X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=lib%2Fdpif-netdev.c;h=3d0c09fdec9d9ce51faea9489828a28b7591c0b2;hb=63be20bee256f305801c0674b29e5773355d2379;hp=443bbcb1334011e2137bc99516877660005f079a;hpb=679ba04cab922f230d6694e4c9fcf2e158b6a9e9;p=sliver-openvswitch.git

diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
index 443bbcb13..3d0c09fde 100644
--- a/lib/dpif-netdev.c
+++ b/lib/dpif-netdev.c
@@ -49,6 +49,7 @@
 #include "odp-util.h"
 #include "ofp-print.h"
 #include "ofpbuf.h"
+#include "ovs-rcu.h"
 #include "packets.h"
 #include "poll-loop.h"
 #include "random.h"
@@ -73,7 +74,6 @@ enum { MAX_FLOWS = 65536 };     /* Maximum number of flows in flow table. */
 enum { DP_NETDEV_HEADROOM = 2 + VLAN_HEADER_LEN };
 
 /* Queues. */
-enum { N_QUEUES = 2 };          /* Number of queues for dpif_recv(). */
 enum { MAX_QUEUE_LEN = 128 };   /* Maximum number of packets per queue. */
 enum { QUEUE_MASK = MAX_QUEUE_LEN - 1 };
 BUILD_ASSERT_DECL(IS_POW2(MAX_QUEUE_LEN));
@@ -90,14 +90,17 @@ struct dp_netdev_upcall {
     struct ofpbuf buf;          /* ofpbuf instance for upcall.packet. */
 };
 
-/* A queue passing packets from a struct dp_netdev to its clients.
+/* A queue passing packets from a struct dp_netdev to its clients (handlers).
  *
  *
  * Thread-safety
  * =============
  *
- * Any access at all requires the owning 'dp_netdev''s queue_mutex. */
+ * Any access at all requires the owning 'dp_netdev''s queue_rwlock and
+ * its own mutex. */
 struct dp_netdev_queue {
+    struct ovs_mutex mutex;
+    struct seq *seq;      /* Incremented whenever a packet is queued. */
     struct dp_netdev_upcall upcalls[MAX_QUEUE_LEN] OVS_GUARDED;
     unsigned int head OVS_GUARDED;
     unsigned int tail OVS_GUARDED;
@@ -118,7 +121,7 @@ struct dp_netdev_queue {
  *    port_rwlock
  *    flow_mutex
  *    cls.rwlock
- *    queue_mutex
+ *    queue_rwlock
  */
 struct dp_netdev {
     const struct dpif_class *const class;
@@ -140,10 +143,12 @@ struct dp_netdev {
 
     /* Queues.
      *
-     * Everything in 'queues' is protected by 'queue_mutex'. */
-    struct ovs_mutex queue_mutex;
-    struct dp_netdev_queue queues[N_QUEUES];
-    struct seq *queue_seq;      /* Incremented whenever a packet is queued. */
+     * 'queue_rwlock' protects the modification of 'handler_queues' and
+     * 'n_handlers'.  The queue elements are protected by its
+     * 'handler_queues''s mutex. */
+    struct fat_rwlock queue_rwlock;
+    struct dp_netdev_queue *handler_queues;
+    uint32_t n_handlers;
 
     /* Statistics.
      *
@@ -245,12 +250,6 @@ struct dp_netdev_flow {
     const struct hmap_node node; /* In owning dp_netdev's 'flow_table'. */
     const struct flow flow;      /* The flow that created this entry. */
 
-    /* Number of references.
-     * The classifier owns one reference.
-     * Any thread trying to keep a rule from being freed should hold its own
-     * reference. */
-    struct ovs_refcount ref_cnt;
-
     /* Protects members marked OVS_GUARDED.
      *
      * Acquire after datapath's flow_mutex. */
@@ -266,12 +265,10 @@ struct dp_netdev_flow {
      * Reading 'actions' requires 'mutex'.
      * Writing 'actions' requires 'mutex' and (to allow for transactions) the
      * datapath's flow_mutex. */
-    struct dp_netdev_actions *actions OVS_GUARDED;
+    OVSRCU_TYPE(struct dp_netdev_actions *) actions;
 };
 
-static struct dp_netdev_flow *dp_netdev_flow_ref(
-    const struct dp_netdev_flow *);
-static void dp_netdev_flow_unref(struct dp_netdev_flow *);
+static void dp_netdev_flow_free(struct dp_netdev_flow *);
 
 /* Contained by struct dp_netdev_flow's 'stats' member.  */
 struct dp_netdev_flow_stats {
@@ -294,8 +291,6 @@ struct dp_netdev_flow_stats {
  * 'flow' is the dp_netdev_flow for which 'flow->actions == actions') or that
  * owns a reference to 'actions->ref_cnt' (or both). */
 struct dp_netdev_actions {
-    struct ovs_refcount ref_cnt;
-
     /* These members are immutable: they do not change during the struct's
      * lifetime.  */
     struct nlattr *actions;     /* Sequence of OVS_ACTION_ATTR_* attributes. */
@@ -304,9 +299,9 @@ struct dp_netdev_actions {
 
 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
                                                    size_t);
-struct dp_netdev_actions *dp_netdev_actions_ref(
-    const struct dp_netdev_actions *);
-void dp_netdev_actions_unref(struct dp_netdev_actions *);
+struct dp_netdev_actions *dp_netdev_flow_get_actions(
+    const struct dp_netdev_flow *);
+static void dp_netdev_actions_free(struct dp_netdev_actions *);
 
 /* A thread that receives packets from some ports, looks them up in the flow
  * table, and executes the actions it finds. */
@@ -338,12 +333,15 @@ static int do_add_port(struct dp_netdev *dp, const char *devname,
     OVS_REQ_WRLOCK(dp->port_rwlock);
 static int do_del_port(struct dp_netdev *dp, odp_port_t port_no)
     OVS_REQ_WRLOCK(dp->port_rwlock);
+static void dp_netdev_destroy_all_queues(struct dp_netdev *dp)
+    OVS_REQ_WRLOCK(dp->queue_rwlock);
 static int dpif_netdev_open(const struct dpif_class *, const char *name,
                             bool create, struct dpif **);
 static int dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *,
-                                    int queue_no, const struct flow *,
-                                    const struct nlattr *userdata)
-    OVS_EXCLUDED(dp->queue_mutex);
+                                      int queue_no, int type,
+                                      const struct flow *,
+                                      const struct nlattr *userdata)
+    OVS_EXCLUDED(dp->queue_rwlock);
 static void dp_netdev_execute_actions(struct dp_netdev *dp,
                                       const struct flow *, struct ofpbuf *,
                                       struct pkt_metadata *,
@@ -461,7 +459,6 @@ create_dp_netdev(const char *name, const struct dpif_class *class,
 {
     struct dp_netdev *dp;
     int error;
-    int i;
 
     dp = xzalloc(sizeof *dp);
     shash_add(&dp_netdevs, name, dp);
@@ -475,13 +472,7 @@ create_dp_netdev(const char *name, const struct dpif_class *class,
     classifier_init(&dp->cls, NULL);
     hmap_init(&dp->flow_table);
 
-    ovs_mutex_init(&dp->queue_mutex);
-    ovs_mutex_lock(&dp->queue_mutex);
-    for (i = 0; i < N_QUEUES; i++) {
-        dp->queues[i].head = dp->queues[i].tail = 0;
-    }
-    ovs_mutex_unlock(&dp->queue_mutex);
-    dp->queue_seq = seq_create();
+    fat_rwlock_init(&dp->queue_rwlock);
 
     ovsthread_stats_init(&dp->stats);
 
@@ -529,20 +520,21 @@ dpif_netdev_open(const struct dpif_class *class, const char *name,
 
 static void
 dp_netdev_purge_queues(struct dp_netdev *dp)
+    OVS_REQ_WRLOCK(dp->queue_rwlock)
 {
     int i;
 
-    ovs_mutex_lock(&dp->queue_mutex);
-    for (i = 0; i < N_QUEUES; i++) {
-        struct dp_netdev_queue *q = &dp->queues[i];
+    for (i = 0; i < dp->n_handlers; i++) {
+        struct dp_netdev_queue *q = &dp->handler_queues[i];
 
+        ovs_mutex_lock(&q->mutex);
         while (q->tail != q->head) {
             struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
             ofpbuf_uninit(&u->upcall.packet);
             ofpbuf_uninit(&u->buf);
         }
+        ovs_mutex_unlock(&q->mutex);
     }
-    ovs_mutex_unlock(&dp->queue_mutex);
 }
 
 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
@@ -573,9 +565,11 @@ dp_netdev_free(struct dp_netdev *dp)
     }
     ovsthread_stats_destroy(&dp->stats);
 
-    dp_netdev_purge_queues(dp);
-    seq_destroy(dp->queue_seq);
-    ovs_mutex_destroy(&dp->queue_mutex);
+    fat_rwlock_wrlock(&dp->queue_rwlock);
+    dp_netdev_destroy_all_queues(dp);
+    fat_rwlock_unlock(&dp->queue_rwlock);
+
+    fat_rwlock_destroy(&dp->queue_rwlock);
 
     classifier_destroy(&dp->cls);
     hmap_destroy(&dp->flow_table);
@@ -870,6 +864,24 @@ dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
     return error;
 }
 
+static void
+dp_netdev_flow_free(struct dp_netdev_flow *flow)
+{
+    struct dp_netdev_flow_stats *bucket;
+    size_t i;
+
+    OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
+        ovs_mutex_destroy(&bucket->mutex);
+        free_cacheline(bucket);
+    }
+    ovsthread_stats_destroy(&flow->stats);
+
+    cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
+    dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
+    ovs_mutex_destroy(&flow->mutex);
+    free(flow);
+}
+
 static void
 dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
     OVS_REQ_WRLOCK(dp->cls.rwlock)
@@ -880,39 +892,7 @@ dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
 
     classifier_remove(&dp->cls, cr);
     hmap_remove(&dp->flow_table, node);
-    dp_netdev_flow_unref(flow);
-}
-
-static struct dp_netdev_flow *
-dp_netdev_flow_ref(const struct dp_netdev_flow *flow_)
-{
-    struct dp_netdev_flow *flow = CONST_CAST(struct dp_netdev_flow *, flow_);
-    if (flow) {
-        ovs_refcount_ref(&flow->ref_cnt);
-    }
-    return flow;
-}
-
-static void
-dp_netdev_flow_unref(struct dp_netdev_flow *flow)
-{
-    if (flow && ovs_refcount_unref(&flow->ref_cnt) == 1) {
-        struct dp_netdev_flow_stats *bucket;
-        size_t i;
-
-        OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
-            ovs_mutex_destroy(&bucket->mutex);
-            free_cacheline(bucket);
-        }
-        ovsthread_stats_destroy(&flow->stats);
-
-        cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
-        ovs_mutex_lock(&flow->mutex);
-        dp_netdev_actions_unref(flow->actions);
-        ovs_mutex_unlock(&flow->mutex);
-        ovs_mutex_destroy(&flow->mutex);
-        free(flow);
-    }
+    ovsrcu_postpone(dp_netdev_flow_free, flow);
 }
 
 static void
@@ -1031,7 +1011,6 @@ dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct flow *flow)
 
     fat_rwlock_rdlock(&dp->cls.rwlock);
     netdev_flow = dp_netdev_flow_cast(classifier_lookup(&dp->cls, flow, NULL));
-    dp_netdev_flow_ref(netdev_flow);
     fat_rwlock_unlock(&dp->cls.rwlock);
 
     return netdev_flow;
@@ -1046,7 +1025,7 @@ dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
     HMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
                              &dp->flow_table) {
         if (flow_equal(&netdev_flow->flow, flow)) {
-            return dp_netdev_flow_ref(netdev_flow);
+            return netdev_flow;
         }
     }
 
@@ -1177,25 +1156,17 @@ dpif_netdev_flow_get(const struct dpif *dpif,
     fat_rwlock_unlock(&dp->cls.rwlock);
 
     if (netdev_flow) {
-        struct dp_netdev_actions *actions = NULL;
-
         if (stats) {
             get_dpif_flow_stats(netdev_flow, stats);
         }
 
-        ovs_mutex_lock(&netdev_flow->mutex);
         if (actionsp) {
-            actions = dp_netdev_actions_ref(netdev_flow->actions);
-        }
-        ovs_mutex_unlock(&netdev_flow->mutex);
-
-        dp_netdev_flow_unref(netdev_flow);
+            struct dp_netdev_actions *actions;
 
-        if (actionsp) {
+            actions = dp_netdev_flow_get_actions(netdev_flow);
             *actionsp = ofpbuf_clone_data(actions->actions, actions->size);
-            dp_netdev_actions_unref(actions);
         }
-    } else {
+     } else {
         error = ENOENT;
     }
 
@@ -1214,14 +1185,13 @@ dp_netdev_flow_add(struct dp_netdev *dp, const struct flow *flow,
 
     netdev_flow = xzalloc(sizeof *netdev_flow);
     *CONST_CAST(struct flow *, &netdev_flow->flow) = *flow;
-    ovs_refcount_init(&netdev_flow->ref_cnt);
 
     ovs_mutex_init(&netdev_flow->mutex);
-    ovs_mutex_lock(&netdev_flow->mutex);
 
     ovsthread_stats_init(&netdev_flow->stats);
 
-    netdev_flow->actions = dp_netdev_actions_create(actions, actions_len);
+    ovsrcu_set(&netdev_flow->actions,
+               dp_netdev_actions_create(actions, actions_len));
 
     match_init(&match, flow, wc);
     cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr),
@@ -1234,8 +1204,6 @@ dp_netdev_flow_add(struct dp_netdev *dp, const struct flow *flow,
                 flow_hash(flow, 0));
     fat_rwlock_unlock(&dp->cls.rwlock);
 
-    ovs_mutex_unlock(&netdev_flow->mutex);
-
     return 0;
 }
 
@@ -1300,10 +1268,8 @@ dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
             new_actions = dp_netdev_actions_create(put->actions,
                                                    put->actions_len);
 
-            ovs_mutex_lock(&netdev_flow->mutex);
-            old_actions = netdev_flow->actions;
-            netdev_flow->actions = new_actions;
-            ovs_mutex_unlock(&netdev_flow->mutex);
+            old_actions = dp_netdev_flow_get_actions(netdev_flow);
+            ovsrcu_set(&netdev_flow->actions, new_actions);
 
             if (put->stats) {
                 get_dpif_flow_stats(netdev_flow, put->stats);
@@ -1312,14 +1278,13 @@ dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
                 clear_stats(netdev_flow);
             }
 
-            dp_netdev_actions_unref(old_actions);
+            ovsrcu_postpone(dp_netdev_actions_free, old_actions);
         } else if (put->flags & DPIF_FP_CREATE) {
             error = EEXIST;
         } else {
             /* Overlapping flow. */
             error = EINVAL;
         }
-        dp_netdev_flow_unref(netdev_flow);
     }
     ovs_mutex_unlock(&dp->flow_mutex);
 
@@ -1347,7 +1312,6 @@ dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
             get_dpif_flow_stats(netdev_flow, del->stats);
         }
         dp_netdev_remove_flow(dp, netdev_flow);
-        dp_netdev_flow_unref(netdev_flow);
     } else {
         error = ENOENT;
     }
@@ -1385,7 +1349,6 @@ dpif_netdev_flow_dump_state_uninit(void *state_)
 {
     struct dp_netdev_flow_state *state = state_;
 
-    dp_netdev_actions_unref(state->actions);
     free(state);
 }
 
@@ -1402,6 +1365,7 @@ dpif_netdev_flow_dump_start(const struct dpif *dpif OVS_UNUSED, void **iterp)
     return 0;
 }
 
+/* XXX the caller must use 'actions' without quiescing */
 static int
 dpif_netdev_flow_dump_next(const struct dpif *dpif, void *iter_, void *state_,
                            const struct nlattr **key, size_t *key_len,
@@ -1424,7 +1388,6 @@ dpif_netdev_flow_dump_next(const struct dpif *dpif, void *iter_, void *state_,
         node = hmap_at_position(&dp->flow_table, &iter->bucket, &iter->offset);
         if (node) {
             netdev_flow = CONTAINER_OF(node, struct dp_netdev_flow, node);
-            dp_netdev_flow_ref(netdev_flow);
         }
         fat_rwlock_unlock(&dp->cls.rwlock);
         if (!node) {
@@ -1462,16 +1425,13 @@ dpif_netdev_flow_dump_next(const struct dpif *dpif, void *iter_, void *state_,
     }
 
     if (actions || stats) {
-        dp_netdev_actions_unref(state->actions);
         state->actions = NULL;
 
-        ovs_mutex_lock(&netdev_flow->mutex);
         if (actions) {
-            state->actions = dp_netdev_actions_ref(netdev_flow->actions);
+            state->actions = dp_netdev_flow_get_actions(netdev_flow);
             *actions = state->actions->actions;
             *actions_len = state->actions->size;
         }
-        ovs_mutex_unlock(&netdev_flow->mutex);
 
         if (stats) {
             get_dpif_flow_stats(netdev_flow, &state->stats);
@@ -1479,8 +1439,6 @@ dpif_netdev_flow_dump_next(const struct dpif *dpif, void *iter_, void *state_,
         }
     }
 
-    dp_netdev_flow_unref(netdev_flow);
-
     return 0;
 }
 
@@ -1517,9 +1475,77 @@ dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
     return 0;
 }
 
+static void
+dp_netdev_destroy_all_queues(struct dp_netdev *dp)
+    OVS_REQ_WRLOCK(dp->queue_rwlock)
+{
+    size_t i;
+
+    dp_netdev_purge_queues(dp);
+
+    for (i = 0; i < dp->n_handlers; i++) {
+        struct dp_netdev_queue *q = &dp->handler_queues[i];
+
+        ovs_mutex_destroy(&q->mutex);
+        seq_destroy(q->seq);
+    }
+    free(dp->handler_queues);
+    dp->handler_queues = NULL;
+    dp->n_handlers = 0;
+}
+
+static void
+dp_netdev_refresh_queues(struct dp_netdev *dp, uint32_t n_handlers)
+    OVS_REQ_WRLOCK(dp->queue_rwlock)
+{
+    if (dp->n_handlers != n_handlers) {
+        size_t i;
+
+        dp_netdev_destroy_all_queues(dp);
+
+        dp->n_handlers = n_handlers;
+        dp->handler_queues = xzalloc(n_handlers * sizeof *dp->handler_queues);
+
+        for (i = 0; i < n_handlers; i++) {
+            struct dp_netdev_queue *q = &dp->handler_queues[i];
+
+            ovs_mutex_init(&q->mutex);
+            q->seq = seq_create();
+        }
+    }
+}
+
+static int
+dpif_netdev_recv_set(struct dpif *dpif, bool enable)
+{
+    struct dp_netdev *dp = get_dp_netdev(dpif);
+
+    if ((dp->handler_queues != NULL) == enable) {
+        return 0;
+    }
+
+    fat_rwlock_wrlock(&dp->queue_rwlock);
+    if (!enable) {
+        dp_netdev_destroy_all_queues(dp);
+    } else {
+        dp_netdev_refresh_queues(dp, 1);
+    }
+    fat_rwlock_unlock(&dp->queue_rwlock);
+
+    return 0;
+}
+
 static int
-dpif_netdev_recv_set(struct dpif *dpif OVS_UNUSED, bool enable OVS_UNUSED)
+dpif_netdev_handlers_set(struct dpif *dpif, uint32_t n_handlers)
 {
+    struct dp_netdev *dp = get_dp_netdev(dpif);
+
+    fat_rwlock_wrlock(&dp->queue_rwlock);
+    if (dp->handler_queues) {
+        dp_netdev_refresh_queues(dp, n_handlers);
+    }
+    fat_rwlock_unlock(&dp->queue_rwlock);
+
     return 0;
 }
 
@@ -1531,62 +1557,86 @@ dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
     return 0;
 }
 
-static struct dp_netdev_queue *
-find_nonempty_queue(struct dp_netdev *dp)
-    OVS_REQUIRES(dp->queue_mutex)
+static bool
+dp_netdev_recv_check(const struct dp_netdev *dp, const uint32_t handler_id)
+    OVS_REQ_RDLOCK(dp->queue_rwlock)
 {
-    int i;
+    static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
 
-    for (i = 0; i < N_QUEUES; i++) {
-        struct dp_netdev_queue *q = &dp->queues[i];
-        if (q->head != q->tail) {
-            return q;
-        }
+    if (!dp->handler_queues) {
+        VLOG_WARN_RL(&rl, "receiving upcall disabled");
+        return false;
     }
-    return NULL;
+
+    if (handler_id >= dp->n_handlers) {
+        VLOG_WARN_RL(&rl, "handler index out of bound");
+        return false;
+    }
+
+    return true;
 }
 
 static int
-dpif_netdev_recv(struct dpif *dpif, struct dpif_upcall *upcall,
-                 struct ofpbuf *buf)
+dpif_netdev_recv(struct dpif *dpif, uint32_t handler_id,
+                 struct dpif_upcall *upcall, struct ofpbuf *buf)
 {
     struct dp_netdev *dp = get_dp_netdev(dpif);
     struct dp_netdev_queue *q;
-    int error;
+    int error = 0;
+
+    fat_rwlock_rdlock(&dp->queue_rwlock);
+
+    if (!dp_netdev_recv_check(dp, handler_id)) {
+        error = EAGAIN;
+        goto out;
+    }
 
-    ovs_mutex_lock(&dp->queue_mutex);
-    q = find_nonempty_queue(dp);
-    if (q) {
+    q = &dp->handler_queues[handler_id];
+    ovs_mutex_lock(&q->mutex);
+    if (q->head != q->tail) {
         struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
 
         *upcall = u->upcall;
 
         ofpbuf_uninit(buf);
         *buf = u->buf;
-
-        error = 0;
     } else {
         error = EAGAIN;
     }
-    ovs_mutex_unlock(&dp->queue_mutex);
+    ovs_mutex_unlock(&q->mutex);
+
+out:
+    fat_rwlock_unlock(&dp->queue_rwlock);
 
     return error;
 }
 
 static void
-dpif_netdev_recv_wait(struct dpif *dpif)
+dpif_netdev_recv_wait(struct dpif *dpif, uint32_t handler_id)
 {
     struct dp_netdev *dp = get_dp_netdev(dpif);
+    struct dp_netdev_queue *q;
     uint64_t seq;
 
-    ovs_mutex_lock(&dp->queue_mutex);
-    seq = seq_read(dp->queue_seq);
-    if (find_nonempty_queue(dp)) {
+    fat_rwlock_rdlock(&dp->queue_rwlock);
+
+    if (!dp_netdev_recv_check(dp, handler_id)) {
+        goto out;
+    }
+
+    q = &dp->handler_queues[handler_id];
+    ovs_mutex_lock(&q->mutex);
+    seq = seq_read(q->seq);
+    if (q->head != q->tail) {
         poll_immediate_wake();
     } else {
-        seq_wait(dp->queue_seq, seq);
+        seq_wait(q->seq, seq);
     }
-    ovs_mutex_unlock(&dp->queue_mutex);
+
+    ovs_mutex_unlock(&q->mutex);
+
+out:
+    fat_rwlock_unlock(&dp->queue_rwlock);
 }
 
 static void
@@ -1594,7 +1644,9 @@ dpif_netdev_recv_purge(struct dpif *dpif)
 {
     struct dpif_netdev *dpif_netdev = dpif_netdev_cast(dpif);
 
+    fat_rwlock_wrlock(&dpif_netdev->dp->queue_rwlock);
     dp_netdev_purge_queues(dpif_netdev->dp);
+    fat_rwlock_unlock(&dpif_netdev->dp->queue_rwlock);
 }
 
 /* Creates and returns a new 'struct dp_netdev_actions', with a reference count
@@ -1606,35 +1658,23 @@ dp_netdev_actions_create(const struct nlattr *actions, size_t size)
     struct dp_netdev_actions *netdev_actions;
 
     netdev_actions = xmalloc(sizeof *netdev_actions);
-    ovs_refcount_init(&netdev_actions->ref_cnt);
     netdev_actions->actions = xmemdup(actions, size);
     netdev_actions->size = size;
 
     return netdev_actions;
 }
 
-/* Increments 'actions''s refcount. */
 struct dp_netdev_actions *
-dp_netdev_actions_ref(const struct dp_netdev_actions *actions_)
+dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
 {
-    struct dp_netdev_actions *actions;
-
-    actions = CONST_CAST(struct dp_netdev_actions *, actions_);
-    if (actions) {
-        ovs_refcount_ref(&actions->ref_cnt);
-    }
-    return actions;
+    return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
 }
 
-/* Decrements 'actions''s refcount and frees 'actions' if the refcount reaches
- * 0. */
-void
-dp_netdev_actions_unref(struct dp_netdev_actions *actions)
+static void
+dp_netdev_actions_free(struct dp_netdev_actions *actions)
 {
-    if (actions && ovs_refcount_unref(&actions->ref_cnt) == 1) {
-        free(actions->actions);
-        free(actions);
-    }
+    free(actions->actions);
+    free(actions);
 }
 
 static void *
@@ -1767,9 +1807,10 @@ dp_netdev_flow_stats_new_cb(void)
 
 static void
 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow,
-                    const struct ofpbuf *packet)
+                    const struct ofpbuf *packet,
+                    const struct flow *key)
 {
-    uint16_t tcp_flags = packet_get_tcp_flags(packet, &netdev_flow->flow);
+    uint16_t tcp_flags = ntohs(key->tcp_flags);
     long long int now = time_msec();
     struct dp_netdev_flow_stats *bucket;
 
@@ -1819,40 +1860,39 @@ dp_netdev_port_input(struct dp_netdev *dp, struct ofpbuf *packet,
     if (netdev_flow) {
         struct dp_netdev_actions *actions;
 
-        dp_netdev_flow_used(netdev_flow, packet);
-
-        ovs_mutex_lock(&netdev_flow->mutex);
-        actions = dp_netdev_actions_ref(netdev_flow->actions);
-        ovs_mutex_unlock(&netdev_flow->mutex);
+        dp_netdev_flow_used(netdev_flow, packet, &key);
 
+        actions = dp_netdev_flow_get_actions(netdev_flow);
         dp_netdev_execute_actions(dp, &key, packet, md,
                                   actions->actions, actions->size);
-        dp_netdev_actions_unref(actions);
-        dp_netdev_flow_unref(netdev_flow);
         dp_netdev_count_packet(dp, DP_STAT_HIT);
-    } else {
+    } else if (dp->handler_queues) {
         dp_netdev_count_packet(dp, DP_STAT_MISS);
-        dp_netdev_output_userspace(dp, packet, DPIF_UC_MISS, &key, NULL);
+        dp_netdev_output_userspace(dp, packet,
+                                   flow_hash_5tuple(&key, 0) % dp->n_handlers,
+                                   DPIF_UC_MISS, &key, NULL);
     }
 }
 
 static int
 dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *packet,
-                           int queue_no, const struct flow *flow,
+                           int queue_no, int type, const struct flow *flow,
                            const struct nlattr *userdata)
-    OVS_EXCLUDED(dp->queue_mutex)
+    OVS_EXCLUDED(dp->queue_rwlock)
 {
-    struct dp_netdev_queue *q = &dp->queues[queue_no];
+    struct dp_netdev_queue *q;
     int error;
 
-    ovs_mutex_lock(&dp->queue_mutex);
+    fat_rwlock_rdlock(&dp->queue_rwlock);
+    q = &dp->handler_queues[queue_no];
+    ovs_mutex_lock(&q->mutex);
     if (q->head - q->tail < MAX_QUEUE_LEN) {
         struct dp_netdev_upcall *u = &q->upcalls[q->head++ & QUEUE_MASK];
         struct dpif_upcall *upcall = &u->upcall;
         struct ofpbuf *buf = &u->buf;
         size_t buf_size;
 
-        upcall->type = queue_no;
+        upcall->type = type;
 
         /* Allocate buffer big enough for everything. */
         buf_size = ODPUTIL_FLOW_KEY_BYTES;
@@ -1877,14 +1917,15 @@ dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *packet,
         upcall->packet = *packet;
         ofpbuf_use(packet, NULL, 0);
 
-        seq_change(dp->queue_seq);
+        seq_change(q->seq);
 
         error = 0;
     } else {
         dp_netdev_count_packet(dp, DP_STAT_LOST);
         error = ENOBUFS;
     }
-    ovs_mutex_unlock(&dp->queue_mutex);
+    ovs_mutex_unlock(&q->mutex);
+    fat_rwlock_unlock(&dp->queue_rwlock);
 
     return error;
 }
@@ -1921,7 +1962,10 @@ dp_execute_cb(void *aux_, struct ofpbuf *packet,
         if (!may_steal) {
             packet = ofpbuf_clone_with_headroom(packet, DP_NETDEV_HEADROOM);
         }
-        dp_netdev_output_userspace(aux->dp, packet, DPIF_UC_ACTION, aux->key,
+        dp_netdev_output_userspace(aux->dp, packet,
+                                   flow_hash_5tuple(aux->key, 0)
+                                       % aux->dp->n_handlers,
+                                   DPIF_UC_ACTION, aux->key,
                                    userdata);
         if (!may_steal) {
             ofpbuf_uninit(packet);
@@ -1984,6 +2028,7 @@ const struct dpif_class dpif_netdev_class = {
     dpif_netdev_execute,
     NULL,                       /* operate */
     dpif_netdev_recv_set,
+    dpif_netdev_handlers_set,
     dpif_netdev_queue_to_priority,
     dpif_netdev_recv,
     dpif_netdev_recv_wait,