linux/net/openvswitch/flow.c

773 lines
20 KiB
C
Raw Normal View History

/*
* Copyright (c) 2007-2014 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.
*
* This program is distributed in the hope that it will be useful, but
* WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
* 02110-1301, USA
*/
#include <linux/uaccess.h>
#include <linux/netdevice.h>
#include <linux/etherdevice.h>
#include <linux/if_ether.h>
#include <linux/if_vlan.h>
#include <net/llc_pdu.h>
#include <linux/kernel.h>
#include <linux/jhash.h>
#include <linux/jiffies.h>
#include <linux/llc.h>
#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>
#include <linux/mpls.h>
#include <linux/sctp.h>
#include <linux/smp.h>
#include <linux/tcp.h>
#include <linux/udp.h>
#include <linux/icmp.h>
#include <linux/icmpv6.h>
#include <linux/rculist.h>
#include <net/ip.h>
#include <net/ip_tunnels.h>
#include <net/ipv6.h>
#include <net/mpls.h>
#include <net/ndisc.h>
#include "conntrack.h"
#include "datapath.h"
#include "flow.h"
#include "flow_netlink.h"
#include "vport.h"
u64 ovs_flow_used_time(unsigned long flow_jiffies)
{
struct timespec cur_ts;
u64 cur_ms, idle_ms;
ktime_get_ts(&cur_ts);
idle_ms = jiffies_to_msecs(jiffies - flow_jiffies);
cur_ms = (u64)cur_ts.tv_sec * MSEC_PER_SEC +
cur_ts.tv_nsec / NSEC_PER_MSEC;
return cur_ms - idle_ms;
}
#define TCP_FLAGS_BE16(tp) (*(__be16 *)&tcp_flag_word(tp) & htons(0x0FFF))
void ovs_flow_stats_update(struct sw_flow *flow, __be16 tcp_flags,
const struct sk_buff *skb)
{
struct flow_stats *stats;
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 <jrajahalme@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com>
2014-03-28 03:42:54 +08:00
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[cpu]);
/* Check if already have CPU-specific stats. */
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 <jrajahalme@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com>
2014-03-28 03:42:54 +08:00
if (likely(stats)) {
spin_lock(&stats->lock);
/* Mark if we write on the pre-allocated stats. */
if (cpu == 0 && unlikely(flow->stats_last_writer != cpu))
flow->stats_last_writer = cpu;
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 <jrajahalme@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com>
2014-03-28 03:42:54 +08:00
} else {
stats = rcu_dereference(flow->stats[0]); /* Pre-allocated. */
spin_lock(&stats->lock);
/* If the current CPU is the only writer on the
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 <jrajahalme@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com>
2014-03-28 03:42:54 +08:00
* pre-allocated stats keep using them.
*/
if (unlikely(flow->stats_last_writer != cpu)) {
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 <jrajahalme@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com>
2014-03-28 03:42:54 +08:00
/* A previous locker may have already allocated the
* stats, so we need to check again. If CPU-specific
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 <jrajahalme@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com>
2014-03-28 03:42:54 +08:00
* stats were already allocated, we update the pre-
* allocated stats as we have already locked them.
*/
if (likely(flow->stats_last_writer != -1) &&
likely(!rcu_access_pointer(flow->stats[cpu]))) {
/* Try to allocate CPU-specific stats. */
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 <jrajahalme@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com>
2014-03-28 03:42:54 +08:00
struct flow_stats *new_stats;
new_stats =
kmem_cache_alloc_node(flow_stats_cache,
mm: remove GFP_THISNODE NOTE: this is not about __GFP_THISNODE, this is only about GFP_THISNODE. GFP_THISNODE is a secret combination of gfp bits that have different behavior than expected. It is a combination of __GFP_THISNODE, __GFP_NORETRY, and __GFP_NOWARN and is special-cased in the page allocator slowpath to fail without trying reclaim even though it may be used in combination with __GFP_WAIT. An example of the problem this creates: commit e97ca8e5b864 ("mm: fix GFP_THISNODE callers and clarify") fixed up many users of GFP_THISNODE that really just wanted __GFP_THISNODE. The problem doesn't end there, however, because even it was a no-op for alloc_misplaced_dst_page(), which also sets __GFP_NORETRY and __GFP_NOWARN, and migrate_misplaced_transhuge_page(), where __GFP_NORETRY and __GFP_NOWAIT is set in GFP_TRANSHUGE. Converting GFP_THISNODE to __GFP_THISNODE is a no-op in these cases since the page allocator special-cases __GFP_THISNODE && __GFP_NORETRY && __GFP_NOWARN. It's time to just remove GFP_THISNODE entirely. We leave __GFP_THISNODE to restrict an allocation to a local node, but remove GFP_THISNODE and its obscurity. Instead, we require that a caller clear __GFP_WAIT if it wants to avoid reclaim. This allows the aforementioned functions to actually reclaim as they should. It also enables any future callers that want to do __GFP_THISNODE but also __GFP_NORETRY && __GFP_NOWARN to reclaim. The rule is simple: if you don't want to reclaim, then don't set __GFP_WAIT. Aside: ovs_flow_stats_update() really wants to avoid reclaim as well, so it is unchanged. Signed-off-by: David Rientjes <rientjes@google.com> Acked-by: Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Lameter <cl@linux.com> Acked-by: Pekka Enberg <penberg@kernel.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Pravin Shelar <pshelar@nicira.com> Cc: Jarno Rajahalme <jrajahalme@nicira.com> Cc: Li Zefan <lizefan@huawei.com> Cc: Greg Thelen <gthelen@google.com> Cc: Tejun Heo <tj@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-04-15 06:46:55 +08:00
GFP_NOWAIT |
__GFP_THISNODE |
__GFP_NOWARN |
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 <jrajahalme@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com>
2014-03-28 03:42:54 +08:00
__GFP_NOMEMALLOC,
node);
if (likely(new_stats)) {
new_stats->used = jiffies;
new_stats->packet_count = 1;
new_stats->byte_count = len;
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 <jrajahalme@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com>
2014-03-28 03:42:54 +08:00
new_stats->tcp_flags = tcp_flags;
spin_lock_init(&new_stats->lock);
rcu_assign_pointer(flow->stats[cpu],
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 <jrajahalme@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com>
2014-03-28 03:42:54 +08:00
new_stats);
goto unlock;
}
}
flow->stats_last_writer = cpu;
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 <jrajahalme@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com>
2014-03-28 03:42:54 +08:00
}
}
stats->used = jiffies;
stats->packet_count++;
stats->byte_count += len;
stats->tcp_flags |= tcp_flags;
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 <jrajahalme@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com>
2014-03-28 03:42:54 +08:00
unlock:
spin_unlock(&stats->lock);
}
/* Must be called with rcu_read_lock or ovs_mutex. */
void ovs_flow_stats_get(const struct sw_flow *flow,
struct ovs_flow_stats *ovs_stats,
unsigned long *used, __be16 *tcp_flags)
{
int cpu;
*used = 0;
*tcp_flags = 0;
memset(ovs_stats, 0, sizeof(*ovs_stats));
/* 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]);
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 <jrajahalme@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com>
2014-03-28 03:42:54 +08:00
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);
}
}
}
/* Called with ovs_mutex. */
void ovs_flow_stats_clear(struct sw_flow *flow)
{
int cpu;
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 <jrajahalme@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com>
2014-03-28 03:42:54 +08:00
/* 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]);
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 <jrajahalme@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com>
2014-03-28 03:42:54 +08:00
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)
{
if (unlikely(skb->len < len))
return -EINVAL;
if (unlikely(!pskb_may_pull(skb, len)))
return -ENOMEM;
return 0;
}
static bool arphdr_ok(struct sk_buff *skb)
{
return pskb_may_pull(skb, skb_network_offset(skb) +
sizeof(struct arp_eth_header));
}
static int check_iphdr(struct sk_buff *skb)
{
unsigned int nh_ofs = skb_network_offset(skb);
unsigned int ip_len;
int err;
err = check_header(skb, nh_ofs + sizeof(struct iphdr));
if (unlikely(err))
return err;
ip_len = ip_hdrlen(skb);
if (unlikely(ip_len < sizeof(struct iphdr) ||
skb->len < nh_ofs + ip_len))
return -EINVAL;
skb_set_transport_header(skb, nh_ofs + ip_len);
return 0;
}
static bool tcphdr_ok(struct sk_buff *skb)
{
int th_ofs = skb_transport_offset(skb);
int tcp_len;
if (unlikely(!pskb_may_pull(skb, th_ofs + sizeof(struct tcphdr))))
return false;
tcp_len = tcp_hdrlen(skb);
if (unlikely(tcp_len < sizeof(struct tcphdr) ||
skb->len < th_ofs + tcp_len))
return false;
return true;
}
static bool udphdr_ok(struct sk_buff *skb)
{
return pskb_may_pull(skb, skb_transport_offset(skb) +
sizeof(struct udphdr));
}
static bool sctphdr_ok(struct sk_buff *skb)
{
return pskb_may_pull(skb, skb_transport_offset(skb) +
sizeof(struct sctphdr));
}
static bool icmphdr_ok(struct sk_buff *skb)
{
return pskb_may_pull(skb, skb_transport_offset(skb) +
sizeof(struct icmphdr));
}
static int parse_ipv6hdr(struct sk_buff *skb, struct sw_flow_key *key)
{
unsigned int nh_ofs = skb_network_offset(skb);
unsigned int nh_len;
int payload_ofs;
struct ipv6hdr *nh;
uint8_t nexthdr;
__be16 frag_off;
int err;
err = check_header(skb, nh_ofs + sizeof(*nh));
if (unlikely(err))
return err;
nh = ipv6_hdr(skb);
nexthdr = nh->nexthdr;
payload_ofs = (u8 *)(nh + 1) - skb->data;
key->ip.proto = NEXTHDR_NONE;
key->ip.tos = ipv6_get_dsfield(nh);
key->ip.ttl = nh->hop_limit;
key->ipv6.label = *(__be32 *)nh & htonl(IPV6_FLOWINFO_FLOWLABEL);
key->ipv6.addr.src = nh->saddr;
key->ipv6.addr.dst = nh->daddr;
payload_ofs = ipv6_skip_exthdr(skb, payload_ofs, &nexthdr, &frag_off);
if (frag_off) {
if (frag_off & htons(~0x7))
key->ip.frag = OVS_FRAG_TYPE_LATER;
else
key->ip.frag = OVS_FRAG_TYPE_FIRST;
} else {
key->ip.frag = OVS_FRAG_TYPE_NONE;
}
/* Delayed handling of error in ipv6_skip_exthdr() as it
* always sets frag_off to a valid value which may be
* used to set key->ip.frag above.
*/
if (unlikely(payload_ofs < 0))
return -EPROTO;
nh_len = payload_ofs - nh_ofs;
skb_set_transport_header(skb, nh_ofs + nh_len);
key->ip.proto = nexthdr;
return nh_len;
}
static bool icmp6hdr_ok(struct sk_buff *skb)
{
return pskb_may_pull(skb, skb_transport_offset(skb) +
sizeof(struct icmp6hdr));
}
/**
* Parse vlan tag from vlan header.
* Returns ERROR on memory error.
* Returns 0 if it encounters a non-vlan or incomplete packet.
* Returns 1 after successfully parsing vlan tag.
*/
static int parse_vlan_tag(struct sk_buff *skb, struct vlan_head *key_vh)
{
struct vlan_head *vh = (struct vlan_head *)skb->data;
if (likely(!eth_type_vlan(vh->tpid)))
return 0;
if (unlikely(skb->len < sizeof(struct vlan_head) + sizeof(__be16)))
return 0;
if (unlikely(!pskb_may_pull(skb, sizeof(struct vlan_head) +
sizeof(__be16))))
return -ENOMEM;
vh = (struct vlan_head *)skb->data;
key_vh->tci = vh->tci | htons(VLAN_TAG_PRESENT);
key_vh->tpid = vh->tpid;
__skb_pull(skb, sizeof(struct vlan_head));
return 1;
}
static int parse_vlan(struct sk_buff *skb, struct sw_flow_key *key)
{
int res;
key->eth.vlan.tci = 0;
key->eth.vlan.tpid = 0;
key->eth.cvlan.tci = 0;
key->eth.cvlan.tpid = 0;
if (skb_vlan_tag_present(skb)) {
key->eth.vlan.tci = htons(skb->vlan_tci);
key->eth.vlan.tpid = skb->vlan_proto;
} else {
/* Parse outer vlan tag in the non-accelerated case. */
res = parse_vlan_tag(skb, &key->eth.vlan);
if (res <= 0)
return res;
}
/* Parse inner vlan tag. */
res = parse_vlan_tag(skb, &key->eth.cvlan);
if (res <= 0)
return res;
return 0;
}
static __be16 parse_ethertype(struct sk_buff *skb)
{
struct llc_snap_hdr {
u8 dsap; /* Always 0xAA */
u8 ssap; /* Always 0xAA */
u8 ctrl;
u8 oui[3];
__be16 ethertype;
};
struct llc_snap_hdr *llc;
__be16 proto;
proto = *(__be16 *) skb->data;
__skb_pull(skb, sizeof(__be16));
if (eth_proto_is_802_3(proto))
return proto;
if (skb->len < sizeof(struct llc_snap_hdr))
return htons(ETH_P_802_2);
if (unlikely(!pskb_may_pull(skb, sizeof(struct llc_snap_hdr))))
return htons(0);
llc = (struct llc_snap_hdr *) skb->data;
if (llc->dsap != LLC_SAP_SNAP ||
llc->ssap != LLC_SAP_SNAP ||
(llc->oui[0] | llc->oui[1] | llc->oui[2]) != 0)
return htons(ETH_P_802_2);
__skb_pull(skb, sizeof(struct llc_snap_hdr));
if (eth_proto_is_802_3(llc->ethertype))
return llc->ethertype;
return htons(ETH_P_802_2);
}
static int parse_icmpv6(struct sk_buff *skb, struct sw_flow_key *key,
int nh_len)
{
struct icmp6hdr *icmp = icmp6_hdr(skb);
/* The ICMPv6 type and code fields use the 16-bit transport port
* fields, so we need to store them in 16-bit network byte order.
*/
key->tp.src = htons(icmp->icmp6_type);
key->tp.dst = htons(icmp->icmp6_code);
memset(&key->ipv6.nd, 0, sizeof(key->ipv6.nd));
if (icmp->icmp6_code == 0 &&
(icmp->icmp6_type == NDISC_NEIGHBOUR_SOLICITATION ||
icmp->icmp6_type == NDISC_NEIGHBOUR_ADVERTISEMENT)) {
int icmp_len = skb->len - skb_transport_offset(skb);
struct nd_msg *nd;
int offset;
/* In order to process neighbor discovery options, we need the
* entire packet.
*/
if (unlikely(icmp_len < sizeof(*nd)))
return 0;
if (unlikely(skb_linearize(skb)))
return -ENOMEM;
nd = (struct nd_msg *)skb_transport_header(skb);
key->ipv6.nd.target = nd->target;
icmp_len -= sizeof(*nd);
offset = 0;
while (icmp_len >= 8) {
struct nd_opt_hdr *nd_opt =
(struct nd_opt_hdr *)(nd->opt + offset);
int opt_len = nd_opt->nd_opt_len * 8;
if (unlikely(!opt_len || opt_len > icmp_len))
return 0;
/* Store the link layer address if the appropriate
* option is provided. It is considered an error if
* the same link layer option is specified twice.
*/
if (nd_opt->nd_opt_type == ND_OPT_SOURCE_LL_ADDR
&& opt_len == 8) {
if (unlikely(!is_zero_ether_addr(key->ipv6.nd.sll)))
goto invalid;
ether_addr_copy(key->ipv6.nd.sll,
&nd->opt[offset+sizeof(*nd_opt)]);
} else if (nd_opt->nd_opt_type == ND_OPT_TARGET_LL_ADDR
&& opt_len == 8) {
if (unlikely(!is_zero_ether_addr(key->ipv6.nd.tll)))
goto invalid;
ether_addr_copy(key->ipv6.nd.tll,
&nd->opt[offset+sizeof(*nd_opt)]);
}
icmp_len -= opt_len;
offset += opt_len;
}
}
return 0;
invalid:
memset(&key->ipv6.nd.target, 0, sizeof(key->ipv6.nd.target));
memset(key->ipv6.nd.sll, 0, sizeof(key->ipv6.nd.sll));
memset(key->ipv6.nd.tll, 0, sizeof(key->ipv6.nd.tll));
return 0;
}
/**
* key_extract - extracts a flow key from an Ethernet frame.
* @skb: sk_buff that contains the frame, with skb->data pointing to the
* Ethernet header
* @key: output flow key
*
* The caller must ensure that skb->len >= ETH_HLEN.
*
* Returns 0 if successful, otherwise a negative errno value.
*
* Initializes @skb header pointers as follows:
*
* - skb->mac_header: the Ethernet header.
*
* - skb->network_header: just past the Ethernet header, or just past the
* VLAN header, to the first byte of the Ethernet payload.
*
* - skb->transport_header: If key->eth.type is ETH_P_IP or ETH_P_IPV6
* on output, then just past the IP header, if one is present and
* of a correct length, otherwise the same as skb->network_header.
* For other key->eth.type values it is left untouched.
*/
static int key_extract(struct sk_buff *skb, struct sw_flow_key *key)
{
int error;
struct ethhdr *eth;
/* Flags are always used as part of stats */
key->tp.flags = 0;
skb_reset_mac_header(skb);
/* Link layer. We are guaranteed to have at least the 14 byte Ethernet
* header in the linear data area.
*/
eth = eth_hdr(skb);
ether_addr_copy(key->eth.src, eth->h_source);
ether_addr_copy(key->eth.dst, eth->h_dest);
__skb_pull(skb, 2 * ETH_ALEN);
/* We are going to push all headers that we pull, so no need to
* update skb->csum here.
*/
if (unlikely(parse_vlan(skb, key)))
return -ENOMEM;
key->eth.type = parse_ethertype(skb);
if (unlikely(key->eth.type == htons(0)))
return -ENOMEM;
skb_reset_network_header(skb);
skb_reset_mac_len(skb);
__skb_push(skb, skb->data - skb_mac_header(skb));
/* Network layer. */
if (key->eth.type == htons(ETH_P_IP)) {
struct iphdr *nh;
__be16 offset;
error = check_iphdr(skb);
if (unlikely(error)) {
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
memset(&key->ip, 0, sizeof(key->ip));
memset(&key->ipv4, 0, sizeof(key->ipv4));
if (error == -EINVAL) {
skb->transport_header = skb->network_header;
error = 0;
}
return error;
}
nh = ip_hdr(skb);
key->ipv4.addr.src = nh->saddr;
key->ipv4.addr.dst = nh->daddr;
key->ip.proto = nh->protocol;
key->ip.tos = nh->tos;
key->ip.ttl = nh->ttl;
offset = nh->frag_off & htons(IP_OFFSET);
if (offset) {
key->ip.frag = OVS_FRAG_TYPE_LATER;
return 0;
}
if (nh->frag_off & htons(IP_MF) ||
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
key->ip.frag = OVS_FRAG_TYPE_FIRST;
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
else
key->ip.frag = OVS_FRAG_TYPE_NONE;
/* Transport layer. */
if (key->ip.proto == IPPROTO_TCP) {
if (tcphdr_ok(skb)) {
struct tcphdr *tcp = tcp_hdr(skb);
key->tp.src = tcp->source;
key->tp.dst = tcp->dest;
key->tp.flags = TCP_FLAGS_BE16(tcp);
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
} else {
memset(&key->tp, 0, sizeof(key->tp));
}
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
} else if (key->ip.proto == IPPROTO_UDP) {
if (udphdr_ok(skb)) {
struct udphdr *udp = udp_hdr(skb);
key->tp.src = udp->source;
key->tp.dst = udp->dest;
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
} else {
memset(&key->tp, 0, sizeof(key->tp));
}
} else if (key->ip.proto == IPPROTO_SCTP) {
if (sctphdr_ok(skb)) {
struct sctphdr *sctp = sctp_hdr(skb);
key->tp.src = sctp->source;
key->tp.dst = sctp->dest;
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
} else {
memset(&key->tp, 0, sizeof(key->tp));
}
} else if (key->ip.proto == IPPROTO_ICMP) {
if (icmphdr_ok(skb)) {
struct icmphdr *icmp = icmp_hdr(skb);
/* The ICMP type and code fields use the 16-bit
* transport port fields, so we need to store
* them in 16-bit network byte order. */
key->tp.src = htons(icmp->type);
key->tp.dst = htons(icmp->code);
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
} else {
memset(&key->tp, 0, sizeof(key->tp));
}
}
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
} else if (key->eth.type == htons(ETH_P_ARP) ||
key->eth.type == htons(ETH_P_RARP)) {
struct arp_eth_header *arp;
bool arp_available = arphdr_ok(skb);
arp = (struct arp_eth_header *)skb_network_header(skb);
if (arp_available &&
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
arp->ar_hrd == htons(ARPHRD_ETHER) &&
arp->ar_pro == htons(ETH_P_IP) &&
arp->ar_hln == ETH_ALEN &&
arp->ar_pln == 4) {
/* We only match on the lower 8 bits of the opcode. */
if (ntohs(arp->ar_op) <= 0xff)
key->ip.proto = ntohs(arp->ar_op);
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
else
key->ip.proto = 0;
memcpy(&key->ipv4.addr.src, arp->ar_sip, sizeof(key->ipv4.addr.src));
memcpy(&key->ipv4.addr.dst, arp->ar_tip, sizeof(key->ipv4.addr.dst));
ether_addr_copy(key->ipv4.arp.sha, arp->ar_sha);
ether_addr_copy(key->ipv4.arp.tha, arp->ar_tha);
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
} else {
memset(&key->ip, 0, sizeof(key->ip));
memset(&key->ipv4, 0, sizeof(key->ipv4));
}
} else if (eth_p_mpls(key->eth.type)) {
size_t stack_len = MPLS_HLEN;
skb_set_inner_network_header(skb, skb->mac_len);
while (1) {
__be32 lse;
error = check_header(skb, skb->mac_len + stack_len);
if (unlikely(error))
return 0;
memcpy(&lse, skb_inner_network_header(skb), MPLS_HLEN);
if (stack_len == MPLS_HLEN)
memcpy(&key->mpls.top_lse, &lse, MPLS_HLEN);
skb_set_inner_network_header(skb, skb->mac_len + stack_len);
if (lse & htonl(MPLS_LS_S_MASK))
break;
stack_len += MPLS_HLEN;
}
} else if (key->eth.type == htons(ETH_P_IPV6)) {
int nh_len; /* IPv6 Header + Extensions */
nh_len = parse_ipv6hdr(skb, key);
if (unlikely(nh_len < 0)) {
switch (nh_len) {
case -EINVAL:
memset(&key->ip, 0, sizeof(key->ip));
memset(&key->ipv6.addr, 0, sizeof(key->ipv6.addr));
/* fall-through */
case -EPROTO:
skb->transport_header = skb->network_header;
error = 0;
break;
default:
error = nh_len;
}
return error;
}
if (key->ip.frag == OVS_FRAG_TYPE_LATER)
return 0;
if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
key->ip.frag = OVS_FRAG_TYPE_FIRST;
/* Transport layer. */
if (key->ip.proto == NEXTHDR_TCP) {
if (tcphdr_ok(skb)) {
struct tcphdr *tcp = tcp_hdr(skb);
key->tp.src = tcp->source;
key->tp.dst = tcp->dest;
key->tp.flags = TCP_FLAGS_BE16(tcp);
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
} else {
memset(&key->tp, 0, sizeof(key->tp));
}
} else if (key->ip.proto == NEXTHDR_UDP) {
if (udphdr_ok(skb)) {
struct udphdr *udp = udp_hdr(skb);
key->tp.src = udp->source;
key->tp.dst = udp->dest;
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
} else {
memset(&key->tp, 0, sizeof(key->tp));
}
} else if (key->ip.proto == NEXTHDR_SCTP) {
if (sctphdr_ok(skb)) {
struct sctphdr *sctp = sctp_hdr(skb);
key->tp.src = sctp->source;
key->tp.dst = sctp->dest;
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
} else {
memset(&key->tp, 0, sizeof(key->tp));
}
} else if (key->ip.proto == NEXTHDR_ICMP) {
if (icmp6hdr_ok(skb)) {
error = parse_icmpv6(skb, key, nh_len);
if (error)
return error;
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
} else {
memset(&key->tp, 0, sizeof(key->tp));
}
}
}
return 0;
}
int ovs_flow_key_update(struct sk_buff *skb, struct sw_flow_key *key)
{
return key_extract(skb, key);
}
int ovs_flow_key_extract(const struct ip_tunnel_info *tun_info,
struct sk_buff *skb, struct sw_flow_key *key)
{
/* Extract metadata from packet. */
if (tun_info) {
key->tun_proto = ip_tunnel_info_af(tun_info);
memcpy(&key->tun_key, &tun_info->key, sizeof(key->tun_key));
if (tun_info->options_len) {
BUILD_BUG_ON((1 << (sizeof(tun_info->options_len) *
8)) - 1
> sizeof(key->tun_opts));
ip_tunnel_info_opts_get(TUN_METADATA_OPTS(key, tun_info->options_len),
tun_info);
key->tun_opts_len = tun_info->options_len;
} else {
key->tun_opts_len = 0;
}
} else {
key->tun_proto = 0;
key->tun_opts_len = 0;
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
memset(&key->tun_key, 0, sizeof(key->tun_key));
}
key->phy.priority = skb->priority;
key->phy.in_port = OVS_CB(skb)->input_vport->port_no;
key->phy.skb_mark = skb->mark;
openvswitch: Add conntrack action Expose the kernel connection tracker via OVS. Userspace components can make use of the CT action to populate the connection state (ct_state) field for a flow. This state can be subsequently matched. Exposed connection states are OVS_CS_F_*: - NEW (0x01) - Beginning of a new connection. - ESTABLISHED (0x02) - Part of an existing connection. - RELATED (0x04) - Related to an established connection. - INVALID (0x20) - Could not track the connection for this packet. - REPLY_DIR (0x40) - This packet is in the reply direction for the flow. - TRACKED (0x80) - This packet has been sent through conntrack. When the CT action is executed by itself, it will send the packet through the connection tracker and populate the ct_state field with one or more of the connection state flags above. The CT action will always set the TRACKED bit. When the COMMIT flag is passed to the conntrack action, this specifies that information about the connection should be stored. This allows subsequent packets for the same (or related) connections to be correlated with this connection. Sending subsequent packets for the connection through conntrack allows the connection tracker to consider the packets as ESTABLISHED, RELATED, and/or REPLY_DIR. The CT action may optionally take a zone to track the flow within. This allows connections with the same 5-tuple to be kept logically separate from connections in other zones. If the zone is specified, then the "ct_zone" match field will be subsequently populated with the zone id. IP fragments are handled by transparently assembling them as part of the CT action. The maximum received unit (MRU) size is tracked so that refragmentation can occur during output. IP frag handling contributed by Andy Zhou. Based on original design by Justin Pettit. Signed-off-by: Joe Stringer <joestringer@nicira.com> Signed-off-by: Justin Pettit <jpettit@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Thomas Graf <tgraf@suug.ch> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-08-27 02:31:48 +08:00
ovs_ct_fill_key(skb, key);
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
key->ovs_flow_hash = 0;
key->mac_proto = MAC_PROTO_ETHERNET;
openvswitch: Eliminate memset() from flow_extract. As new protocols are added, the size of the flow key tends to increase although few protocols care about all of the fields. In order to optimize this for hashing and matching, OVS uses a variable length portion of the key. However, when fields are extracted from the packet we must still zero out the entire key. This is no longer necessary now that OVS implements masking. Any fields (or holes in the structure) which are not part of a given protocol will be by definition not part of the mask and zeroed out during lookup. Furthermore, since masking already uses variable length keys this zeroing operation automatically benefits as well. In principle, the only thing that needs to be done at this point is remove the memset() at the beginning of flow. However, some fields assume that they are initialized to zero, which now must be done explicitly. In addition, in the event of an error we must also zero out corresponding fields to signal that there is no valid data present. These increase the total amount of code but very little of it is executed in non-error situations. Removing the memset() reduces the profile of ovs_flow_extract() from 0.64% to 0.56% when tested with large packets on a 10G link. Suggested-by: Pravin Shelar <pshelar@nicira.com> Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: Andy Zhou <azhou@nicira.com> Acked-by: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-04 06:35:29 +08:00
key->recirc_id = 0;
return key_extract(skb, key);
}
int ovs_flow_key_extract_userspace(struct net *net, const struct nlattr *attr,
struct sk_buff *skb,
struct sw_flow_key *key, bool log)
{
int err;
/* Extract metadata from netlink attributes. */
err = ovs_nla_get_flow_metadata(net, attr, key, log);
if (err)
return err;
return key_extract(skb, key);
}