Use percpu allocator for stats due to objection to stats array.
But percpu allocator is not designed for high churn allocation/
deallcation. so we need to avoid allocating percpu flow for
short lived flows. One cheaper way to detect flow is by checking
if 5-tuple used in RSS are masked or not. if any one of them is
masked, flow is likely shared across CPU where percpu stat
should be more scalable. And that flow should be relatively
long lived flow.
Signed-off-by: Pravin B Shelar <pshe...@nicira.com>
---
datapath/datapath.c | 42 +++++++++++------------
datapath/flow.c | 85 +++++++++++++++++++++++++++++-----------------
datapath/flow.h | 17 +++++++--
datapath/flow_netlink.c | 26 ++++++++++++--
datapath/flow_netlink.h | 3 +-
datapath/flow_table.c | 33 ++++++++++++++++--
datapath/flow_table.h | 2 +-
7 files changed, 140 insertions(+), 68 deletions(-)
diff --git a/datapath/datapath.c b/datapath/datapath.c
index d0a8595..41f2f8b 100644
--- a/datapath/datapath.c
+++ b/datapath/datapath.c
@@ -494,7 +494,7 @@ static int ovs_packet_cmd_execute(struct sk_buff *skb,
struct genl_info *info)
packet->protocol = htons(ETH_P_802_2);
/* Build an sw_flow for sending this packet. */
- flow = ovs_flow_alloc();
+ flow = ovs_flow_alloc(false);
err = PTR_ERR(flow);
if (IS_ERR(flow))
goto err_kfree_skb;
@@ -629,7 +629,9 @@ static int ovs_flow_cmd_fill_info(struct sw_flow *flow,
struct datapath *dp,
{
const int skb_orig_len = skb->len;
struct nlattr *start;
- struct sw_flow_stats flow_stats;
+ struct ovs_flow_stats stats;
+ __be16 tcp_flags;
+ unsigned long used;
struct ovs_header *ovs_header;
struct nlattr *nla;
int err;
@@ -660,25 +662,17 @@ static int ovs_flow_cmd_fill_info(struct sw_flow *flow,
struct datapath *dp,
nla_nest_end(skb, nla);
- ovs_flow_stats_get(flow, &flow_stats);
- if (flow_stats.used &&
- nla_put_u64(skb, OVS_FLOW_ATTR_USED,
ovs_flow_used_time(flow_stats.used)))
+ ovs_flow_stats_get(flow, &used, &stats, &tcp_flags);
+ if (used &&
+ nla_put_u64(skb, OVS_FLOW_ATTR_USED, ovs_flow_used_time(used)))
goto nla_put_failure;
- if (flow_stats.packet_count) {
- struct ovs_flow_stats stats = {
- .n_packets = flow_stats.packet_count,
- .n_bytes = flow_stats.byte_count,
- };
-
- if (nla_put(skb, OVS_FLOW_ATTR_STATS,
- sizeof(struct ovs_flow_stats), &stats))
- goto nla_put_failure;
- }
+ if (stats.n_packets &&
+ nla_put(skb, OVS_FLOW_ATTR_STATS, sizeof(struct ovs_flow_stats),
&stats))
+ goto nla_put_failure;
- if ((u8)ntohs(flow_stats.tcp_flags) &&
- nla_put_u8(skb, OVS_FLOW_ATTR_TCP_FLAGS,
- (u8)ntohs(flow_stats.tcp_flags)))
+ if ((u8)ntohs(tcp_flags) &&
+ nla_put_u8(skb, OVS_FLOW_ATTR_TCP_FLAGS, (u8)ntohs(tcp_flags)))
goto nla_put_failure;
/* If OVS_FLOW_ATTR_ACTIONS doesn't fit, skip dumping the actions if
@@ -750,6 +744,7 @@ static int ovs_flow_cmd_new_or_set(struct sk_buff *skb,
struct genl_info *info)
struct nlattr **a = info->attrs;
struct ovs_header *ovs_header = info->userhdr;
struct sw_flow_key key, masked_key;
+ bool wc_5tupple = false;
struct sw_flow *flow = NULL;
struct sw_flow_mask mask;
struct sk_buff *reply;
@@ -765,7 +760,8 @@ static int ovs_flow_cmd_new_or_set(struct sk_buff *skb,
struct genl_info *info)
ovs_match_init(&match, &key, &mask);
error = ovs_nla_get_match(&match,
- a[OVS_FLOW_ATTR_KEY], a[OVS_FLOW_ATTR_MASK]);
+ a[OVS_FLOW_ATTR_KEY], a[OVS_FLOW_ATTR_MASK],
+ &wc_5tupple);
if (error)
goto error;
@@ -803,7 +799,7 @@ static int ovs_flow_cmd_new_or_set(struct sk_buff *skb,
struct genl_info *info)
goto err_unlock_ovs;
/* Allocate flow. */
- flow = ovs_flow_alloc();
+ flow = ovs_flow_alloc(wc_5tupple);
if (IS_ERR(flow)) {
error = PTR_ERR(flow);
goto err_unlock_ovs;
@@ -879,6 +875,7 @@ static int ovs_flow_cmd_get(struct sk_buff *skb, struct
genl_info *info)
{
struct nlattr **a = info->attrs;
struct ovs_header *ovs_header = info->userhdr;
+ bool wc_5tupple = false;
struct sw_flow_key key;
struct sk_buff *reply;
struct sw_flow *flow;
@@ -892,7 +889,7 @@ static int ovs_flow_cmd_get(struct sk_buff *skb, struct
genl_info *info)
}
ovs_match_init(&match, &key, NULL);
- err = ovs_nla_get_match(&match, a[OVS_FLOW_ATTR_KEY], NULL);
+ err = ovs_nla_get_match(&match, a[OVS_FLOW_ATTR_KEY], NULL,
&wc_5tupple);
if (err)
return err;
@@ -927,6 +924,7 @@ static int ovs_flow_cmd_del(struct sk_buff *skb, struct
genl_info *info)
{
struct nlattr **a = info->attrs;
struct ovs_header *ovs_header = info->userhdr;
+ bool wc_5tupple = false;
struct sw_flow_key key;
struct sk_buff *reply;
struct sw_flow *flow;
@@ -947,7 +945,7 @@ static int ovs_flow_cmd_del(struct sk_buff *skb, struct
genl_info *info)
}
ovs_match_init(&match, &key, NULL);
- err = ovs_nla_get_match(&match, a[OVS_FLOW_ATTR_KEY], NULL);
+ err = ovs_nla_get_match(&match, a[OVS_FLOW_ATTR_KEY], NULL,
&wc_5tupple);
if (err)
goto unlock;
diff --git a/datapath/flow.c b/datapath/flow.c
index 57eb6b5..4353a4f 100644
--- a/datapath/flow.c
+++ b/datapath/flow.c
@@ -64,48 +64,63 @@ u64 ovs_flow_used_time(unsigned long flow_jiffies)
void ovs_flow_stats_update(struct sw_flow *flow, struct sk_buff *skb)
{
- struct sw_flow_stats *stats = &flow->stats[smp_processor_id()];
+ struct flow_stats *stat;
__be16 tcp_flags = 0;
- if ((flow->key.eth.type == htons(ETH_P_IP) ||
- flow->key.eth.type == htons(ETH_P_IPV6)) &&
- flow->key.ip.proto == IPPROTO_TCP &&
- likely(skb->len >= skb_transport_offset(skb) + sizeof(struct
tcphdr))) {
- tcp_flags = TCP_FLAGS_BE16(tcp_hdr(skb));
- }
-
- spin_lock(&stats->lock);
- stats->used = jiffies;
- stats->packet_count++;
- stats->byte_count += skb->len;
- stats->tcp_flags |= tcp_flags;
- spin_unlock(&stats->lock);
+ if (!flow->stats.is_percpu) {
+ stat = flow->stats.stat;
+
+ if ((flow->key.eth.type == htons(ETH_P_IP) ||
+ flow->key.eth.type == htons(ETH_P_IPV6)) &&
+ flow->key.ip.proto == IPPROTO_TCP &&
+ likely(skb->len >= skb_transport_offset(skb) +
sizeof(struct tcphdr)))
+ tcp_flags = TCP_FLAGS_BE16(tcp_hdr(skb));
+ } else
+ stat = this_cpu_ptr(flow->stats.cpu_stats);
+
+ spin_lock(&stat->lock);
+ stat->used = jiffies;
+ stat->packet_count++;
+ stat->byte_count += skb->len;
+ if (!flow->stats.is_percpu)
+ flow->stats.tcp_flags |= tcp_flags;
+ spin_unlock(&stat->lock);
}
-void ovs_flow_stats_get(struct sw_flow *flow, struct sw_flow_stats *res)
+void ovs_flow_stats_get(struct sw_flow *flow, unsigned long *used,
+ struct ovs_flow_stats *stats, __be16 *tcp_flags)
{
int cpu, cur_cpu;
- memset(res, 0, sizeof(*res));
+ if (!flow->stats.is_percpu) {
+ *used = flow->stats.stat->used;
+ stats->n_packets = flow->stats.stat->packet_count;
+ stats->n_bytes = flow->stats.stat->byte_count;
+ *tcp_flags = flow->stats.tcp_flags;
+ return;
+ }
+
+ *used = 0;
+ *tcp_flags = 0;
+ memset(stats, 0, sizeof(*stats));
cur_cpu = get_cpu();
for_each_possible_cpu(cpu) {
- struct sw_flow_stats *stats = &flow->stats[cpu];
+ struct flow_stats *stat;
if (cpu == cur_cpu)
local_bh_disable();
- spin_lock(&stats->lock);
- if (time_after(stats->used, res->used))
- res->used = stats->used;
- res->packet_count += stats->packet_count;
- res->byte_count += stats->byte_count;
- res->tcp_flags |= stats->tcp_flags;
- spin_unlock(&stats->lock);
+ stat = per_cpu_ptr(flow->stats.cpu_stats, cpu);
+ spin_lock(&stat->lock);
+ if (time_after(stat->used, *used))
+ *used = stat->used;
+ stats->n_packets += stat->packet_count;
+ stats->n_bytes += stat->byte_count;
+ spin_unlock(&stat->lock);
if (cpu == cur_cpu)
local_bh_enable();
-
}
put_cpu();
}
@@ -114,19 +129,25 @@ void ovs_flow_stats_clear(struct sw_flow *flow)
{
int cpu, cur_cpu;
+ if (!flow->stats.is_percpu) {
+ memset(flow->stats.stat, 0, sizeof(*flow->stats.stat));
+ flow->stats.tcp_flags = 0;
+ return;
+ }
+
cur_cpu = get_cpu();
for_each_possible_cpu(cpu) {
- struct sw_flow_stats *stats = &flow->stats[cpu];
+ struct flow_stats *stat;
if (cpu == cur_cpu)
local_bh_disable();
- spin_lock(&stats->lock);
- stats->used = 0;
- stats->packet_count = 0;
- stats->byte_count = 0;
- stats->tcp_flags = 0;
- spin_unlock(&stats->lock);
+ stat = per_cpu_ptr(flow->stats.cpu_stats, cpu);
+ spin_lock(&stat->lock);
+ stat->used = 0;
+ stat->packet_count = 0;
+ stat->byte_count = 0;
+ spin_unlock(&stat->lock);
if (cpu == cur_cpu)
local_bh_enable();
diff --git a/datapath/flow.h b/datapath/flow.h
index 6b68cf1..2405638 100644
--- a/datapath/flow.h
+++ b/datapath/flow.h
@@ -149,13 +149,21 @@ struct sw_flow_actions {
struct nlattr actions[];
};
-struct sw_flow_stats {
+struct flow_stats {
u64 packet_count; /* Number of packets matched. */
u64 byte_count; /* Number of bytes matched. */
unsigned long used; /* Last used time (in jiffies). */
spinlock_t lock; /* Lock for atomic stats update. */
+};
+
+struct sw_flow_stats {
__be16 tcp_flags; /* Union of seen TCP flags. */
-} ____cacheline_aligned_in_smp;
+ bool is_percpu;
+ union {
+ struct flow_stats *stat;
+ struct flow_stats __percpu *cpu_stats;
+ };
+};
struct sw_flow {
struct rcu_head rcu;
@@ -166,7 +174,7 @@ struct sw_flow {
struct sw_flow_key unmasked_key;
struct sw_flow_mask *mask;
struct sw_flow_actions __rcu *sf_acts;
- struct sw_flow_stats stats[];
+ struct sw_flow_stats stats;
};
struct arp_eth_header {
@@ -184,7 +192,8 @@ struct arp_eth_header {
} __packed;
void ovs_flow_stats_update(struct sw_flow *flow, struct sk_buff *skb);
-void ovs_flow_stats_get(struct sw_flow *flow, struct sw_flow_stats *res);
+void ovs_flow_stats_get(struct sw_flow *flow, unsigned long *used,
+ struct ovs_flow_stats *stats, __be16 *tcp_flags);
void ovs_flow_stats_clear(struct sw_flow *flow);
u64 ovs_flow_used_time(unsigned long flow_jiffies);
diff --git a/datapath/flow_netlink.c b/datapath/flow_netlink.c
index 75c72b3..631aa59 100644
--- a/datapath/flow_netlink.c
+++ b/datapath/flow_netlink.c
@@ -488,7 +488,8 @@ static int metadata_from_nlattrs(struct sw_flow_match
*match, u64 *attrs,
}
static int ovs_key_from_nlattrs(struct sw_flow_match *match, u64 attrs,
- const struct nlattr **a, bool is_mask)
+ const struct nlattr **a, bool is_mask,
+ bool *wc_5tupple)
{
int err;
u64 orig_attrs = attrs;
@@ -567,6 +568,12 @@ static int ovs_key_from_nlattrs(struct sw_flow_match
*match, u64 attrs,
SW_FLOW_KEY_PUT(match, ipv4.addr.dst,
ipv4_key->ipv4_dst, is_mask);
attrs &= ~(1ULL << OVS_KEY_ATTR_IPV4);
+
+ if (is_mask) {
+ if (!ipv4_key->ipv4_proto ||
+ !ipv4_key->ipv4_src || !ipv4_key->ipv4_dst)
+ *wc_5tupple = true;
+ }
}
if (attrs & (1ULL << OVS_KEY_ATTR_IPV6)) {
@@ -598,6 +605,13 @@ static int ovs_key_from_nlattrs(struct sw_flow_match
*match, u64 attrs,
is_mask);
attrs &= ~(1ULL << OVS_KEY_ATTR_IPV6);
+
+ if (is_mask) {
+ if (!ipv6_key->ipv6_proto ||
+ is_all_zero((u8 *)ipv6_key->ipv6_src,
sizeof(match->key->ipv6.addr.src)) ||
+ is_all_zero((u8 *)ipv6_key->ipv6_dst,
sizeof(match->key->ipv6.addr.dst)))
+ *wc_5tupple = true;
+ }
}
if (attrs & (1ULL << OVS_KEY_ATTR_ARP)) {
@@ -640,6 +654,9 @@ static int ovs_key_from_nlattrs(struct sw_flow_match
*match, u64 attrs,
tcp_key->tcp_dst, is_mask);
}
attrs &= ~(1ULL << OVS_KEY_ATTR_TCP);
+
+ if (is_mask && (!tcp_key->tcp_src || !tcp_key->tcp_dst))
+ *wc_5tupple = true;
}
if (attrs & (1ULL << OVS_KEY_ATTR_TCP_FLAGS)) {
@@ -757,7 +774,8 @@ static void sw_flow_mask_set(struct sw_flow_mask *mask,
*/
int ovs_nla_get_match(struct sw_flow_match *match,
const struct nlattr *key,
- const struct nlattr *mask)
+ const struct nlattr *mask,
+ bool *wc_5tupple)
{
const struct nlattr *a[OVS_KEY_ATTR_MAX + 1];
const struct nlattr *encap;
@@ -803,7 +821,7 @@ int ovs_nla_get_match(struct sw_flow_match *match,
}
}
- err = ovs_key_from_nlattrs(match, key_attrs, a, false);
+ err = ovs_key_from_nlattrs(match, key_attrs, a, false, wc_5tupple);
if (err)
return err;
@@ -844,7 +862,7 @@ int ovs_nla_get_match(struct sw_flow_match *match,
}
}
- err = ovs_key_from_nlattrs(match, mask_attrs, a, true);
+ err = ovs_key_from_nlattrs(match, mask_attrs, a, true,
wc_5tupple);
if (err)
return err;
} else {
diff --git a/datapath/flow_netlink.h b/datapath/flow_netlink.h
index 4401510..8467931 100644
--- a/datapath/flow_netlink.h
+++ b/datapath/flow_netlink.h
@@ -46,7 +46,8 @@ int ovs_nla_get_flow_metadata(struct sw_flow *flow,
const struct nlattr *attr);
int ovs_nla_get_match(struct sw_flow_match *match,
const struct nlattr *,
- const struct nlattr *);
+ const struct nlattr *,
+ bool *wc_5tupple);
int ovs_nla_copy_actions(const struct nlattr *attr,
const struct sw_flow_key *key, int depth,
diff --git a/datapath/flow_table.c b/datapath/flow_table.c
index 29cfcbe..e59a8fe 100644
--- a/datapath/flow_table.c
+++ b/datapath/flow_table.c
@@ -73,7 +73,7 @@ void ovs_flow_mask_key(struct sw_flow_key *dst, const struct
sw_flow_key *src,
*d++ = *s++ & *m++;
}
-struct sw_flow *ovs_flow_alloc(void)
+struct sw_flow *ovs_flow_alloc(bool percpu_stats)
{
struct sw_flow *flow;
int cpu;
@@ -85,11 +85,32 @@ struct sw_flow *ovs_flow_alloc(void)
flow->sf_acts = NULL;
flow->mask = NULL;
- memset(flow->stats, 0, num_possible_cpus() * sizeof(struct
sw_flow_stats));
- for_each_possible_cpu(cpu)
- spin_lock_init(&flow->stats[cpu].lock);
+ flow->stats.is_percpu = percpu_stats;
+
+ if (!percpu_stats) {
+ flow->stats.tcp_flags = 0;
+
+ flow->stats.stat = kzalloc(sizeof(*flow->stats.stat),
GFP_KERNEL);
+ if (!flow->stats.stat)
+ goto err;
+ return flow;
+ }
+
+ flow->stats.cpu_stats = alloc_percpu(struct flow_stats);
+ if (!flow->stats.cpu_stats)
+ goto err;
+
+ for_each_possible_cpu(cpu) {
+ struct flow_stats *cpu_stats;
+
+ cpu_stats = per_cpu_ptr(flow->stats.cpu_stats, cpu);
+ spin_lock_init(&cpu_stats->lock);
+ }
return flow;
+err:
+ kfree(flow);
+ return ERR_PTR(-ENOMEM);
}
int ovs_flow_tbl_count(struct flow_table *table)
@@ -123,6 +144,10 @@ static struct flex_array *alloc_buckets(unsigned int
n_buckets)
static void flow_free(struct sw_flow *flow)
{
kfree((struct sf_flow_acts __force *)flow->sf_acts);
+ if (flow->stats.is_percpu)
+ free_percpu(flow->stats.cpu_stats);
+ else
+ kfree(flow->stats.stat);
kmem_cache_free(flow_cache, flow);
}
diff --git a/datapath/flow_table.h b/datapath/flow_table.h
index f54aa82..1996e34 100644
--- a/datapath/flow_table.h
+++ b/datapath/flow_table.h
@@ -55,7 +55,7 @@ struct flow_table {
int ovs_flow_init(void);
void ovs_flow_exit(void);
-struct sw_flow *ovs_flow_alloc(void);
+struct sw_flow *ovs_flow_alloc(bool percpu_stats);
void ovs_flow_free(struct sw_flow *, bool deferred);
int ovs_flow_tbl_init(struct flow_table *);
--
1.7.1
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
