From 63e7959c4b9bd6f791061c460a22d9ee32ae2240 Mon Sep 17 00:00:00 2001 From: Jarno Rajahalme Date: Thu, 27 Mar 2014 12:42:54 -0700 Subject: [PATCH] openvswitch: Per NUMA node flow stats. Keep kernel flow stats for each NUMA node rather than each (logical) CPU. This avoids using the per-CPU allocator and removes most of the kernel-side OVS locking overhead otherwise on the top of perf reports and allows OVS to scale better with higher number of threads. With 9 handlers and 4 revalidators netperf TCP_CRR test flow setup rate doubles on a server with two hyper-threaded physical CPUs (16 logical cores each) compared to the current OVS master. Tested with non-trivial flow table with a TCP port match rule forcing all new connections with unique port numbers to OVS userspace. The IP addresses are still wildcarded, so the kernel flows are not considered as exact match 5-tuple flows. This type of flows can be expected to appear in large numbers as the result of more effective wildcarding made possible by improvements in OVS userspace flow classifier. Perf results for this test (master): Events: 305K cycles + 8.43% ovs-vswitchd [kernel.kallsyms] [k] mutex_spin_on_owner + 5.64% ovs-vswitchd [kernel.kallsyms] [k] __ticket_spin_lock + 4.75% ovs-vswitchd ovs-vswitchd [.] find_match_wc + 3.32% ovs-vswitchd libpthread-2.15.so [.] pthread_mutex_lock + 2.61% ovs-vswitchd [kernel.kallsyms] [k] pcpu_alloc_area + 2.19% ovs-vswitchd ovs-vswitchd [.] flow_hash_in_minimask_range + 2.03% swapper [kernel.kallsyms] [k] intel_idle + 1.84% ovs-vswitchd libpthread-2.15.so [.] pthread_mutex_unlock + 1.64% ovs-vswitchd ovs-vswitchd [.] classifier_lookup + 1.58% ovs-vswitchd libc-2.15.so [.] 0x7f4e6 + 1.07% ovs-vswitchd [kernel.kallsyms] [k] memset + 1.03% netperf [kernel.kallsyms] [k] __ticket_spin_lock + 0.92% swapper [kernel.kallsyms] [k] __ticket_spin_lock ... And after this patch: Events: 356K cycles + 6.85% ovs-vswitchd ovs-vswitchd [.] find_match_wc + 4.63% ovs-vswitchd libpthread-2.15.so [.] pthread_mutex_lock + 3.06% ovs-vswitchd [kernel.kallsyms] [k] __ticket_spin_lock + 2.81% ovs-vswitchd ovs-vswitchd [.] flow_hash_in_minimask_range + 2.51% ovs-vswitchd libpthread-2.15.so [.] pthread_mutex_unlock + 2.27% ovs-vswitchd ovs-vswitchd [.] classifier_lookup + 1.84% ovs-vswitchd libc-2.15.so [.] 0x15d30f + 1.74% ovs-vswitchd [kernel.kallsyms] [k] mutex_spin_on_owner + 1.47% swapper [kernel.kallsyms] [k] intel_idle + 1.34% ovs-vswitchd ovs-vswitchd [.] flow_hash_in_minimask + 1.33% ovs-vswitchd ovs-vswitchd [.] rule_actions_unref + 1.16% ovs-vswitchd ovs-vswitchd [.] hindex_node_with_hash + 1.16% ovs-vswitchd ovs-vswitchd [.] do_xlate_actions + 1.09% ovs-vswitchd ovs-vswitchd [.] ofproto_rule_ref + 1.01% netperf [kernel.kallsyms] [k] __ticket_spin_lock ... There is a small increase in kernel spinlock overhead due to the same spinlock being shared between multiple cores of the same physical CPU, but that is barely visible in the netperf TCP_CRR test performance (maybe ~1% performance drop, hard to tell exactly due to variance in the test results), when testing for kernel module throughput (with no userspace activity, handful of kernel flows). On flow setup, a single stats instance is allocated (for the NUMA node 0). As CPUs from multiple NUMA nodes start updating stats, new NUMA-node specific stats instances are allocated. This allocation on the packet processing code path is made to never block or look for emergency memory pools, minimizing the allocation latency. If the allocation fails, the existing preallocated stats instance is used. Also, if only CPUs from one NUMA-node are updating the preallocated stats instance, no additional stats instances are allocated. This eliminates the need to pre-allocate stats instances that will not be used, also relieving the stats reader from the burden of reading stats that are never used. Signed-off-by: Jarno Rajahalme Acked-by: Pravin B Shelar Signed-off-by: Jesse Gross --- net/openvswitch/flow.c | 119 ++++++++++++++++++++++------------- net/openvswitch/flow.h | 10 ++- net/openvswitch/flow_table.c | 46 ++++++++++---- net/openvswitch/flow_table.h | 2 + 4 files changed, 122 insertions(+), 55 deletions(-) diff --git a/net/openvswitch/flow.c b/net/openvswitch/flow.c index aad7a8da70b1..432f04d5c896 100644 --- a/net/openvswitch/flow.c +++ b/net/openvswitch/flow.c @@ -65,8 +65,9 @@ void ovs_flow_stats_update(struct sw_flow *flow, struct sk_buff *skb) { struct flow_stats *stats; __be16 tcp_flags = 0; + int node = numa_node_id(); - stats = this_cpu_ptr(flow->stats); + stats = rcu_dereference(flow->stats[node]); if ((flow->key.eth.type == htons(ETH_P_IP) || flow->key.eth.type == htons(ETH_P_IPV6)) && @@ -76,68 +77,102 @@ void ovs_flow_stats_update(struct sw_flow *flow, struct sk_buff *skb) tcp_flags = TCP_FLAGS_BE16(tcp_hdr(skb)); } - spin_lock(&stats->lock); + /* Check if already have node-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; + } else { + stats = rcu_dereference(flow->stats[0]); /* Pre-allocated. */ + spin_lock(&stats->lock); + + /* If the current NUMA-node is the only writer on the + * pre-allocated stats keep using them. + */ + if (unlikely(flow->stats_last_writer != node)) { + /* A previous locker may have already allocated the + * stats, so we need to check again. If node-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_dereference(flow->stats[node]))) { + /* Try to allocate node-specific stats. */ + struct flow_stats *new_stats; + + new_stats = + kmem_cache_alloc_node(flow_stats_cache, + GFP_THISNODE | + __GFP_NOMEMALLOC, + node); + if (likely(new_stats)) { + new_stats->used = jiffies; + new_stats->packet_count = 1; + new_stats->byte_count = skb->len; + new_stats->tcp_flags = tcp_flags; + spin_lock_init(&new_stats->lock); + + rcu_assign_pointer(flow->stats[node], + new_stats); + goto unlock; + } + } + flow->stats_last_writer = node; + } + } + stats->used = jiffies; stats->packet_count++; stats->byte_count += skb->len; stats->tcp_flags |= tcp_flags; - spin_unlock(&stats->lock); -} - -static void stats_read(struct flow_stats *stats, - struct ovs_flow_stats *ovs_stats, - unsigned long *used, __be16 *tcp_flags) -{ - spin_lock(&stats->lock); - if (!*used || time_after(stats->used, *used)) - *used = stats->used; - *tcp_flags |= stats->tcp_flags; - ovs_stats->n_packets += stats->packet_count; - ovs_stats->n_bytes += stats->byte_count; +unlock: spin_unlock(&stats->lock); } void ovs_flow_stats_get(struct sw_flow *flow, struct ovs_flow_stats *ovs_stats, unsigned long *used, __be16 *tcp_flags) { - int cpu; + int node; *used = 0; *tcp_flags = 0; memset(ovs_stats, 0, sizeof(*ovs_stats)); - local_bh_disable(); - - for_each_possible_cpu(cpu) { - struct flow_stats *stats; + for_each_node(node) { + struct flow_stats *stats = rcu_dereference(flow->stats[node]); - stats = per_cpu_ptr(flow->stats.cpu_stats, cpu); - stats_read(stats, ovs_stats, used, tcp_flags); + if (stats) { + /* Local CPU may write on non-local stats, so we must + * block bottom-halves here. + */ + spin_lock_bh(&stats->lock); + if (!*used || time_after(stats->used, *used)) + *used = stats->used; + *tcp_flags |= stats->tcp_flags; + ovs_stats->n_packets += stats->packet_count; + ovs_stats->n_bytes += stats->byte_count; + spin_unlock_bh(&stats->lock); + } } - - local_bh_enable(); -} - -static void stats_reset(struct flow_stats *stats) -{ - spin_lock(&stats->lock); - stats->used = 0; - stats->packet_count = 0; - stats->byte_count = 0; - stats->tcp_flags = 0; - spin_unlock(&stats->lock); } void ovs_flow_stats_clear(struct sw_flow *flow) { - int cpu; - - local_bh_disable(); - - for_each_possible_cpu(cpu) - stats_reset(per_cpu_ptr(flow->stats, cpu)); - - local_bh_enable(); + int node; + + for_each_node(node) { + struct flow_stats *stats = rcu_dereference(flow->stats[node]); + + if (stats) { + spin_lock_bh(&stats->lock); + stats->used = 0; + stats->packet_count = 0; + stats->byte_count = 0; + stats->tcp_flags = 0; + spin_unlock_bh(&stats->lock); + } + } } static int check_header(struct sk_buff *skb, int len) diff --git a/net/openvswitch/flow.h b/net/openvswitch/flow.h index 9c0dd8aa3117..ddcebc53224f 100644 --- a/net/openvswitch/flow.h +++ b/net/openvswitch/flow.h @@ -159,12 +159,18 @@ struct sw_flow { struct rcu_head rcu; struct hlist_node hash_node[2]; u32 hash; - + int stats_last_writer; /* NUMA-node id of the last writer on + * 'stats[0]'. + */ struct sw_flow_key key; struct sw_flow_key unmasked_key; struct sw_flow_mask *mask; struct sw_flow_actions __rcu *sf_acts; - struct flow_stats __percpu *stats; + struct flow_stats __rcu *stats[]; /* One for each NUMA node. First one + * is allocated at flow creation time, + * the rest are allocated on demand + * while holding the 'stats[0].lock'. + */ }; struct arp_eth_header { diff --git a/net/openvswitch/flow_table.c b/net/openvswitch/flow_table.c index aa92da23053d..d8ef37b937bd 100644 --- a/net/openvswitch/flow_table.c +++ b/net/openvswitch/flow_table.c @@ -48,6 +48,7 @@ #define REHASH_INTERVAL (10 * 60 * HZ) static struct kmem_cache *flow_cache; +struct kmem_cache *flow_stats_cache __read_mostly; static u16 range_n_bytes(const struct sw_flow_key_range *range) { @@ -75,7 +76,8 @@ void ovs_flow_mask_key(struct sw_flow_key *dst, const struct sw_flow_key *src, struct sw_flow *ovs_flow_alloc(void) { struct sw_flow *flow; - int cpu; + struct flow_stats *stats; + int node; flow = kmem_cache_alloc(flow_cache, GFP_KERNEL); if (!flow) @@ -83,17 +85,22 @@ struct sw_flow *ovs_flow_alloc(void) flow->sf_acts = NULL; flow->mask = NULL; + flow->stats_last_writer = NUMA_NO_NODE; - flow->stats = alloc_percpu(struct flow_stats); - if (!flow->stats) + /* Initialize the default stat node. */ + stats = kmem_cache_alloc_node(flow_stats_cache, + GFP_KERNEL | __GFP_ZERO, 0); + if (!stats) goto err; - for_each_possible_cpu(cpu) { - struct flow_stats *cpu_stats; + spin_lock_init(&stats->lock); + + RCU_INIT_POINTER(flow->stats[0], stats); + + for_each_node(node) + if (node != 0) + RCU_INIT_POINTER(flow->stats[node], NULL); - cpu_stats = per_cpu_ptr(flow->stats, cpu); - spin_lock_init(&cpu_stats->lock); - } return flow; err: kmem_cache_free(flow_cache, flow); @@ -130,8 +137,13 @@ static struct flex_array *alloc_buckets(unsigned int n_buckets) static void flow_free(struct sw_flow *flow) { + int node; + kfree((struct sf_flow_acts __force *)flow->sf_acts); - free_percpu(flow->stats); + for_each_node(node) + if (flow->stats[node]) + kmem_cache_free(flow_stats_cache, + (struct flow_stats __force *)flow->stats[node]); kmem_cache_free(flow_cache, flow); } @@ -586,16 +598,28 @@ int ovs_flow_init(void) BUILD_BUG_ON(__alignof__(struct sw_flow_key) % __alignof__(long)); BUILD_BUG_ON(sizeof(struct sw_flow_key) % sizeof(long)); - flow_cache = kmem_cache_create("sw_flow", sizeof(struct sw_flow), 0, - 0, NULL); + flow_cache = kmem_cache_create("sw_flow", sizeof(struct sw_flow) + + (num_possible_nodes() + * sizeof(struct flow_stats *)), + 0, 0, NULL); if (flow_cache == NULL) return -ENOMEM; + flow_stats_cache + = kmem_cache_create("sw_flow_stats", sizeof(struct flow_stats), + 0, SLAB_HWCACHE_ALIGN, NULL); + if (flow_stats_cache == NULL) { + kmem_cache_destroy(flow_cache); + flow_cache = NULL; + return -ENOMEM; + } + return 0; } /* Uninitializes the flow module. */ void ovs_flow_exit(void) { + kmem_cache_destroy(flow_stats_cache); kmem_cache_destroy(flow_cache); } diff --git a/net/openvswitch/flow_table.h b/net/openvswitch/flow_table.h index c26c59a7ab57..ca8a5820f615 100644 --- a/net/openvswitch/flow_table.h +++ b/net/openvswitch/flow_table.h @@ -52,6 +52,8 @@ struct flow_table { unsigned int count; }; +extern struct kmem_cache *flow_stats_cache; + int ovs_flow_init(void); void ovs_flow_exit(void); -- 2.20.1