openvswitch: Add meter infrastructure
authorAndy Zhou <azhou@ovn.org>
Fri, 10 Nov 2017 20:09:42 +0000 (12:09 -0800)
committerDavid S. Miller <davem@davemloft.net>
Mon, 13 Nov 2017 01:37:07 +0000 (10:37 +0900)
OVS kernel datapath so far does not support Openflow meter action.
This is the first stab at adding kernel datapath meter support.
This implementation supports only drop band type.

Signed-off-by: Andy Zhou <azhou@ovn.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/openvswitch/Makefile
net/openvswitch/datapath.c
net/openvswitch/datapath.h
net/openvswitch/meter.c [new file with mode: 0644]
net/openvswitch/meter.h [new file with mode: 0644]

index 299f447..41109c3 100644 (file)
@@ -12,6 +12,7 @@ openvswitch-y := \
        flow.o \
        flow_netlink.o \
        flow_table.o \
+       meter.o \
        vport.o \
        vport-internal_dev.o \
        vport-netdev.o
index 6e09803..0dab33f 100644 (file)
@@ -55,6 +55,7 @@
 #include "flow.h"
 #include "flow_table.h"
 #include "flow_netlink.h"
+#include "meter.h"
 #include "vport-internal_dev.h"
 #include "vport-netdev.h"
 
@@ -174,6 +175,7 @@ static void destroy_dp_rcu(struct rcu_head *rcu)
        ovs_flow_tbl_destroy(&dp->table);
        free_percpu(dp->stats_percpu);
        kfree(dp->ports);
+       ovs_meters_exit(dp);
        kfree(dp);
 }
 
@@ -1572,6 +1574,10 @@ static int ovs_dp_cmd_new(struct sk_buff *skb, struct genl_info *info)
        for (i = 0; i < DP_VPORT_HASH_BUCKETS; i++)
                INIT_HLIST_HEAD(&dp->ports[i]);
 
+       err = ovs_meters_init(dp);
+       if (err)
+               goto err_destroy_ports_array;
+
        /* Set up our datapath device. */
        parms.name = nla_data(a[OVS_DP_ATTR_NAME]);
        parms.type = OVS_VPORT_TYPE_INTERNAL;
@@ -1600,7 +1606,7 @@ static int ovs_dp_cmd_new(struct sk_buff *skb, struct genl_info *info)
                                ovs_dp_reset_user_features(skb, info);
                }
 
-               goto err_destroy_ports_array;
+               goto err_destroy_meters;
        }
 
        err = ovs_dp_cmd_fill_info(dp, reply, info->snd_portid,
@@ -1615,8 +1621,10 @@ static int ovs_dp_cmd_new(struct sk_buff *skb, struct genl_info *info)
        ovs_notify(&dp_datapath_genl_family, reply, info);
        return 0;
 
-err_destroy_ports_array:
+err_destroy_meters:
        ovs_unlock();
+       ovs_meters_exit(dp);
+err_destroy_ports_array:
        kfree(dp->ports);
 err_destroy_percpu:
        free_percpu(dp->stats_percpu);
@@ -2265,6 +2273,7 @@ static struct genl_family * const dp_genl_families[] = {
        &dp_vport_genl_family,
        &dp_flow_genl_family,
        &dp_packet_genl_family,
+       &dp_meter_genl_family,
 };
 
 static void dp_unregister_genl(int n_families)
@@ -2445,3 +2454,4 @@ MODULE_ALIAS_GENL_FAMILY(OVS_DATAPATH_FAMILY);
 MODULE_ALIAS_GENL_FAMILY(OVS_VPORT_FAMILY);
 MODULE_ALIAS_GENL_FAMILY(OVS_FLOW_FAMILY);
 MODULE_ALIAS_GENL_FAMILY(OVS_PACKET_FAMILY);
+MODULE_ALIAS_GENL_FAMILY(OVS_METER_FAMILY);
index 954c4ed..5d2997b 100644 (file)
@@ -92,6 +92,9 @@ struct datapath {
        u32 user_features;
 
        u32 max_headroom;
+
+       /* Switch meters. */
+       struct hlist_head *meters;
 };
 
 /**
diff --git a/net/openvswitch/meter.c b/net/openvswitch/meter.c
new file mode 100644 (file)
index 0000000..2a5ba35
--- /dev/null
@@ -0,0 +1,604 @@
+/*
+ * Copyright (c) 2017 Nicira, Inc.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ */
+
+#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
+
+#include <linux/if.h>
+#include <linux/skbuff.h>
+#include <linux/ip.h>
+#include <linux/kernel.h>
+#include <linux/openvswitch.h>
+#include <linux/netlink.h>
+#include <linux/rculist.h>
+
+#include <net/netlink.h>
+#include <net/genetlink.h>
+
+#include "datapath.h"
+#include "meter.h"
+
+#define METER_HASH_BUCKETS 1024
+
+static const struct nla_policy meter_policy[OVS_METER_ATTR_MAX + 1] = {
+       [OVS_METER_ATTR_ID] = { .type = NLA_U32, },
+       [OVS_METER_ATTR_KBPS] = { .type = NLA_FLAG },
+       [OVS_METER_ATTR_STATS] = { .len = sizeof(struct ovs_flow_stats) },
+       [OVS_METER_ATTR_BANDS] = { .type = NLA_NESTED },
+       [OVS_METER_ATTR_USED] = { .type = NLA_U64 },
+       [OVS_METER_ATTR_CLEAR] = { .type = NLA_FLAG },
+       [OVS_METER_ATTR_MAX_METERS] = { .type = NLA_U32 },
+       [OVS_METER_ATTR_MAX_BANDS] = { .type = NLA_U32 },
+};
+
+static const struct nla_policy band_policy[OVS_BAND_ATTR_MAX + 1] = {
+       [OVS_BAND_ATTR_TYPE] = { .type = NLA_U32, },
+       [OVS_BAND_ATTR_RATE] = { .type = NLA_U32, },
+       [OVS_BAND_ATTR_BURST] = { .type = NLA_U32, },
+       [OVS_BAND_ATTR_STATS] = { .len = sizeof(struct ovs_flow_stats) },
+};
+
+static void rcu_free_ovs_meter_callback(struct rcu_head *rcu)
+{
+       struct dp_meter *meter = container_of(rcu, struct dp_meter, rcu);
+
+       kfree(meter);
+}
+
+static void ovs_meter_free(struct dp_meter *meter)
+{
+       if (!meter)
+               return;
+
+       call_rcu(&meter->rcu, rcu_free_ovs_meter_callback);
+}
+
+static struct hlist_head *meter_hash_bucket(const struct datapath *dp,
+                                           u32 meter_id)
+{
+       return &dp->meters[meter_id & (METER_HASH_BUCKETS - 1)];
+}
+
+/* Call with ovs_mutex or RCU read lock. */
+static struct dp_meter *lookup_meter(const struct datapath *dp,
+                                    u32 meter_id)
+{
+       struct dp_meter *meter;
+       struct hlist_head *head;
+
+       head = meter_hash_bucket(dp, meter_id);
+       hlist_for_each_entry_rcu(meter, head, dp_hash_node) {
+               if (meter->id == meter_id)
+                       return meter;
+       }
+       return NULL;
+}
+
+static void attach_meter(struct datapath *dp, struct dp_meter *meter)
+{
+       struct hlist_head *head = meter_hash_bucket(dp, meter->id);
+
+       hlist_add_head_rcu(&meter->dp_hash_node, head);
+}
+
+static void detach_meter(struct dp_meter *meter)
+{
+       ASSERT_OVSL();
+       if (meter)
+               hlist_del_rcu(&meter->dp_hash_node);
+}
+
+static struct sk_buff *
+ovs_meter_cmd_reply_start(struct genl_info *info, u8 cmd,
+                         struct ovs_header **ovs_reply_header)
+{
+       struct sk_buff *skb;
+       struct ovs_header *ovs_header = info->userhdr;
+
+       skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
+       if (!skb)
+               return ERR_PTR(-ENOMEM);
+
+       *ovs_reply_header = genlmsg_put(skb, info->snd_portid,
+                                       info->snd_seq,
+                                       &dp_meter_genl_family, 0, cmd);
+       if (!ovs_reply_header) {
+               nlmsg_free(skb);
+               return ERR_PTR(-EMSGSIZE);
+       }
+       (*ovs_reply_header)->dp_ifindex = ovs_header->dp_ifindex;
+
+       return skb;
+}
+
+static int ovs_meter_cmd_reply_stats(struct sk_buff *reply, u32 meter_id,
+                                    struct dp_meter *meter)
+{
+       struct nlattr *nla;
+       struct dp_meter_band *band;
+       u16 i;
+
+       if (nla_put_u32(reply, OVS_METER_ATTR_ID, meter_id))
+               goto error;
+
+       if (!meter)
+               return 0;
+
+       if (nla_put(reply, OVS_METER_ATTR_STATS,
+                   sizeof(struct ovs_flow_stats), &meter->stats) ||
+           nla_put_u64_64bit(reply, OVS_METER_ATTR_USED, meter->used,
+                             OVS_METER_ATTR_PAD))
+               goto error;
+
+       nla = nla_nest_start(reply, OVS_METER_ATTR_BANDS);
+       if (!nla)
+               goto error;
+
+       band = meter->bands;
+
+       for (i = 0; i < meter->n_bands; ++i, ++band) {
+               struct nlattr *band_nla;
+
+               band_nla = nla_nest_start(reply, OVS_BAND_ATTR_UNSPEC);
+               if (!band_nla || nla_put(reply, OVS_BAND_ATTR_STATS,
+                                        sizeof(struct ovs_flow_stats),
+                                        &band->stats))
+                       goto error;
+               nla_nest_end(reply, band_nla);
+       }
+       nla_nest_end(reply, nla);
+
+       return 0;
+error:
+       return -EMSGSIZE;
+}
+
+static int ovs_meter_cmd_features(struct sk_buff *skb, struct genl_info *info)
+{
+       struct sk_buff *reply;
+       struct ovs_header *ovs_reply_header;
+       struct nlattr *nla, *band_nla;
+       int err;
+
+       reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_FEATURES,
+                                         &ovs_reply_header);
+       if (!reply)
+               return PTR_ERR(reply);
+
+       if (nla_put_u32(reply, OVS_METER_ATTR_MAX_METERS, U32_MAX) ||
+           nla_put_u32(reply, OVS_METER_ATTR_MAX_BANDS, DP_MAX_BANDS))
+               goto nla_put_failure;
+
+       nla = nla_nest_start(reply, OVS_METER_ATTR_BANDS);
+       if (!nla)
+               goto nla_put_failure;
+
+       band_nla = nla_nest_start(reply, OVS_BAND_ATTR_UNSPEC);
+       if (!band_nla)
+               goto nla_put_failure;
+       /* Currently only DROP band type is supported. */
+       if (nla_put_u32(reply, OVS_BAND_ATTR_TYPE, OVS_METER_BAND_TYPE_DROP))
+               goto nla_put_failure;
+       nla_nest_end(reply, band_nla);
+       nla_nest_end(reply, nla);
+
+       genlmsg_end(reply, ovs_reply_header);
+       return genlmsg_reply(reply, info);
+
+nla_put_failure:
+       nlmsg_free(reply);
+       err = -EMSGSIZE;
+       return err;
+}
+
+static struct dp_meter *dp_meter_create(struct nlattr **a)
+{
+       struct nlattr *nla;
+       int rem;
+       u16 n_bands = 0;
+       struct dp_meter *meter;
+       struct dp_meter_band *band;
+       int err;
+
+       /* Validate attributes, count the bands. */
+       if (!a[OVS_METER_ATTR_BANDS])
+               return ERR_PTR(-EINVAL);
+
+       nla_for_each_nested(nla, a[OVS_METER_ATTR_BANDS], rem)
+               if (++n_bands > DP_MAX_BANDS)
+                       return ERR_PTR(-EINVAL);
+
+       /* Allocate and set up the meter before locking anything. */
+       meter = kzalloc(n_bands * sizeof(struct dp_meter_band) +
+                       sizeof(*meter), GFP_KERNEL);
+       if (!meter)
+               return ERR_PTR(-ENOMEM);
+
+       meter->used = div_u64(ktime_get_ns(), 1000 * 1000);
+       meter->kbps = a[OVS_METER_ATTR_KBPS] ? 1 : 0;
+       meter->keep_stats = !a[OVS_METER_ATTR_CLEAR];
+       spin_lock_init(&meter->lock);
+       if (meter->keep_stats && a[OVS_METER_ATTR_STATS]) {
+               meter->stats = *(struct ovs_flow_stats *)
+                       nla_data(a[OVS_METER_ATTR_STATS]);
+       }
+       meter->n_bands = n_bands;
+
+       /* Set up meter bands. */
+       band = meter->bands;
+       nla_for_each_nested(nla, a[OVS_METER_ATTR_BANDS], rem) {
+               struct nlattr *attr[OVS_BAND_ATTR_MAX + 1];
+               u32 band_max_delta_t;
+
+               err = nla_parse((struct nlattr **)&attr, OVS_BAND_ATTR_MAX,
+                               nla_data(nla), nla_len(nla), band_policy,
+                               NULL);
+               if (err)
+                       goto exit_free_meter;
+
+               if (!attr[OVS_BAND_ATTR_TYPE] ||
+                   !attr[OVS_BAND_ATTR_RATE] ||
+                   !attr[OVS_BAND_ATTR_BURST]) {
+                       err = -EINVAL;
+                       goto exit_free_meter;
+               }
+
+               band->type = nla_get_u32(attr[OVS_BAND_ATTR_TYPE]);
+               band->rate = nla_get_u32(attr[OVS_BAND_ATTR_RATE]);
+               band->burst_size = nla_get_u32(attr[OVS_BAND_ATTR_BURST]);
+               /* Figure out max delta_t that is enough to fill any bucket.
+                * Keep max_delta_t size to the bucket units:
+                * pkts => 1/1000 packets, kilobits => bits.
+                */
+               band_max_delta_t = (band->burst_size + band->rate) * 1000;
+               /* Start with a full bucket. */
+               band->bucket = band_max_delta_t;
+               if (band_max_delta_t > meter->max_delta_t)
+                       meter->max_delta_t = band_max_delta_t;
+               band++;
+       }
+
+       return meter;
+
+exit_free_meter:
+       kfree(meter);
+       return ERR_PTR(err);
+}
+
+static int ovs_meter_cmd_set(struct sk_buff *skb, struct genl_info *info)
+{
+       struct nlattr **a = info->attrs;
+       struct dp_meter *meter, *old_meter;
+       struct sk_buff *reply;
+       struct ovs_header *ovs_reply_header;
+       struct ovs_header *ovs_header = info->userhdr;
+       struct datapath *dp;
+       int err;
+       u32 meter_id;
+       bool failed;
+
+       meter = dp_meter_create(a);
+       if (IS_ERR_OR_NULL(meter))
+               return PTR_ERR(meter);
+
+       reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_SET,
+                                         &ovs_reply_header);
+       if (IS_ERR(reply)) {
+               err = PTR_ERR(reply);
+               goto exit_free_meter;
+       }
+
+       ovs_lock();
+       dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
+       if (!dp) {
+               err = -ENODEV;
+               goto exit_unlock;
+       }
+
+       if (!a[OVS_METER_ATTR_ID]) {
+               err = -ENODEV;
+               goto exit_unlock;
+       }
+
+       meter_id = nla_get_u32(a[OVS_METER_ATTR_ID]);
+
+       /* Cannot fail after this. */
+       old_meter = lookup_meter(dp, meter_id);
+       detach_meter(old_meter);
+       attach_meter(dp, meter);
+       ovs_unlock();
+
+       /* Build response with the meter_id and stats from
+        * the old meter, if any.
+        */
+       failed = nla_put_u32(reply, OVS_METER_ATTR_ID, meter_id);
+       WARN_ON(failed);
+       if (old_meter) {
+               spin_lock_bh(&old_meter->lock);
+               if (old_meter->keep_stats) {
+                       err = ovs_meter_cmd_reply_stats(reply, meter_id,
+                                                       old_meter);
+                       WARN_ON(err);
+               }
+               spin_unlock_bh(&old_meter->lock);
+               ovs_meter_free(old_meter);
+       }
+
+       genlmsg_end(reply, ovs_reply_header);
+       return genlmsg_reply(reply, info);
+
+exit_unlock:
+       ovs_unlock();
+       nlmsg_free(reply);
+exit_free_meter:
+       kfree(meter);
+       return err;
+}
+
+static int ovs_meter_cmd_get(struct sk_buff *skb, struct genl_info *info)
+{
+       struct nlattr **a = info->attrs;
+       u32 meter_id;
+       struct ovs_header *ovs_header = info->userhdr;
+       struct ovs_header *ovs_reply_header;
+       struct datapath *dp;
+       int err;
+       struct sk_buff *reply;
+       struct dp_meter *meter;
+
+       if (!a[OVS_METER_ATTR_ID])
+               return -EINVAL;
+
+       meter_id = nla_get_u32(a[OVS_METER_ATTR_ID]);
+
+       reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_GET,
+                                         &ovs_reply_header);
+       if (IS_ERR(reply))
+               return PTR_ERR(reply);
+
+       ovs_lock();
+
+       dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
+       if (!dp) {
+               err = -ENODEV;
+               goto exit_unlock;
+       }
+
+       /* Locate meter, copy stats. */
+       meter = lookup_meter(dp, meter_id);
+       if (!meter) {
+               err = -ENOENT;
+               goto exit_unlock;
+       }
+
+       spin_lock_bh(&meter->lock);
+       err = ovs_meter_cmd_reply_stats(reply, meter_id, meter);
+       spin_unlock_bh(&meter->lock);
+       if (err)
+               goto exit_unlock;
+
+       ovs_unlock();
+
+       genlmsg_end(reply, ovs_reply_header);
+       return genlmsg_reply(reply, info);
+
+exit_unlock:
+       ovs_unlock();
+       nlmsg_free(reply);
+       return err;
+}
+
+static int ovs_meter_cmd_del(struct sk_buff *skb, struct genl_info *info)
+{
+       struct nlattr **a = info->attrs;
+       u32 meter_id;
+       struct ovs_header *ovs_header = info->userhdr;
+       struct ovs_header *ovs_reply_header;
+       struct datapath *dp;
+       int err;
+       struct sk_buff *reply;
+       struct dp_meter *old_meter;
+
+       if (!a[OVS_METER_ATTR_ID])
+               return -EINVAL;
+       meter_id = nla_get_u32(a[OVS_METER_ATTR_ID]);
+
+       reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_DEL,
+                                         &ovs_reply_header);
+       if (IS_ERR(reply))
+               return PTR_ERR(reply);
+
+       ovs_lock();
+
+       dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
+       if (!dp) {
+               err = -ENODEV;
+               goto exit_unlock;
+       }
+
+       old_meter = lookup_meter(dp, meter_id);
+       if (old_meter) {
+               spin_lock_bh(&old_meter->lock);
+               err = ovs_meter_cmd_reply_stats(reply, meter_id, old_meter);
+               WARN_ON(err);
+               spin_unlock_bh(&old_meter->lock);
+               detach_meter(old_meter);
+       }
+       ovs_unlock();
+       ovs_meter_free(old_meter);
+       genlmsg_end(reply, ovs_reply_header);
+       return genlmsg_reply(reply, info);
+
+exit_unlock:
+       ovs_unlock();
+       nlmsg_free(reply);
+       return err;
+}
+
+/* Meter action execution.
+ *
+ * Return true 'meter_id' drop band is triggered. The 'skb' should be
+ * dropped by the caller'.
+ */
+bool ovs_meter_execute(struct datapath *dp, struct sk_buff *skb,
+                      struct sw_flow_key *key, u32 meter_id)
+{
+       struct dp_meter *meter;
+       struct dp_meter_band *band;
+       long long int now_ms = div_u64(ktime_get_ns(), 1000 * 1000);
+       long long int long_delta_ms;
+       u32 delta_ms;
+       u32 cost;
+       int i, band_exceeded_max = -1;
+       u32 band_exceeded_rate = 0;
+
+       meter = lookup_meter(dp, meter_id);
+       /* Do not drop the packet when there is no meter. */
+       if (!meter)
+               return false;
+
+       /* Lock the meter while using it. */
+       spin_lock(&meter->lock);
+
+       long_delta_ms = (now_ms - meter->used); /* ms */
+
+       /* Make sure delta_ms will not be too large, so that bucket will not
+        * wrap around below.
+        */
+       delta_ms = (long_delta_ms > (long long int)meter->max_delta_t)
+                  ? meter->max_delta_t : (u32)long_delta_ms;
+
+       /* Update meter statistics.
+        */
+       meter->used = now_ms;
+       meter->stats.n_packets += 1;
+       meter->stats.n_bytes += skb->len;
+
+       /* Bucket rate is either in kilobits per second, or in packets per
+        * second.  We maintain the bucket in the units of either bits or
+        * 1/1000th of a packet, correspondingly.
+        * Then, when rate is multiplied with milliseconds, we get the
+        * bucket units:
+        * msec * kbps = bits, and
+        * msec * packets/sec = 1/1000 packets.
+        *
+        * 'cost' is the number of bucket units in this packet.
+        */
+       cost = (meter->kbps) ? skb->len * 8 : 1000;
+
+       /* Update all bands and find the one hit with the highest rate. */
+       for (i = 0; i < meter->n_bands; ++i) {
+               long long int max_bucket_size;
+
+               band = &meter->bands[i];
+               max_bucket_size = (band->burst_size + band->rate) * 1000;
+
+               band->bucket += delta_ms * band->rate;
+               if (band->bucket > max_bucket_size)
+                       band->bucket = max_bucket_size;
+
+               if (band->bucket >= cost) {
+                       band->bucket -= cost;
+               } else if (band->rate > band_exceeded_rate) {
+                       band_exceeded_rate = band->rate;
+                       band_exceeded_max = i;
+               }
+       }
+
+       if (band_exceeded_max >= 0) {
+               /* Update band statistics. */
+               band = &meter->bands[band_exceeded_max];
+               band->stats.n_packets += 1;
+               band->stats.n_bytes += skb->len;
+
+               /* Drop band triggered, let the caller drop the 'skb'.  */
+               if (band->type == OVS_METER_BAND_TYPE_DROP) {
+                       spin_unlock(&meter->lock);
+                       return true;
+               }
+       }
+
+       spin_unlock(&meter->lock);
+       return false;
+}
+
+static struct genl_ops dp_meter_genl_ops[] = {
+       { .cmd = OVS_METER_CMD_FEATURES,
+               .flags = 0,               /* OK for unprivileged users. */
+               .policy = meter_policy,
+               .doit = ovs_meter_cmd_features
+       },
+       { .cmd = OVS_METER_CMD_SET,
+               .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN
+                                          *  privilege.
+                                          */
+               .policy = meter_policy,
+               .doit = ovs_meter_cmd_set,
+       },
+       { .cmd = OVS_METER_CMD_GET,
+               .flags = 0,               /* OK for unprivileged users. */
+               .policy = meter_policy,
+               .doit = ovs_meter_cmd_get,
+       },
+       { .cmd = OVS_METER_CMD_DEL,
+               .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN
+                                          *  privilege.
+                                          */
+               .policy = meter_policy,
+               .doit = ovs_meter_cmd_del
+       },
+};
+
+static const struct genl_multicast_group ovs_meter_multicast_group = {
+       .name = OVS_METER_MCGROUP,
+};
+
+struct genl_family dp_meter_genl_family __ro_after_init = {
+       .hdrsize = sizeof(struct ovs_header),
+       .name = OVS_METER_FAMILY,
+       .version = OVS_METER_VERSION,
+       .maxattr = OVS_METER_ATTR_MAX,
+       .netnsok = true,
+       .parallel_ops = true,
+       .ops = dp_meter_genl_ops,
+       .n_ops = ARRAY_SIZE(dp_meter_genl_ops),
+       .mcgrps = &ovs_meter_multicast_group,
+       .n_mcgrps = 1,
+       .module = THIS_MODULE,
+};
+
+int ovs_meters_init(struct datapath *dp)
+{
+       int i;
+
+       dp->meters = kmalloc_array(METER_HASH_BUCKETS,
+                                  sizeof(struct hlist_head), GFP_KERNEL);
+
+       if (!dp->meters)
+               return -ENOMEM;
+
+       for (i = 0; i < METER_HASH_BUCKETS; i++)
+               INIT_HLIST_HEAD(&dp->meters[i]);
+
+       return 0;
+}
+
+void ovs_meters_exit(struct datapath *dp)
+{
+       int i;
+
+       for (i = 0; i < METER_HASH_BUCKETS; i++) {
+               struct hlist_head *head = &dp->meters[i];
+               struct dp_meter *meter;
+               struct hlist_node *n;
+
+               hlist_for_each_entry_safe(meter, n, head, dp_hash_node)
+                       kfree(meter);
+       }
+
+       kfree(dp->meters);
+}
diff --git a/net/openvswitch/meter.h b/net/openvswitch/meter.h
new file mode 100644 (file)
index 0000000..964ace2
--- /dev/null
@@ -0,0 +1,54 @@
+/*
+ * Copyright (c) 2017 Nicira, Inc.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ */
+
+#ifndef METER_H
+#define METER_H 1
+
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/kernel.h>
+#include <linux/netlink.h>
+#include <linux/openvswitch.h>
+#include <linux/genetlink.h>
+#include <linux/skbuff.h>
+
+#include "flow.h"
+struct datapath;
+
+#define DP_MAX_BANDS           1
+
+struct dp_meter_band {
+       u32 type;
+       u32 rate;
+       u32 burst_size;
+       u32 bucket; /* 1/1000 packets, or in bits */
+       struct ovs_flow_stats stats;
+};
+
+struct dp_meter {
+       spinlock_t lock;    /* Per meter lock */
+       struct rcu_head rcu;
+       struct hlist_node dp_hash_node; /*Element in datapath->meters
+                                        * hash table.
+                                        */
+       u32 id;
+       u16 kbps:1, keep_stats:1;
+       u16 n_bands;
+       u32 max_delta_t;
+       u64 used;
+       struct ovs_flow_stats stats;
+       struct dp_meter_band bands[];
+};
+
+extern struct genl_family dp_meter_genl_family;
+int ovs_meters_init(struct datapath *dp);
+void ovs_meters_exit(struct datapath *dp);
+bool ovs_meter_execute(struct datapath *dp, struct sk_buff *skb,
+                      struct sw_flow_key *key, u32 meter_id);
+
+#endif /* meter.h */