[ovs-dev,branch-2.7,02/25] datapath: use percpu flow stats

Submitted by Jarno Rajahalme on March 15, 2017, 11:31 p.m.

Details

Message ID 1489620689-122370-3-git-send-email-jarno@ovn.org
State Superseded
Headers show

Commit Message

Jarno Rajahalme March 15, 2017, 11:31 p.m.
From: Thadeu Lima de Souza Cascardo <cascardo@redhat.com>

Upstream commit:
    commit db74a3335e0f645e3139c80bcfc90feb01d8e304
    Author: Thadeu Lima de Souza Cascardo <cascardo@redhat.com>
    Date:   Thu Sep 15 19:11:53 2016 -0300

    openvswitch: use percpu flow stats

    Instead of using flow stats per NUMA node, use it per CPU. When using
    megaflows, the stats lock can be a bottleneck in scalability.

    On a E5-2690 12-core system, usual throughput went from ~4Mpps to
    ~15Mpps when forwarding between two 40GbE ports with a single flow
    configured on the datapath.

    This has been tested on a system with possible CPUs 0-7,16-23. After
    module removal, there were no corruption on the slab cache.

    Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@redhat.com>
    Cc: pravin shelar <pshelar@ovn.org>
    Acked-by: Pravin B Shelar <pshelar@ovn.org>
    Signed-off-by: David S. Miller <davem@davemloft.net>

Signed-off-by: Jarno Rajahalme <jarno@ovn.org>
Signed-off-by: Joe Stringer <joe@ovn.org>
---
 datapath/flow.c       | 42 ++++++++++++++++++++++--------------------
 datapath/flow.h       |  4 ++--
 datapath/flow_table.c | 26 +++++++++-----------------
 3 files changed, 33 insertions(+), 39 deletions(-)

Patch hide | download patch | download mbox

diff --git a/datapath/flow.c b/datapath/flow.c
index 6d56644..58b0e13 100644
--- a/datapath/flow.c
+++ b/datapath/flow.c
@@ -29,6 +29,7 @@ 
 #include <linux/module.h>
 #include <linux/in.h>
 #include <linux/rcupdate.h>
+#include <linux/cpumask.h>
 #include <linux/if_arp.h>
 #include <linux/ip.h>
 #include <linux/ipv6.h>
@@ -71,32 +72,33 @@  void ovs_flow_stats_update(struct sw_flow *flow, __be16 tcp_flags,
 {
 	struct flow_stats *stats;
 	int node = numa_node_id();
+	int cpu = smp_processor_id();
 	int len = skb->len + (skb_vlan_tag_present(skb) ? VLAN_HLEN : 0);
 
-	stats = rcu_dereference(flow->stats[node]);
+	stats = rcu_dereference(flow->stats[cpu]);
 
-	/* Check if already have node-specific stats. */
+	/* Check if already have CPU-specific stats. */
 	if (likely(stats)) {
 		spin_lock(&stats->lock);
 		/* Mark if we write on the pre-allocated stats. */
-		if (node == 0 && unlikely(flow->stats_last_writer != node))
-			flow->stats_last_writer = node;
+		if (cpu == 0 && unlikely(flow->stats_last_writer != cpu))
+			flow->stats_last_writer = cpu;
 	} else {
 		stats = rcu_dereference(flow->stats[0]); /* Pre-allocated. */
 		spin_lock(&stats->lock);
 
-		/* If the current NUMA-node is the only writer on the
+		/* If the current CPU is the only writer on the
 		 * pre-allocated stats keep using them.
 		 */
-		if (unlikely(flow->stats_last_writer != node)) {
+		if (unlikely(flow->stats_last_writer != cpu)) {
 			/* A previous locker may have already allocated the
-			 * stats, so we need to check again.  If node-specific
+			 * stats, so we need to check again.  If CPU-specific
 			 * stats were already allocated, we update the pre-
 			 * allocated stats as we have already locked them.
 			 */
-			if (likely(flow->stats_last_writer != NUMA_NO_NODE)
-			    && likely(!rcu_access_pointer(flow->stats[node]))) {
-				/* Try to allocate node-specific stats. */
+			if (likely(flow->stats_last_writer != -1) &&
+			    likely(!rcu_access_pointer(flow->stats[cpu]))) {
+				/* Try to allocate CPU-specific stats. */
 				struct flow_stats *new_stats;
 
 				new_stats =
@@ -113,12 +115,12 @@  void ovs_flow_stats_update(struct sw_flow *flow, __be16 tcp_flags,
 					new_stats->tcp_flags = tcp_flags;
 					spin_lock_init(&new_stats->lock);
 
-					rcu_assign_pointer(flow->stats[node],
+					rcu_assign_pointer(flow->stats[cpu],
 							   new_stats);
 					goto unlock;
 				}
 			}
-			flow->stats_last_writer = node;
+			flow->stats_last_writer = cpu;
 		}
 	}
 
@@ -135,15 +137,15 @@  void ovs_flow_stats_get(const struct sw_flow *flow,
 			struct ovs_flow_stats *ovs_stats,
 			unsigned long *used, __be16 *tcp_flags)
 {
-	int node;
+	int cpu;
 
 	*used = 0;
 	*tcp_flags = 0;
 	memset(ovs_stats, 0, sizeof(*ovs_stats));
 
-	/* We open code this to make sure node 0 is always considered */
-	for (node = 0; node < MAX_NUMNODES; node = next_node(node, node_possible_map)) {
-		struct flow_stats *stats = rcu_dereference_ovsl(flow->stats[node]);
+	/* We open code this to make sure cpu 0 is always considered */
+	for (cpu = 0; cpu < nr_cpu_ids; cpu = cpumask_next(cpu, cpu_possible_mask)) {
+		struct flow_stats *stats = rcu_dereference_ovsl(flow->stats[cpu]);
 
 		if (stats) {
 			/* Local CPU may write on non-local stats, so we must
@@ -163,11 +165,11 @@  void ovs_flow_stats_get(const struct sw_flow *flow,
 /* Called with ovs_mutex. */
 void ovs_flow_stats_clear(struct sw_flow *flow)
 {
-	int node;
+	int cpu;
 
-	/* We open code this to make sure node 0 is always considered */
-	for (node = 0; node < MAX_NUMNODES; node = next_node(node, node_possible_map)) {
-		struct flow_stats *stats = ovsl_dereference(flow->stats[node]);
+	/* We open code this to make sure cpu 0 is always considered */
+	for (cpu = 0; cpu < nr_cpu_ids; cpu = cpumask_next(cpu, cpu_possible_mask)) {
+		struct flow_stats *stats = ovsl_dereference(flow->stats[cpu]);
 
 		if (stats) {
 			spin_lock_bh(&stats->lock);
diff --git a/datapath/flow.h b/datapath/flow.h
index 2dd0696..3203527 100644
--- a/datapath/flow.h
+++ b/datapath/flow.h
@@ -171,14 +171,14 @@  struct sw_flow {
 		struct hlist_node node[2];
 		u32 hash;
 	} flow_table, ufid_table;
-	int stats_last_writer;		/* NUMA-node id of the last writer on
+	int stats_last_writer;		/* CPU id of the last writer on
 					 * 'stats[0]'.
 					 */
 	struct sw_flow_key key;
 	struct sw_flow_id id;
 	struct sw_flow_mask *mask;
 	struct sw_flow_actions __rcu *sf_acts;
-	struct flow_stats __rcu *stats[]; /* One for each NUMA node.  First one
+	struct flow_stats __rcu *stats[]; /* One for each CPU.  First one
 					   * is allocated at flow creation time,
 					   * the rest are allocated on demand
 					   * while holding the 'stats[0].lock'.
diff --git a/datapath/flow_table.c b/datapath/flow_table.c
index 3829b92..6fe3739 100644
--- a/datapath/flow_table.c
+++ b/datapath/flow_table.c
@@ -31,6 +31,7 @@ 
 #include <linux/module.h>
 #include <linux/in.h>
 #include <linux/rcupdate.h>
+#include <linux/cpumask.h>
 #include <linux/if_arp.h>
 #include <linux/ip.h>
 #include <linux/ipv6.h>
@@ -85,17 +86,12 @@  struct sw_flow *ovs_flow_alloc(void)
 {
 	struct sw_flow *flow;
 	struct flow_stats *stats;
-	int node;
 
-	flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
+	flow = kmem_cache_zalloc(flow_cache, GFP_KERNEL);
 	if (!flow)
 		return ERR_PTR(-ENOMEM);
 
-	flow->sf_acts = NULL;
-	flow->mask = NULL;
-	flow->id.ufid_len = 0;
-	flow->id.unmasked_key = NULL;
-	flow->stats_last_writer = NUMA_NO_NODE;
+	flow->stats_last_writer = -1;
 
 	/* Initialize the default stat node. */
 	stats = kmem_cache_alloc_node(flow_stats_cache,
@@ -108,10 +104,6 @@  struct sw_flow *ovs_flow_alloc(void)
 
 	RCU_INIT_POINTER(flow->stats[0], stats);
 
-	for_each_node(node)
-		if (node != 0)
-			RCU_INIT_POINTER(flow->stats[node], NULL);
-
 	return flow;
 err:
 	kmem_cache_free(flow_cache, flow);
@@ -148,17 +140,17 @@  static struct flex_array *alloc_buckets(unsigned int n_buckets)
 
 static void flow_free(struct sw_flow *flow)
 {
-	int node;
+	int cpu;
 
 	if (ovs_identifier_is_key(&flow->id))
 		kfree(flow->id.unmasked_key);
 	if (flow->sf_acts)
 		ovs_nla_free_flow_actions((struct sw_flow_actions __force *)flow->sf_acts);
-	/* We open code this to make sure node 0 is always considered */
-	for (node = 0; node < MAX_NUMNODES; node = next_node(node, node_possible_map))
-		if (flow->stats[node])
+	/* We open code this to make sure cpu 0 is always considered */
+	for (cpu = 0; cpu < nr_cpu_ids; cpu = cpumask_next(cpu, cpu_possible_mask))
+		if (flow->stats[cpu])
 			kmem_cache_free(flow_stats_cache,
-					rcu_dereference_raw(flow->stats[node]));
+					rcu_dereference_raw(flow->stats[cpu]));
 	kmem_cache_free(flow_cache, flow);
 }
 
@@ -985,7 +977,7 @@  int ovs_flow_init(void)
 	BUILD_BUG_ON(sizeof(struct sw_flow_key) % sizeof(long));
 
 	flow_cache = kmem_cache_create("sw_flow", sizeof(struct sw_flow)
-				       + (nr_node_ids
+				       + (nr_cpu_ids
 					  * sizeof(struct flow_stats *)),
 				       0, 0, NULL);
 	if (flow_cache == NULL)