openvswitch: Introduce per-cpu upcall dispatch

The Open vSwitch kernel module uses the upcall mechanism to send
packets from kernel space to user space when it misses in the kernel
space flow table. The upcall sends packets via a Netlink socket.
Currently, a Netlink socket is created for every vport. In this way,
there is a 1:1 mapping between a vport and a Netlink socket.
When a packet is received by a vport, if it needs to be sent to
user space, it is sent via the corresponding Netlink socket.

This mechanism, with various iterations of the corresponding user
space code, has seen some limitations and issues:

* On systems with a large number of vports, there is a correspondingly
large number of Netlink sockets which can limit scaling.
(https://bugzilla.redhat.com/show_bug.cgi?id=1526306)
* Packet reordering on upcalls.
(https://bugzilla.redhat.com/show_bug.cgi?id=1844576)
* A thundering herd issue.
(https://bugzilla.redhat.com/show_bug.cgi?id=1834444)

This patch introduces an alternative, feature-negotiated, upcall
mode using a per-cpu dispatch rather than a per-vport dispatch.

In this mode, the Netlink socket to be used for the upcall is
selected based on the CPU of the thread that is executing the upcall.
In this way, it resolves the issues above as:

a) The number of Netlink sockets scales with the number of CPUs
rather than the number of vports.
b) Ordering per-flow is maintained as packets are distributed to
CPUs based on mechanisms such as RSS and flows are distributed
to a single user space thread.
c) Packets from a flow can only wake up one user space thread.

The corresponding user space code can be found at:
https://mail.openvswitch.org/pipermail/ovs-dev/2021-July/385139.html

Bugzilla: https://bugzilla.redhat.com/1844576
Signed-off-by: Mark Gray <mark.d.gray@redhat.com>
Acked-by: Flavio Leitner <fbl@sysclose.org>
Acked-by: Pravin B Shelar <pshelar@ovn.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

3 files changed, 95 insertions, 3 deletions

diff --git a/net/openvswitch/actions.c b/net/openvswitch/actions.c
index ef15d9eb4774..f79679746c62 100644
--- a/net/openvswitch/actions.c
+++ b/net/openvswitch/actions.c
@@ -924,7 +924,11 @@ static int output_userspace(struct datapath *dp, struct sk_buff *skb,
 			break;
 
 		case OVS_USERSPACE_ATTR_PID:
-			upcall.portid = nla_get_u32(a);
+			if (dp->user_features & OVS_DP_F_DISPATCH_UPCALL_PER_CPU)
+				upcall.portid =
+				   ovs_dp_get_upcall_portid(dp, smp_processor_id());
+			else
+				upcall.portid = nla_get_u32(a);
 			break;
 
 		case OVS_USERSPACE_ATTR_EGRESS_TUN_PORT: {
diff --git a/net/openvswitch/datapath.c b/net/openvswitch/datapath.c
index bc164b35e67d..7a4edafdc685 100644
--- a/net/openvswitch/datapath.c
+++ b/net/openvswitch/datapath.c
@@ -133,6 +133,8 @@ static int queue_userspace_packet(struct datapath *dp, struct sk_buff *,
 
 static void ovs_dp_masks_rebalance(struct work_struct *work);
 
+static int ovs_dp_set_upcall_portids(struct datapath *, const struct nlattr *);
+
 /* Must be called with rcu_read_lock or ovs_mutex. */
 const char *ovs_dp_name(const struct datapath *dp)
 {
@@ -166,6 +168,7 @@ static void destroy_dp_rcu(struct rcu_head *rcu)
 	free_percpu(dp->stats_percpu);
 	kfree(dp->ports);
 	ovs_meters_exit(dp);
+	kfree(dp->upcall_portids);
 	kfree(dp);
 }
 
@@ -239,7 +242,12 @@ void ovs_dp_process_packet(struct sk_buff *skb, struct sw_flow_key *key)
 
 		memset(&upcall, 0, sizeof(upcall));
 		upcall.cmd = OVS_PACKET_CMD_MISS;
-		upcall.portid = ovs_vport_find_upcall_portid(p, skb);
+
+		if (dp->user_features & OVS_DP_F_DISPATCH_UPCALL_PER_CPU)
+			upcall.portid = ovs_dp_get_upcall_portid(dp, smp_processor_id());
+		else
+			upcall.portid = ovs_vport_find_upcall_portid(p, skb);
+
 		upcall.mru = OVS_CB(skb)->mru;
 		error = ovs_dp_upcall(dp, skb, key, &upcall, 0);
 		if (unlikely(error))
@@ -1594,16 +1602,67 @@ static void ovs_dp_reset_user_features(struct sk_buff *skb,
 
 DEFINE_STATIC_KEY_FALSE(tc_recirc_sharing_support);
 
+static int ovs_dp_set_upcall_portids(struct datapath *dp,
+			      const struct nlattr *ids)
+{
+	struct dp_nlsk_pids *old, *dp_nlsk_pids;
+
+	if (!nla_len(ids) || nla_len(ids) % sizeof(u32))
+		return -EINVAL;
+
+	old = ovsl_dereference(dp->upcall_portids);
+
+	dp_nlsk_pids = kmalloc(sizeof(*dp_nlsk_pids) + nla_len(ids),
+			       GFP_KERNEL);
+	if (!dp_nlsk_pids)
+		return -ENOMEM;
+
+	dp_nlsk_pids->n_pids = nla_len(ids) / sizeof(u32);
+	nla_memcpy(dp_nlsk_pids->pids, ids, nla_len(ids));
+
+	rcu_assign_pointer(dp->upcall_portids, dp_nlsk_pids);
+
+	kfree_rcu(old, rcu);
+
+	return 0;
+}
+
+u32 ovs_dp_get_upcall_portid(const struct datapath *dp, uint32_t cpu_id)
+{
+	struct dp_nlsk_pids *dp_nlsk_pids;
+
+	dp_nlsk_pids = rcu_dereference(dp->upcall_portids);
+
+	if (dp_nlsk_pids) {
+		if (cpu_id < dp_nlsk_pids->n_pids) {
+			return dp_nlsk_pids->pids[cpu_id];
+		} else if (dp_nlsk_pids->n_pids > 0 && cpu_id >= dp_nlsk_pids->n_pids) {
+			/* If the number of netlink PIDs is mismatched with the number of
+			 * CPUs as seen by the kernel, log this and send the upcall to an
+			 * arbitrary socket (0) in order to not drop packets
+			 */
+			pr_info_ratelimited("cpu_id mismatch with handler threads");
+			return dp_nlsk_pids->pids[cpu_id % dp_nlsk_pids->n_pids];
+		} else {
+			return 0;
+		}
+	} else {
+		return 0;
+	}
+}
+
 static int ovs_dp_change(struct datapath *dp, struct nlattr *a[])
 {
 	u32 user_features = 0;
+	int err;
 
 	if (a[OVS_DP_ATTR_USER_FEATURES]) {
 		user_features = nla_get_u32(a[OVS_DP_ATTR_USER_FEATURES]);
 
 		if (user_features & ~(OVS_DP_F_VPORT_PIDS |
 				      OVS_DP_F_UNALIGNED |
-				      OVS_DP_F_TC_RECIRC_SHARING))
+				      OVS_DP_F_TC_RECIRC_SHARING |
+				      OVS_DP_F_DISPATCH_UPCALL_PER_CPU))
 			return -EOPNOTSUPP;
 
 #if !IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
@@ -1624,6 +1683,15 @@ static int ovs_dp_change(struct datapath *dp, struct nlattr *a[])
 
 	dp->user_features = user_features;
 
+	if (dp->user_features & OVS_DP_F_DISPATCH_UPCALL_PER_CPU &&
+	    a[OVS_DP_ATTR_PER_CPU_PIDS]) {
+		/* Upcall Netlink Port IDs have been updated */
+		err = ovs_dp_set_upcall_portids(dp,
+						a[OVS_DP_ATTR_PER_CPU_PIDS]);
+		if (err)
+			return err;
+	}
+
 	if (dp->user_features & OVS_DP_F_TC_RECIRC_SHARING)
 		static_branch_enable(&tc_recirc_sharing_support);
 	else
diff --git a/net/openvswitch/datapath.h b/net/openvswitch/datapath.h
index 38f7d3e66ca6..fcfe6cb46441 100644
--- a/net/openvswitch/datapath.h
+++ b/net/openvswitch/datapath.h
@@ -51,6 +51,21 @@ struct dp_stats_percpu {
 };
 
 /**
+ * struct dp_nlsk_pids - array of netlink portids of for a datapath.
+ *                       This is used when OVS_DP_F_DISPATCH_UPCALL_PER_CPU
+ *                       is enabled and must be protected by rcu.
+ * @rcu: RCU callback head for deferred destruction.
+ * @n_pids: Size of @pids array.
+ * @pids: Array storing the Netlink socket PIDs indexed by CPU ID for packets
+ *       that miss the flow table.
+ */
+struct dp_nlsk_pids {
+	struct rcu_head rcu;
+	u32 n_pids;
+	u32 pids[];
+};
+
+/**
  * struct datapath - datapath for flow-based packet switching
  * @rcu: RCU callback head for deferred destruction.
  * @list_node: Element in global 'dps' list.
@@ -61,6 +76,7 @@ struct dp_stats_percpu {
  * @net: Reference to net namespace.
  * @max_headroom: the maximum headroom of all vports in this datapath; it will
  * be used by all the internal vports in this dp.
+ * @upcall_portids: RCU protected 'struct dp_nlsk_pids'.
  *
  * Context: See the comment on locking at the top of datapath.c for additional
  * locking information.
@@ -87,6 +103,8 @@ struct datapath {
 
 	/* Switch meters. */
 	struct dp_meter_table meter_tbl;
+
+	struct dp_nlsk_pids __rcu *upcall_portids;
 };
 
 /**
@@ -243,6 +261,8 @@ int ovs_dp_upcall(struct datapath *, struct sk_buff *,
 		  const struct sw_flow_key *, const struct dp_upcall_info *,
 		  uint32_t cutlen);
 
+u32 ovs_dp_get_upcall_portid(const struct datapath *dp, uint32_t cpu_id);
+
 const char *ovs_dp_name(const struct datapath *dp);
 struct sk_buff *ovs_vport_cmd_build_info(struct vport *vport, struct net *net,
 					 u32 portid, u32 seq, u8 cmd);