2011-10-26 04:26:31 +02:00
|
|
|
/*
|
2014-09-16 04:37:25 +02:00
|
|
|
* Copyright (c) 2007-2014 Nicira, Inc.
|
2011-10-26 04:26:31 +02:00
|
|
|
*
|
|
|
|
* 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/if_arp.h>
|
|
|
|
#include <linux/ip.h>
|
|
|
|
#include <linux/ipv6.h>
|
2013-08-22 21:30:48 +02:00
|
|
|
#include <linux/sctp.h>
|
2013-10-30 01:22:21 +01:00
|
|
|
#include <linux/smp.h>
|
2011-10-26 04:26:31 +02:00
|
|
|
#include <linux/tcp.h>
|
|
|
|
#include <linux/udp.h>
|
|
|
|
#include <linux/icmp.h>
|
|
|
|
#include <linux/icmpv6.h>
|
|
|
|
#include <linux/rculist.h>
|
|
|
|
#include <net/ip.h>
|
2013-06-18 02:50:18 +02:00
|
|
|
#include <net/ip_tunnels.h>
|
2011-10-26 04:26:31 +02:00
|
|
|
#include <net/ipv6.h>
|
|
|
|
#include <net/ndisc.h>
|
|
|
|
|
2014-09-16 04:20:31 +02:00
|
|
|
#include "datapath.h"
|
|
|
|
#include "flow.h"
|
|
|
|
#include "flow_netlink.h"
|
|
|
|
|
2013-10-04 03:16:47 +02:00
|
|
|
u64 ovs_flow_used_time(unsigned long flow_jiffies)
|
2013-08-08 05:01:00 +02:00
|
|
|
{
|
2013-10-04 03:16:47 +02:00
|
|
|
struct timespec cur_ts;
|
|
|
|
u64 cur_ms, idle_ms;
|
2013-08-08 05:01:00 +02:00
|
|
|
|
2013-10-04 03:16:47 +02:00
|
|
|
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;
|
2013-08-08 05:01:00 +02:00
|
|
|
|
2013-10-04 03:16:47 +02:00
|
|
|
return cur_ms - idle_ms;
|
2013-08-27 22:02:21 +02:00
|
|
|
}
|
|
|
|
|
2013-10-23 10:40:44 +02:00
|
|
|
#define TCP_FLAGS_BE16(tp) (*(__be16 *)&tcp_flag_word(tp) & htons(0x0FFF))
|
2013-08-08 05:01:00 +02:00
|
|
|
|
2014-05-07 01:48:38 +02:00
|
|
|
void ovs_flow_stats_update(struct sw_flow *flow, __be16 tcp_flags,
|
|
|
|
struct sk_buff *skb)
|
2013-08-08 05:01:00 +02:00
|
|
|
{
|
2013-10-30 01:22:21 +01:00
|
|
|
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>
2014-03-27 20:42:54 +01:00
|
|
|
int node = numa_node_id();
|
2013-08-08 05:01:00 +02:00
|
|
|
|
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>
2014-03-27 20:42:54 +01:00
|
|
|
stats = rcu_dereference(flow->stats[node]);
|
2013-10-30 01:22:21 +01:00
|
|
|
|
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>
2014-03-27 20:42:54 +01:00
|
|
|
/* Check if already have node-specific stats. */
|
|
|
|
if (likely(stats)) {
|
|
|
|
spin_lock(&stats->lock);
|
|
|
|
/* Mark if we write on the pre-allocated stats. */
|
|
|
|
if (node == 0 && unlikely(flow->stats_last_writer != node))
|
|
|
|
flow->stats_last_writer = node;
|
|
|
|
} else {
|
|
|
|
stats = rcu_dereference(flow->stats[0]); /* Pre-allocated. */
|
|
|
|
spin_lock(&stats->lock);
|
|
|
|
|
|
|
|
/* If the current NUMA-node is the only writer on the
|
|
|
|
* pre-allocated stats keep using them.
|
|
|
|
*/
|
|
|
|
if (unlikely(flow->stats_last_writer != node)) {
|
|
|
|
/* A previous locker may have already allocated the
|
|
|
|
* stats, so we need to check again. If node-specific
|
|
|
|
* stats were already allocated, we update the pre-
|
|
|
|
* allocated stats as we have already locked them.
|
|
|
|
*/
|
|
|
|
if (likely(flow->stats_last_writer != NUMA_NO_NODE)
|
2014-08-17 15:29:43 +02:00
|
|
|
&& likely(!rcu_access_pointer(flow->stats[node]))) {
|
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>
2014-03-27 20:42:54 +01:00
|
|
|
/* Try to allocate node-specific stats. */
|
|
|
|
struct flow_stats *new_stats;
|
|
|
|
|
|
|
|
new_stats =
|
|
|
|
kmem_cache_alloc_node(flow_stats_cache,
|
|
|
|
GFP_THISNODE |
|
|
|
|
__GFP_NOMEMALLOC,
|
|
|
|
node);
|
|
|
|
if (likely(new_stats)) {
|
|
|
|
new_stats->used = jiffies;
|
|
|
|
new_stats->packet_count = 1;
|
|
|
|
new_stats->byte_count = skb->len;
|
|
|
|
new_stats->tcp_flags = tcp_flags;
|
|
|
|
spin_lock_init(&new_stats->lock);
|
|
|
|
|
|
|
|
rcu_assign_pointer(flow->stats[node],
|
|
|
|
new_stats);
|
|
|
|
goto unlock;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
flow->stats_last_writer = node;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-10-30 01:22:21 +01:00
|
|
|
stats->used = jiffies;
|
|
|
|
stats->packet_count++;
|
|
|
|
stats->byte_count += skb->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>
2014-03-27 20:42:54 +01:00
|
|
|
unlock:
|
2013-10-30 01:22:21 +01:00
|
|
|
spin_unlock(&stats->lock);
|
|
|
|
}
|
|
|
|
|
2014-05-05 23:17:28 +02:00
|
|
|
/* 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,
|
2013-10-30 01:22:21 +01:00
|
|
|
unsigned long *used, __be16 *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>
2014-03-27 20:42:54 +01:00
|
|
|
int node;
|
2013-10-30 01:22:21 +01:00
|
|
|
|
|
|
|
*used = 0;
|
|
|
|
*tcp_flags = 0;
|
|
|
|
memset(ovs_stats, 0, sizeof(*ovs_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>
2014-03-27 20:42:54 +01:00
|
|
|
for_each_node(node) {
|
2014-05-05 23:17:28 +02:00
|
|
|
struct flow_stats *stats = rcu_dereference_ovsl(flow->stats[node]);
|
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>
2014-03-27 20:42:54 +01:00
|
|
|
|
|
|
|
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);
|
|
|
|
}
|
2013-10-30 01:22:21 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-05-05 23:17:28 +02:00
|
|
|
/* Called with ovs_mutex. */
|
2013-10-30 01:22:21 +01:00
|
|
|
void ovs_flow_stats_clear(struct sw_flow *flow)
|
|
|
|
{
|
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>
2014-03-27 20:42:54 +01:00
|
|
|
int node;
|
|
|
|
|
|
|
|
for_each_node(node) {
|
2014-05-05 23:17:28 +02:00
|
|
|
struct flow_stats *stats = ovsl_dereference(flow->stats[node]);
|
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>
2014-03-27 20:42:54 +01:00
|
|
|
|
|
|
|
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);
|
|
|
|
}
|
|
|
|
}
|
2013-08-08 05:01:00 +02:00
|
|
|
}
|
|
|
|
|
2011-10-26 04:26:31 +02:00
|
|
|
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));
|
|
|
|
}
|
|
|
|
|
2013-08-22 21:30:48 +02:00
|
|
|
static bool sctphdr_ok(struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
return pskb_may_pull(skb, skb_transport_offset(skb) +
|
|
|
|
sizeof(struct sctphdr));
|
|
|
|
}
|
|
|
|
|
2011-10-26 04:26:31 +02:00
|
|
|
static bool icmphdr_ok(struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
return pskb_may_pull(skb, skb_transport_offset(skb) +
|
|
|
|
sizeof(struct icmphdr));
|
|
|
|
}
|
|
|
|
|
2013-08-08 05:01:00 +02:00
|
|
|
static int parse_ipv6hdr(struct sk_buff *skb, struct sw_flow_key *key)
|
2011-10-26 04:26:31 +02:00
|
|
|
{
|
|
|
|
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 (unlikely(payload_ofs < 0))
|
|
|
|
return -EINVAL;
|
|
|
|
|
|
|
|
if (frag_off) {
|
|
|
|
if (frag_off & htons(~0x7))
|
|
|
|
key->ip.frag = OVS_FRAG_TYPE_LATER;
|
|
|
|
else
|
|
|
|
key->ip.frag = OVS_FRAG_TYPE_FIRST;
|
2014-10-17 22:56:31 +02:00
|
|
|
} else {
|
|
|
|
key->ip.frag = OVS_FRAG_TYPE_NONE;
|
2011-10-26 04:26:31 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
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));
|
|
|
|
}
|
|
|
|
|
|
|
|
static int parse_vlan(struct sk_buff *skb, struct sw_flow_key *key)
|
|
|
|
{
|
|
|
|
struct qtag_prefix {
|
|
|
|
__be16 eth_type; /* ETH_P_8021Q */
|
|
|
|
__be16 tci;
|
|
|
|
};
|
|
|
|
struct qtag_prefix *qp;
|
|
|
|
|
|
|
|
if (unlikely(skb->len < sizeof(struct qtag_prefix) + sizeof(__be16)))
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
if (unlikely(!pskb_may_pull(skb, sizeof(struct qtag_prefix) +
|
|
|
|
sizeof(__be16))))
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
qp = (struct qtag_prefix *) skb->data;
|
|
|
|
key->eth.tci = qp->tci | htons(VLAN_TAG_PRESENT);
|
|
|
|
__skb_pull(skb, sizeof(struct qtag_prefix));
|
|
|
|
|
|
|
|
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));
|
|
|
|
|
2013-03-28 05:38:25 +01:00
|
|
|
if (ntohs(proto) >= ETH_P_802_3_MIN)
|
2011-10-26 04:26:31 +02:00
|
|
|
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));
|
2013-02-19 20:10:30 +01:00
|
|
|
|
2013-03-28 05:38:25 +01:00
|
|
|
if (ntohs(llc->ethertype) >= ETH_P_802_3_MIN)
|
2013-02-19 20:10:30 +01:00
|
|
|
return llc->ethertype;
|
|
|
|
|
|
|
|
return htons(ETH_P_802_2);
|
2011-10-26 04:26:31 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
static int parse_icmpv6(struct sk_buff *skb, struct sw_flow_key *key,
|
2013-08-08 05:01:00 +02:00
|
|
|
int nh_len)
|
2011-10-26 04:26:31 +02:00
|
|
|
{
|
|
|
|
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.
|
|
|
|
*/
|
2014-05-05 18:54:49 +02:00
|
|
|
key->tp.src = htons(icmp->icmp6_type);
|
|
|
|
key->tp.dst = htons(icmp->icmp6_code);
|
2014-10-17 22:56:31 +02:00
|
|
|
memset(&key->ipv6.nd, 0, sizeof(key->ipv6.nd));
|
2011-10-26 04:26:31 +02:00
|
|
|
|
|
|
|
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)))
|
2013-08-08 05:01:00 +02:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
if (unlikely(skb_linearize(skb)))
|
|
|
|
return -ENOMEM;
|
2011-10-26 04:26:31 +02:00
|
|
|
|
|
|
|
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))
|
2013-08-08 05:01:00 +02:00
|
|
|
return 0;
|
2011-10-26 04:26:31 +02:00
|
|
|
|
|
|
|
/* 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;
|
2014-02-18 20:15:45 +01:00
|
|
|
ether_addr_copy(key->ipv6.nd.sll,
|
|
|
|
&nd->opt[offset+sizeof(*nd_opt)]);
|
2011-10-26 04:26:31 +02:00
|
|
|
} 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;
|
2014-02-18 20:15:45 +01:00
|
|
|
ether_addr_copy(key->ipv6.nd.tll,
|
|
|
|
&nd->opt[offset+sizeof(*nd_opt)]);
|
2011-10-26 04:26:31 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
icmp_len -= opt_len;
|
|
|
|
offset += opt_len;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-08-08 05:01:00 +02:00
|
|
|
return 0;
|
2011-10-26 04:26:31 +02:00
|
|
|
|
|
|
|
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));
|
|
|
|
|
2013-08-08 05:01:00 +02:00
|
|
|
return 0;
|
2011-10-26 04:26:31 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
2014-09-16 04:20:31 +02:00
|
|
|
* key_extract - extracts a flow key from an Ethernet frame.
|
2011-10-26 04:26:31 +02:00
|
|
|
* @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.
|
|
|
|
*
|
2013-06-03 19:01:14 +02:00
|
|
|
* - skb->transport_header: If key->eth.type is ETH_P_IP or ETH_P_IPV6
|
2011-10-26 04:26:31 +02:00
|
|
|
* on output, then just past the IP header, if one is present and
|
|
|
|
* of a correct length, otherwise the same as skb->network_header.
|
2013-06-03 19:01:14 +02:00
|
|
|
* For other key->eth.type values it is left untouched.
|
2011-10-26 04:26:31 +02:00
|
|
|
*/
|
2014-09-16 04:20:31 +02:00
|
|
|
static int key_extract(struct sk_buff *skb, struct sw_flow_key *key)
|
2011-10-26 04:26:31 +02:00
|
|
|
{
|
2013-08-08 05:01:00 +02:00
|
|
|
int error;
|
2011-10-26 04:26:31 +02:00
|
|
|
struct ethhdr *eth;
|
|
|
|
|
2014-10-04 00:35:33 +02:00
|
|
|
/* Flags are always used as part of stats */
|
|
|
|
key->tp.flags = 0;
|
|
|
|
|
2011-10-26 04:26:31 +02:00
|
|
|
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);
|
2014-02-18 20:15:45 +01:00
|
|
|
ether_addr_copy(key->eth.src, eth->h_source);
|
|
|
|
ether_addr_copy(key->eth.dst, eth->h_dest);
|
2011-10-26 04:26:31 +02:00
|
|
|
|
|
|
|
__skb_pull(skb, 2 * ETH_ALEN);
|
2013-06-13 20:11:44 +02:00
|
|
|
/* We are going to push all headers that we pull, so no need to
|
|
|
|
* update skb->csum here.
|
|
|
|
*/
|
2011-10-26 04:26:31 +02:00
|
|
|
|
2014-10-04 00:35:29 +02:00
|
|
|
key->eth.tci = 0;
|
2011-10-26 04:26:31 +02:00
|
|
|
if (vlan_tx_tag_present(skb))
|
|
|
|
key->eth.tci = htons(skb->vlan_tci);
|
|
|
|
else if (eth->h_proto == htons(ETH_P_8021Q))
|
|
|
|
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_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)) {
|
2014-10-04 00:35:29 +02:00
|
|
|
memset(&key->ip, 0, sizeof(key->ip));
|
|
|
|
memset(&key->ipv4, 0, sizeof(key->ipv4));
|
2011-10-26 04:26:31 +02:00
|
|
|
if (error == -EINVAL) {
|
|
|
|
skb->transport_header = skb->network_header;
|
|
|
|
error = 0;
|
|
|
|
}
|
2013-08-08 05:01:00 +02:00
|
|
|
return error;
|
2011-10-26 04:26:31 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
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;
|
2013-08-08 05:01:00 +02:00
|
|
|
return 0;
|
2011-10-26 04:26:31 +02:00
|
|
|
}
|
|
|
|
if (nh->frag_off & htons(IP_MF) ||
|
2014-10-04 00:35:29 +02:00
|
|
|
skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
|
2011-10-26 04:26:31 +02:00
|
|
|
key->ip.frag = OVS_FRAG_TYPE_FIRST;
|
2014-10-04 00:35:29 +02:00
|
|
|
else
|
|
|
|
key->ip.frag = OVS_FRAG_TYPE_NONE;
|
2011-10-26 04:26:31 +02:00
|
|
|
|
|
|
|
/* Transport layer. */
|
|
|
|
if (key->ip.proto == IPPROTO_TCP) {
|
|
|
|
if (tcphdr_ok(skb)) {
|
|
|
|
struct tcphdr *tcp = tcp_hdr(skb);
|
2014-05-05 18:54:49 +02:00
|
|
|
key->tp.src = tcp->source;
|
|
|
|
key->tp.dst = tcp->dest;
|
|
|
|
key->tp.flags = TCP_FLAGS_BE16(tcp);
|
2014-10-04 00:35:29 +02:00
|
|
|
} else {
|
|
|
|
memset(&key->tp, 0, sizeof(key->tp));
|
2011-10-26 04:26:31 +02:00
|
|
|
}
|
2014-10-04 00:35:29 +02:00
|
|
|
|
2011-10-26 04:26:31 +02:00
|
|
|
} else if (key->ip.proto == IPPROTO_UDP) {
|
|
|
|
if (udphdr_ok(skb)) {
|
|
|
|
struct udphdr *udp = udp_hdr(skb);
|
2014-05-05 18:54:49 +02:00
|
|
|
key->tp.src = udp->source;
|
|
|
|
key->tp.dst = udp->dest;
|
2014-10-04 00:35:29 +02:00
|
|
|
} else {
|
|
|
|
memset(&key->tp, 0, sizeof(key->tp));
|
2011-10-26 04:26:31 +02:00
|
|
|
}
|
2013-08-22 21:30:48 +02:00
|
|
|
} else if (key->ip.proto == IPPROTO_SCTP) {
|
|
|
|
if (sctphdr_ok(skb)) {
|
|
|
|
struct sctphdr *sctp = sctp_hdr(skb);
|
2014-05-05 18:54:49 +02:00
|
|
|
key->tp.src = sctp->source;
|
|
|
|
key->tp.dst = sctp->dest;
|
2014-10-04 00:35:29 +02:00
|
|
|
} else {
|
|
|
|
memset(&key->tp, 0, sizeof(key->tp));
|
2013-08-22 21:30:48 +02:00
|
|
|
}
|
2011-10-26 04:26:31 +02:00
|
|
|
} 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. */
|
2014-05-05 18:54:49 +02:00
|
|
|
key->tp.src = htons(icmp->type);
|
|
|
|
key->tp.dst = htons(icmp->code);
|
2014-10-04 00:35:29 +02:00
|
|
|
} else {
|
|
|
|
memset(&key->tp, 0, sizeof(key->tp));
|
2011-10-26 04:26:31 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-10-04 00:35:29 +02:00
|
|
|
} else if (key->eth.type == htons(ETH_P_ARP) ||
|
|
|
|
key->eth.type == htons(ETH_P_RARP)) {
|
2011-10-26 04:26:31 +02:00
|
|
|
struct arp_eth_header *arp;
|
2014-10-17 08:03:08 +02:00
|
|
|
bool arp_available = arphdr_ok(skb);
|
2011-10-26 04:26:31 +02:00
|
|
|
|
|
|
|
arp = (struct arp_eth_header *)skb_network_header(skb);
|
|
|
|
|
2014-10-17 08:03:08 +02:00
|
|
|
if (arp_available &&
|
2014-10-04 00:35:29 +02:00
|
|
|
arp->ar_hrd == htons(ARPHRD_ETHER) &&
|
|
|
|
arp->ar_pro == htons(ETH_P_IP) &&
|
|
|
|
arp->ar_hln == ETH_ALEN &&
|
|
|
|
arp->ar_pln == 4) {
|
2011-10-26 04:26:31 +02:00
|
|
|
|
|
|
|
/* We only match on the lower 8 bits of the opcode. */
|
|
|
|
if (ntohs(arp->ar_op) <= 0xff)
|
|
|
|
key->ip.proto = ntohs(arp->ar_op);
|
2014-10-04 00:35:29 +02:00
|
|
|
else
|
|
|
|
key->ip.proto = 0;
|
|
|
|
|
2012-10-30 23:50:28 +01:00
|
|
|
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));
|
2014-02-18 20:15:45 +01:00
|
|
|
ether_addr_copy(key->ipv4.arp.sha, arp->ar_sha);
|
|
|
|
ether_addr_copy(key->ipv4.arp.tha, arp->ar_tha);
|
2014-10-04 00:35:29 +02:00
|
|
|
} else {
|
|
|
|
memset(&key->ip, 0, sizeof(key->ip));
|
|
|
|
memset(&key->ipv4, 0, sizeof(key->ipv4));
|
2011-10-26 04:26:31 +02:00
|
|
|
}
|
|
|
|
} else if (key->eth.type == htons(ETH_P_IPV6)) {
|
|
|
|
int nh_len; /* IPv6 Header + Extensions */
|
|
|
|
|
2013-08-08 05:01:00 +02:00
|
|
|
nh_len = parse_ipv6hdr(skb, key);
|
2011-10-26 04:26:31 +02:00
|
|
|
if (unlikely(nh_len < 0)) {
|
2014-10-04 00:35:29 +02:00
|
|
|
memset(&key->ip, 0, sizeof(key->ip));
|
|
|
|
memset(&key->ipv6.addr, 0, sizeof(key->ipv6.addr));
|
2013-08-08 05:01:00 +02:00
|
|
|
if (nh_len == -EINVAL) {
|
2011-10-26 04:26:31 +02:00
|
|
|
skb->transport_header = skb->network_header;
|
2013-08-08 05:01:00 +02:00
|
|
|
error = 0;
|
|
|
|
} else {
|
2011-10-26 04:26:31 +02:00
|
|
|
error = nh_len;
|
2013-08-08 05:01:00 +02:00
|
|
|
}
|
|
|
|
return error;
|
2011-10-26 04:26:31 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
if (key->ip.frag == OVS_FRAG_TYPE_LATER)
|
2013-08-08 05:01:00 +02:00
|
|
|
return 0;
|
2011-10-26 04:26:31 +02:00
|
|
|
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);
|
2014-05-05 18:54:49 +02:00
|
|
|
key->tp.src = tcp->source;
|
|
|
|
key->tp.dst = tcp->dest;
|
|
|
|
key->tp.flags = TCP_FLAGS_BE16(tcp);
|
2014-10-04 00:35:29 +02:00
|
|
|
} else {
|
|
|
|
memset(&key->tp, 0, sizeof(key->tp));
|
2011-10-26 04:26:31 +02:00
|
|
|
}
|
|
|
|
} else if (key->ip.proto == NEXTHDR_UDP) {
|
|
|
|
if (udphdr_ok(skb)) {
|
|
|
|
struct udphdr *udp = udp_hdr(skb);
|
2014-05-05 18:54:49 +02:00
|
|
|
key->tp.src = udp->source;
|
|
|
|
key->tp.dst = udp->dest;
|
2014-10-04 00:35:29 +02:00
|
|
|
} else {
|
|
|
|
memset(&key->tp, 0, sizeof(key->tp));
|
2011-10-26 04:26:31 +02:00
|
|
|
}
|
2013-08-22 21:30:48 +02:00
|
|
|
} else if (key->ip.proto == NEXTHDR_SCTP) {
|
|
|
|
if (sctphdr_ok(skb)) {
|
|
|
|
struct sctphdr *sctp = sctp_hdr(skb);
|
2014-05-05 18:54:49 +02:00
|
|
|
key->tp.src = sctp->source;
|
|
|
|
key->tp.dst = sctp->dest;
|
2014-10-04 00:35:29 +02:00
|
|
|
} else {
|
|
|
|
memset(&key->tp, 0, sizeof(key->tp));
|
2013-08-22 21:30:48 +02:00
|
|
|
}
|
2011-10-26 04:26:31 +02:00
|
|
|
} else if (key->ip.proto == NEXTHDR_ICMP) {
|
|
|
|
if (icmp6hdr_ok(skb)) {
|
2013-08-08 05:01:00 +02:00
|
|
|
error = parse_icmpv6(skb, key, nh_len);
|
|
|
|
if (error)
|
|
|
|
return error;
|
2014-10-04 00:35:29 +02:00
|
|
|
} else {
|
|
|
|
memset(&key->tp, 0, sizeof(key->tp));
|
2011-10-26 04:26:31 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2013-08-08 05:01:00 +02:00
|
|
|
return 0;
|
2011-10-26 04:26:31 +02:00
|
|
|
}
|
2014-09-16 04:20:31 +02:00
|
|
|
|
2014-09-16 04:37:25 +02:00
|
|
|
int ovs_flow_key_update(struct sk_buff *skb, struct sw_flow_key *key)
|
|
|
|
{
|
|
|
|
return key_extract(skb, key);
|
|
|
|
}
|
|
|
|
|
2014-10-04 00:35:31 +02:00
|
|
|
int ovs_flow_key_extract(struct ovs_tunnel_info *tun_info,
|
2014-09-16 04:28:44 +02:00
|
|
|
struct sk_buff *skb, struct sw_flow_key *key)
|
2014-09-16 04:20:31 +02:00
|
|
|
{
|
|
|
|
/* Extract metadata from packet. */
|
2014-10-04 00:35:33 +02:00
|
|
|
if (tun_info) {
|
2014-10-04 00:35:31 +02:00
|
|
|
memcpy(&key->tun_key, &tun_info->tunnel, sizeof(key->tun_key));
|
2014-10-04 00:35:33 +02:00
|
|
|
|
|
|
|
if (tun_info->options) {
|
|
|
|
BUILD_BUG_ON((1 << (sizeof(tun_info->options_len) *
|
|
|
|
8)) - 1
|
|
|
|
> sizeof(key->tun_opts));
|
|
|
|
memcpy(GENEVE_OPTS(key, tun_info->options_len),
|
|
|
|
tun_info->options, tun_info->options_len);
|
|
|
|
key->tun_opts_len = tun_info->options_len;
|
|
|
|
} else {
|
|
|
|
key->tun_opts_len = 0;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
key->tun_opts_len = 0;
|
2014-10-04 00:35:29 +02:00
|
|
|
memset(&key->tun_key, 0, sizeof(key->tun_key));
|
2014-10-04 00:35:33 +02:00
|
|
|
}
|
2014-09-16 04:20:31 +02:00
|
|
|
|
|
|
|
key->phy.priority = skb->priority;
|
|
|
|
key->phy.in_port = OVS_CB(skb)->input_vport->port_no;
|
|
|
|
key->phy.skb_mark = skb->mark;
|
2014-10-04 00:35:29 +02:00
|
|
|
key->ovs_flow_hash = 0;
|
|
|
|
key->recirc_id = 0;
|
|
|
|
|
2014-09-16 04:20:31 +02:00
|
|
|
return key_extract(skb, key);
|
|
|
|
}
|
|
|
|
|
|
|
|
int ovs_flow_key_extract_userspace(const struct nlattr *attr,
|
|
|
|
struct sk_buff *skb,
|
|
|
|
struct sw_flow_key *key)
|
|
|
|
{
|
|
|
|
int err;
|
|
|
|
|
|
|
|
/* Extract metadata from netlink attributes. */
|
|
|
|
err = ovs_nla_get_flow_metadata(attr, key);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
|
|
|
|
return key_extract(skb, key);
|
|
|
|
}
|