rule_destroy__(rule);
+    ofproto_delete_rule(ofproto, cls, rule);
 }
 
 /* Returns true if 'rule' has an OpenFlow OFPAT_OUTPUT or OFPAT_ENQUEUE action
@@ -2245,12 +2260,11 @@ ofoperation_has_out_port(const struct ofoperation *op, ofp_port_t out_port)
 
     switch (op->type) {
     case OFOPERATION_ADD:
-        return op->victim && ofproto_rule_has_out_port(op->victim, out_port);
-
     case OFOPERATION_DELETE:
         return false;
 
     case OFOPERATION_MODIFY:
+    case OFOPERATION_REPLACE:
         return ofpacts_output_to_port(op->ofpacts, op->ofpacts_len, out_port);
     }
 
@@ -2462,8 +2476,8 @@ ofproto_check_ofpacts(struct ofproto *ofproto,
     enum ofperr error;
     uint32_t mid;
 
-    error = ofpacts_check(ofpacts, ofpacts_len, flow, ofproto->max_ports,
-                          table_id);
+    error = ofpacts_check(ofpacts, ofpacts_len, flow,
+                          u16_to_ofp(ofproto->max_ports), table_id);
     if (error) {
         return error;
     }
@@ -2500,7 +2514,7 @@ handle_packet_out(struct ofconn *ofconn, const struct ofp_header *oh)
     if (error) {
         goto exit_free_ofpacts;
     }
-    if (ofp_to_u16(po.in_port) >= ofp_to_u16(p->max_ports)
+    if (ofp_to_u16(po.in_port) >= p->max_ports
         && ofp_to_u16(po.in_port) < ofp_to_u16(OFPP_MAX)) {
         error = OFPERR_OFPBRC_BAD_PORT;
         goto exit_free_ofpacts;
@@ -3072,17 +3086,17 @@ handle_flow_stats_request(struct ofconn *ofconn,
                                                &fs.byte_count);
         fs.ofpacts = rule->ofpacts;
         fs.ofpacts_len = rule->ofpacts_len;
-        fs.flags = 0;
 
         ovs_mutex_lock(&rule->timeout_mutex);
         fs.idle_timeout = rule->idle_timeout;
         fs.hard_timeout = rule->hard_timeout;
         ovs_mutex_unlock(&rule->timeout_mutex);
 
+        fs.flags = 0;
         if (rule->send_flow_removed) {
-            fs.flags |= OFPFF_SEND_FLOW_REM;
-            /* FIXME: Implement OF 1.3 flags OFPFF13_NO_PKT_COUNTS
-               and OFPFF13_NO_BYT_COUNTS */
+            fs.flags |= OFPUTIL_FF_SEND_FLOW_REM;
+            /* FIXME: Implement OFPUTIL_FF_NO_PKT_COUNTS and
+               OFPUTIL_FF_NO_BYT_COUNTS. */
         }
         ofputil_append_flow_stats_reply(&fs, &replies);
     }
@@ -3337,6 +3351,34 @@ is_flow_deletion_pending(const struct ofproto *ofproto,
     return false;
 }
 
+static enum ofperr
+evict_rule_from_table(struct ofproto *ofproto, struct oftable *table)
+{
+    struct rule *rule;
+    size_t n_rules;
+
+    ovs_rwlock_rdlock(&table->cls.rwlock);
+    n_rules = classifier_count(&table->cls);
+    ovs_rwlock_unlock(&table->cls.rwlock);
+
+    if (n_rules < table->max_flows) {
+        return 0;
+    } else if (!choose_rule_to_evict(table, &rule)) {
+        return OFPERR_OFPFMFC_TABLE_FULL;
+    } else if (rule->pending) {
+        ovs_rwlock_unlock(&rule->evict);
+        return OFPROTO_POSTPONE;
+    } else {
+        struct ofopgroup *group;
+
+        group = ofopgroup_create_unattached(ofproto);
+        delete_flow__(rule, group, OFPRR_EVICTION);
+        ofopgroup_submit(group);
+
+        return 0;
+    }
+}
+
 /* Implements OFPFC_ADD and the cases for OFPFC_MODIFY and OFPFC_MODIFY_STRICT
  * in which no matching flow already exists in the flow table.
  *
@@ -3356,10 +3398,9 @@ add_flow(struct ofproto *ofproto, struct ofconn *ofconn,
 {
     struct oftable *table;
     struct ofopgroup *group;
-    struct rule *victim;
+    struct cls_rule cr;
     struct rule *rule;
     uint8_t table_id;
-    bool overlaps;
     int error;
 
     error = check_table_id(ofproto, fm->table_id);
@@ -3392,6 +3433,28 @@ add_flow(struct ofproto *ofproto, struct ofconn *ofconn,
         return OFPERR_OFPBRC_EPERM;
     }
 
+    cls_rule_init(&cr, &fm->match, fm->priority);
+
+    /* Transform "add" into "modify" if there's an existing identical flow. */
+    ovs_rwlock_rdlock(&table->cls.rwlock);
+    rule = rule_from_cls_rule(classifier_find_rule_exactly(&table->cls, &cr));
+    ovs_rwlock_unlock(&table->cls.rwlock);
+    if (rule) {
+        cls_rule_destroy(&cr);
+        if (!rule_is_modifiable(rule)) {
+            return OFPERR_OFPBRC_EPERM;
+        } else if (rule->pending) {
+            return OFPROTO_POSTPONE;
+        } else {
+            struct list rules;
+
+            list_init(&rules);
+            list_push_back(&rules, &rule->ofproto_node);
+            fm->modify_cookie = true;
+            return modify_flows__(ofproto, ofconn, fm, request, &rules);
+        }
+    }
+
     /* Verify actions. */
     error = ofproto_check_ofpacts(ofproto, fm->ofpacts, fm->ofpacts_len,
                                   &fm->match.flow, table_id);
@@ -3399,49 +3462,57 @@ add_flow(struct ofproto *ofproto, struct ofconn *ofconn,
         return error;
     }
 
-    /* Allocate new rule and initialize classifier rule. */
-    rule = ofproto->ofproto_class->rule_alloc();
-    if (!rule) {
-        VLOG_WARN_RL(&rl, "%s: failed to create rule (%s)",
-                     ofproto->name, ovs_strerror(error));
-        return ENOMEM;
-    }
-    cls_rule_init(&rule->cr, &fm->match, fm->priority);
-
     /* Serialize against pending deletion. */
-    if (is_flow_deletion_pending(ofproto, &rule->cr, table_id)) {
-        cls_rule_destroy(&rule->cr);
-        ofproto->ofproto_class->rule_dealloc(rule);
+    if (is_flow_deletion_pending(ofproto, &cr, table_id)) {
+        cls_rule_destroy(&cr);
         return OFPROTO_POSTPONE;
     }
 
     /* Check for overlap, if requested. */
-    ovs_rwlock_rdlock(&table->cls.rwlock);
-    overlaps = classifier_rule_overlaps(&table->cls, &rule->cr);
-    ovs_rwlock_unlock(&table->cls.rwlock);
-    if (fm->flags & OFPFF_CHECK_OVERLAP && overlaps) {
-        cls_rule_destroy(&rule->cr);
-        ofproto->ofproto_class->rule_dealloc(rule);
-        return OFPERR_OFPFMFC_OVERLAP;
+    if (fm->flags & OFPUTIL_FF_CHECK_OVERLAP) {
+        bool overlaps;
+
+        ovs_rwlock_rdlock(&table->cls.rwlock);
+        overlaps = classifier_rule_overlaps(&table->cls, &cr);
+        ovs_rwlock_unlock(&table->cls.rwlock);
+
+        if (overlaps) {
+            cls_rule_destroy(&cr);
+            return OFPERR_OFPFMFC_OVERLAP;
+        }
+    }
+
+    /* If necessary, evict an existing rule to clear out space. */
+    error = evict_rule_from_table(ofproto, table);
+    if (error) {
+        cls_rule_destroy(&cr);
+        return error;
     }
 
-    /* FIXME: Implement OFPFF12_RESET_COUNTS */
+    /* Allocate new rule. */
+    rule = ofproto->ofproto_class->rule_alloc();
+    if (!rule) {
+        cls_rule_destroy(&cr);
+        VLOG_WARN_RL(&rl, "%s: failed to create rule (%s)",
+                     ofproto->name, ovs_strerror(error));
+        return ENOMEM;
+    }
 
+    /* Initialize base state. */
     rule->ofproto = ofproto;
+    cls_rule_move(&rule->cr, &cr);
     rule->pending = NULL;
     rule->flow_cookie = fm->new_cookie;
     rule->created = rule->modified = rule->used = time_msec();
 
-    ovs_mutex_init(&rule->timeout_mutex, OVS_MUTEX_ADAPTIVE);
+    ovs_mutex_init(&rule->timeout_mutex);
     ovs_mutex_lock(&rule->timeout_mutex);
     rule->idle_timeout = fm->idle_timeout;
     rule->hard_timeout = fm->hard_timeout;
     ovs_mutex_unlock(&rule->timeout_mutex);
 
     rule->table_id = table - ofproto->tables;
-    rule->send_flow_removed = (fm->flags & OFPFF_SEND_FLOW_REM) != 0;
-    /* FIXME: Implement OF 1.3 flags OFPFF13_NO_PKT_COUNTS
-       and OFPFF13_NO_BYT_COUNTS */
+    rule->send_flow_removed = (fm->flags & OFPUTIL_FF_SEND_FLOW_REM) != 0;
     rule->ofpacts = xmemdup(fm->ofpacts, fm->ofpacts_len);
     rule->ofpacts_len = fm->ofpacts_len;
     rule->meter_id = find_meter(rule->ofpacts, rule->ofpacts_len);
@@ -3453,65 +3524,21 @@ add_flow(struct ofproto *ofproto, struct ofconn *ofconn,
     rule->modify_seqno = 0;
     ovs_rwlock_init(&rule->evict);
 
-    /* Insert new rule. */
-    victim = oftable_replace_rule(rule);
-    if (victim && !rule_is_modifiable(victim)) {
-        error = OFPERR_OFPBRC_EPERM;
-    } else if (victim && victim->pending) {
-        error = OFPROTO_POSTPONE;
-    } else {
-        struct ofoperation *op;
-        struct rule *evict;
-        size_t n_rules;
-
-        ovs_rwlock_rdlock(&table->cls.rwlock);
-        n_rules = classifier_count(&table->cls);
-        ovs_rwlock_unlock(&table->cls.rwlock);
-        if (n_rules > table->max_flows) {
-            ovs_rwlock_rdlock(&rule->evict);
-            if (choose_rule_to_evict(table, &evict)) {
-                ovs_rwlock_unlock(&rule->evict);
-                ovs_rwlock_unlock(&evict->evict);
-                if (evict->pending) {
-                    error = OFPROTO_POSTPONE;
-                    goto exit;
-                }
-            } else {
-                ovs_rwlock_unlock(&rule->evict);
-                error = OFPERR_OFPFMFC_TABLE_FULL;
-                goto exit;
-            }
-        } else {
-            evict = NULL;
-        }
-
-        group = ofopgroup_create(ofproto, ofconn, request, fm->buffer_id);
-        op = ofoperation_create(group, rule, OFOPERATION_ADD, 0);
-        op->victim = victim;
-
-        error = ofproto->ofproto_class->rule_construct(rule);
-        if (error) {
-            op->group->n_running--;
-            ofoperation_destroy(rule->pending);
-        } else if (evict) {
-            /* It would be better if we maintained the lock we took in
-             * choose_rule_to_evict() earlier, but that confuses the thread
-             * safety analysis, and this code is fragile enough that we really
-             * need it.  In the worst case, we'll have to block a little while
-             * before we perform the eviction, which doesn't seem like a big
-             * problem. */
-            ovs_rwlock_wrlock(&evict->evict);
-            delete_flow__(evict, group, OFPRR_EVICTION);
-        }
-        ofopgroup_submit(group);
-    }
-
-exit:
-    /* Back out if an error occurred. */
+    /* Construct rule, initializing derived state. */
+    error = ofproto->ofproto_class->rule_construct(rule);
     if (error) {
-        oftable_substitute_rule(rule, victim);
         ofproto_rule_destroy__(rule);
+        return error;
     }
+
+    /* Insert rule. */
+    oftable_insert_rule(rule);
+
+    group = ofopgroup_create(ofproto, ofconn, request, fm->buffer_id);
+    ofoperation_create(group, rule, OFOPERATION_ADD, 0);
+    ofproto->ofproto_class->rule_insert(rule);
+    ofopgroup_submit(group);
+
     return error;
 }
 \f
@@ -3529,17 +3556,20 @@ modify_flows__(struct ofproto *ofproto, struct ofconn *ofconn,
                struct ofputil_flow_mod *fm, const struct ofp_header *request,
                struct list *rules)
 {
+    enum ofoperation_type type;
     struct ofopgroup *group;
     struct rule *rule;
     enum ofperr error;
 
+    type = fm->command == OFPFC_ADD ? OFOPERATION_REPLACE : OFOPERATION_MODIFY;
     group = ofopgroup_create(ofproto, ofconn, request, fm->buffer_id);
     error = OFPERR_OFPBRC_EPERM;
     LIST_FOR_EACH (rule, ofproto_node, rules) {
         struct ofoperation *op;
         bool actions_changed;
+        bool reset_counters;
 
-        /* FIXME: Implement OFPFF12_RESET_COUNTS */
+        /* FIXME: Implement OFPFUTIL_FF_RESET_COUNTS */
 
         if (rule_is_modifiable(rule)) {
             /* At least one rule is modifiable, don't report EPERM error. */
@@ -3550,7 +3580,7 @@ modify_flows__(struct ofproto *ofproto, struct ofconn *ofconn,
 
         /* Verify actions. */
         error = ofpacts_check(fm->ofpacts, fm->ofpacts_len, &fm->match.flow,
-                              ofproto->max_ports, rule->table_id);
+                              u16_to_ofp(ofproto->max_ports), rule->table_id);
         if (error) {
             return error;
         }
@@ -3558,19 +3588,39 @@ modify_flows__(struct ofproto *ofproto, struct ofconn *ofconn,
         actions_changed = !ofpacts_equal(fm->ofpacts, fm->ofpacts_len,
                                          rule->ofpacts, rule->ofpacts_len);
 
-        op = ofoperation_create(group, rule, OFOPERATION_MODIFY, 0);
+        op = ofoperation_create(group, rule, type, 0);
 
         if (fm->modify_cookie && fm->new_cookie != htonll(UINT64_MAX)) {
             ofproto_rule_change_cookie(ofproto, rule, fm->new_cookie);
         }
-        if (actions_changed) {
+        if (type == OFOPERATION_REPLACE) {
+            ovs_mutex_lock(&rule->timeout_mutex);
+            rule->idle_timeout = fm->idle_timeout;
+            rule->hard_timeout = fm->hard_timeout;
+            ovs_mutex_unlock(&rule->timeout_mutex);
+
+            rule->send_flow_removed = (fm->flags
+                                       & OFPUTIL_FF_SEND_FLOW_REM) != 0;
+
+            if (fm->idle_timeout || fm->hard_timeout) {
+                if (!rule->eviction_group) {
+                    eviction_group_add_rule(rule);
+                }
+            } else {
+                eviction_group_remove_rule(rule);
+            }
+        }
+
+        reset_counters = (fm->flags & OFPUTIL_FF_RESET_COUNTS) != 0;
+        if (actions_changed || reset_counters) {
             op->ofpacts = rule->ofpacts;
             op->ofpacts_len = rule->ofpacts_len;
             op->meter_id = rule->meter_id;
             rule->ofpacts = xmemdup(fm->ofpacts, fm->ofpacts_len);
             rule->ofpacts_len = fm->ofpacts_len;
             rule->meter_id = find_meter(rule->ofpacts, rule->ofpacts_len);
-            rule->ofproto->ofproto_class->rule_modify_actions(rule);
+            rule->ofproto->ofproto_class->rule_modify_actions(rule,
+                                                              reset_counters);
         } else {
             ofoperation_complete(op, 0);
         }
@@ -3655,7 +3705,7 @@ delete_flow__(struct rule *rule, struct ofopgroup *group,
 
     ofoperation_create(group, rule, OFOPERATION_DELETE, reason);
     oftable_remove_rule(rule);
-    ofproto->ofproto_class->rule_destruct(rule);
+    ofproto->ofproto_class->rule_delete(rule);
 }
 
 /* Deletes the rules listed in 'rules'.
@@ -3769,17 +3819,54 @@ void
 ofproto_rule_expire(struct rule *rule, uint8_t reason)
 {
     struct ofproto *ofproto = rule->ofproto;
-    struct ofopgroup *group;
+    struct classifier *cls = &ofproto->tables[rule->table_id].cls;
 
     ovs_assert(reason == OFPRR_HARD_TIMEOUT || reason == OFPRR_IDLE_TIMEOUT);
-
     ofproto_rule_send_removed(rule, reason);
 
-    group = ofopgroup_create_unattached(ofproto);
-    ofoperation_create(group, rule, OFOPERATION_DELETE, reason);
-    oftable_remove_rule(rule);
-    ofproto->ofproto_class->rule_destruct(rule);
-    ofopgroup_submit(group);
+    ovs_rwlock_wrlock(&cls->rwlock);
+    ofproto_delete_rule(ofproto, cls, rule);
+    ovs_rwlock_unlock(&cls->rwlock);
+}
+
+/* Reduces '*timeout' to no more than 'max'.  A value of zero in either case
+ * means "infinite". */
+static void
+reduce_timeout(uint16_t max, uint16_t *timeout)
+{
+    if (max && (!*timeout || *timeout > max)) {
+        *timeout = max;
+    }
+}
+
+/* If 'idle_timeout' is nonzero, and 'rule' has no idle timeout or an idle
+ * timeout greater than 'idle_timeout', lowers 'rule''s idle timeout to
+ * 'idle_timeout' seconds.  Similarly for 'hard_timeout'.
+ *
+ * Suitable for implementing OFPACT_FIN_TIMEOUT. */
+void
+ofproto_rule_reduce_timeouts(struct rule *rule,
+                             uint16_t idle_timeout, uint16_t hard_timeout)
+    OVS_EXCLUDED(rule->ofproto->expirable_mutex, rule->timeout_mutex)
+{
+    if (!idle_timeout && !hard_timeout) {
+        return;
+    }
+
+    ovs_mutex_lock(&rule->ofproto->expirable_mutex);
+    if (list_is_empty(&rule->expirable)) {
+        list_insert(&rule->ofproto->expirable, &rule->expirable);
+    }
+    ovs_mutex_unlock(&rule->ofproto->expirable_mutex);
+
+    ovs_mutex_lock(&rule->timeout_mutex);
+    reduce_timeout(idle_timeout, &rule->idle_timeout);
+    reduce_timeout(hard_timeout, &rule->hard_timeout);
+    ovs_mutex_unlock(&rule->timeout_mutex);
+
+    if (!rule->eviction_group) {
+        eviction_group_add_rule(rule);
+    }
 }
 \f
 static enum ofperr
@@ -4032,7 +4119,7 @@ ofproto_compose_flow_refresh_update(const struct rule *rule,
     struct ofputil_flow_update fu;
     struct match match;
 
-    if (op && op->type == OFOPERATION_ADD && !op->victim) {
+    if (op && op->type == OFOPERATION_ADD) {
         /* We'll report the final flow when the operation completes.  Reporting
          * it now would cause a duplicate report later. */
         return;
@@ -4061,12 +4148,10 @@ ofproto_compose_flow_refresh_update(const struct rule *rule,
          * actions, so that when the operation commits we report the change. */
         switch (op->type) {
         case OFOPERATION_ADD:
-            /* We already verified that there was a victim. */
-            fu.ofpacts = op->victim->ofpacts;
-            fu.ofpacts_len = op->victim->ofpacts_len;
-            break;
+            NOT_REACHED();
 
         case OFOPERATION_MODIFY:
+        case OFOPERATION_REPLACE:
             if (op->ofpacts) {
                 fu.ofpacts = op->ofpacts;
                 fu.ofpacts_len = op->ofpacts_len;
@@ -4877,15 +4962,27 @@ ofopgroup_complete(struct ofopgroup *group)
                   && rule->flow_cookie == op->flow_cookie))) {
             /* Check that we can just cast from ofoperation_type to
              * nx_flow_update_event. */
-            BUILD_ASSERT_DECL((enum nx_flow_update_event) OFOPERATION_ADD
-                              == NXFME_ADDED);
-            BUILD_ASSERT_DECL((enum nx_flow_update_event) OFOPERATION_DELETE
-                              == NXFME_DELETED);
-            BUILD_ASSERT_DECL((enum nx_flow_update_event) OFOPERATION_MODIFY
-                              == NXFME_MODIFIED);
-
-            ofmonitor_report(ofproto->connmgr, rule,
-                             (enum nx_flow_update_event) op->type,
+            enum nx_flow_update_event event_type;
+
+            switch (op->type) {
+            case OFOPERATION_ADD:
+            case OFOPERATION_REPLACE:
+                event_type = NXFME_ADDED;
+                break;
+
+            case OFOPERATION_DELETE:
+                event_type = NXFME_DELETED;
+                break;
+
+            case OFOPERATION_MODIFY:
+                event_type = NXFME_MODIFIED;
+                break;
+
+            default:
+                NOT_REACHED();
+            }
+
+            ofmonitor_report(ofproto->connmgr, rule, event_type,
                              op->reason, abbrev_ofconn, abbrev_xid);
         }
 
@@ -4896,7 +4993,6 @@ ofopgroup_complete(struct ofopgroup *group)
             if (!op->error) {
                 uint16_t vid_mask;
 
-                ofproto_rule_destroy__(op->victim);
                 vid_mask = minimask_get_vid_mask(&rule->cr.match.mask);
                 if (vid_mask == VLAN_VID_MASK) {
                     if (ofproto->vlan_bitmap) {
@@ -4910,7 +5006,8 @@ ofopgroup_complete(struct ofopgroup *group)
                     }
                 }
             } else {
-                oftable_substitute_rule(rule, op->victim);
+                ovs_rwlock_wrlock(&rule->evict);
+                oftable_remove_rule(rule);
                 ofproto_rule_destroy__(rule);
             }
             break;
@@ -4922,10 +5019,20 @@ ofopgroup_complete(struct ofopgroup *group)
             break;
 
         case OFOPERATION_MODIFY:
+        case OFOPERATION_REPLACE:
             if (!op->error) {
-                rule->modified = time_msec();
+                long long int now = time_msec();
+
+                rule->modified = now;
+                if (op->type == OFOPERATION_REPLACE) {
+                    rule->created = rule->used = now;
+                }
             } else {
                 ofproto_rule_change_cookie(ofproto, rule, op->flow_cookie);
+                ovs_mutex_lock(&rule->timeout_mutex);
+                rule->idle_timeout = op->idle_timeout;
+                rule->hard_timeout = op->hard_timeout;
+                ovs_mutex_unlock(&rule->timeout_mutex);
                 if (op->ofpacts) {
                     free(rule->ofpacts);
                     rule->ofpacts = op->ofpacts;
@@ -4933,6 +5040,7 @@ ofopgroup_complete(struct ofopgroup *group)
                     op->ofpacts = NULL;
                     op->ofpacts_len = 0;
                 }
+                rule->send_flow_removed = op->send_flow_removed;
             }
             break;
 
@@ -4986,6 +5094,11 @@ ofoperation_create(struct ofopgroup *group, struct rule *rule,
     op->type = type;
     op->reason = reason;
     op->flow_cookie = rule->flow_cookie;
+    ovs_mutex_lock(&rule->timeout_mutex);
+    op->idle_timeout = rule->idle_timeout;
+    op->hard_timeout = rule->hard_timeout;
+    ovs_mutex_unlock(&rule->timeout_mutex);
+    op->send_flow_removed = rule->send_flow_removed;
 
     group->n_running++;
 
@@ -5017,14 +5130,7 @@ ofoperation_destroy(struct ofoperation *op)
  * indicate success or an OpenFlow error code on failure.
  *
  * If 'error' is 0, indicating success, the operation will be committed
- * permanently to the flow table.  There is one interesting subcase:
- *
- *   - If 'op' is an "add flow" operation that is replacing an existing rule in
- *     the flow table (the "victim" rule) by a new one, then the caller must
- *     have uninitialized any derived state in the victim rule, as in step 5 in
- *     the "Life Cycle" in ofproto/ofproto-provider.h.  ofoperation_complete()
- *     performs steps 6 and 7 for the victim rule, most notably by calling its
- *     ->rule_dealloc() function.
+ * permanently to the flow table.
  *
  * If 'error' is nonzero, then generally the operation will be rolled back:
  *
@@ -5056,13 +5162,6 @@ ofoperation_complete(struct ofoperation *op, enum ofperr error)
         ofopgroup_complete(group);
     }
 }
-
-struct rule *
-ofoperation_get_victim(struct ofoperation *op)
-{
-    ovs_assert(op->type == OFOPERATION_ADD);
-    return op->victim;
-}
 \f
 static uint64_t
 pick_datapath_id(const struct ofproto *ofproto)
@@ -5120,7 +5219,7 @@ choose_rule_to_evict(struct oftable *table, struct rule **rulep)
      *     group has no evictable rules.
      *
      *   - The outer loop can exit only if table's 'max_flows' is all filled up
-     *     by unevictable rules'. */
+     *     by unevictable rules. */
     HEAP_FOR_EACH (evg, size_node, &table->eviction_groups_by_size) {
         struct rule *rule;
 
@@ -5173,7 +5272,7 @@ ofproto_evict(struct ofproto *ofproto)
             ofoperation_create(group, rule,
                                OFOPERATION_DELETE, OFPRR_EVICTION);
             oftable_remove_rule(rule);
-            ofproto->ofproto_class->rule_destruct(rule);
+            ofproto->ofproto_class->rule_delete(rule);
         }
     }
     ofopgroup_submit(group);
@@ -5504,15 +5603,13 @@ oftable_remove_rule(struct rule *rule)
     ovs_rwlock_unlock(&table->cls.rwlock);
 }
 
-/* Inserts 'rule' into its oftable.  Removes any existing rule from 'rule''s
- * oftable that has an identical cls_rule.  Returns the rule that was removed,
- * if any, and otherwise NULL. */
-static struct rule *
-oftable_replace_rule(struct rule *rule)
+/* Inserts 'rule' into its oftable, which must not already contain any rule for
+ * the same cls_rule. */
+static void
+oftable_insert_rule(struct rule *rule)
 {
     struct ofproto *ofproto = rule->ofproto;
     struct oftable *table = &ofproto->tables[rule->table_id];
-    struct rule *victim;
     bool may_expire;
 
     ovs_mutex_lock(&rule->timeout_mutex);
@@ -5530,35 +5627,9 @@ oftable_replace_rule(struct rule *rule)
         list_insert(&meter->rules, &rule->meter_list_node);
     }
     ovs_rwlock_wrlock(&table->cls.rwlock);
-    victim = rule_from_cls_rule(classifier_replace(&table->cls, &rule->cr));
+    classifier_insert(&table->cls, &rule->cr);
     ovs_rwlock_unlock(&table->cls.rwlock);
-    if (victim) {
-        if (victim->meter_id) {
-            list_remove(&victim->meter_list_node);
-        }
-        cookies_remove(ofproto, victim);
-
-        ovs_mutex_lock(&ofproto->expirable_mutex);
-        if (!list_is_empty(&victim->expirable)) {
-            list_remove(&victim->expirable);
-        }
-        ovs_mutex_unlock(&ofproto->expirable_mutex);
-        eviction_group_remove_rule(victim);
-    }
     eviction_group_add_rule(rule);
-    return victim;
-}
-
-/* Removes 'old' from its oftable then, if 'new' is nonnull, inserts 'new'. */
-static void
-oftable_substitute_rule(struct rule *old, struct rule *new)
-{
-    if (new) {
-        oftable_replace_rule(new);
-    } else {
-        ovs_rwlock_wrlock(&old->evict);
-        oftable_remove_rule(old);
-    }
 }
 \f
 /* unixctl commands. */