Re: Reducing NET_LOCK() contention
On 2.8.2017. 11:00, Alexandr Nedvedicky wrote: > Hello, > > On Wed, Aug 02, 2017 at 10:10:51AM +0200, Martin Pieuchot wrote: >> On 18/07/17(Tue) 15:55, Martin Pieuchot wrote: >>> When forwarding a lot of traffic with 10G interfaces contention on the >>> NET_LOCK() is "visible". Each time you type "ifconfig" you can go grab >>> a coffee... >>> >>> The problem has a name: pf_purge_thread(). This thread is created by >>> default and run even if you don't have PF enabled. Every `hz' it wakes >>> up, grab the lock and go to sleep. >>> >>> Since the execution of this thread is serialized with the `softnet' task, >>> it makes more sense to execute it in the same context. This reduce the >>> NET_LOCK() contention and implicitly preempt the packet processing. >>> >>> Diff below improves the situation with PF disabled, I didn't test with >>> PF enabled. >> >> Updated diff that includes sashan@ suggestions and do not stop the purge >> task when PF is disabled. Otherwise some states are not purged until PF >> is re-enabled. This can be optimized later. >> > > thank Hrvoje for spotting the glitch. i should have spotted it with the first diff, but then i haven't disabled pf while generating traffic ...
Re: Reducing NET_LOCK() contention
On 2.8.2017. 10:10, Martin Pieuchot wrote: > On 18/07/17(Tue) 15:55, Martin Pieuchot wrote: >> When forwarding a lot of traffic with 10G interfaces contention on the >> NET_LOCK() is "visible". Each time you type "ifconfig" you can go grab >> a coffee... >> >> The problem has a name: pf_purge_thread(). This thread is created by >> default and run even if you don't have PF enabled. Every `hz' it wakes >> up, grab the lock and go to sleep. >> >> Since the execution of this thread is serialized with the `softnet' task, >> it makes more sense to execute it in the same context. This reduce the >> NET_LOCK() contention and implicitly preempt the packet processing. >> >> Diff below improves the situation with PF disabled, I didn't test with >> PF enabled. > Updated diff that includes sashan@ suggestions and do not stop the purge > task when PF is disabled. Otherwise some states are not purged until PF > is re-enabled. This can be optimized later. Hi all, with this diff states are purged as expected and ifconfig, netstat and arp outputs are little faster then before.
Re: Reducing NET_LOCK() contention
Hello, On Wed, Aug 02, 2017 at 10:10:51AM +0200, Martin Pieuchot wrote: > On 18/07/17(Tue) 15:55, Martin Pieuchot wrote: > > When forwarding a lot of traffic with 10G interfaces contention on the > > NET_LOCK() is "visible". Each time you type "ifconfig" you can go grab > > a coffee... > > > > The problem has a name: pf_purge_thread(). This thread is created by > > default and run even if you don't have PF enabled. Every `hz' it wakes > > up, grab the lock and go to sleep. > > > > Since the execution of this thread is serialized with the `softnet' task, > > it makes more sense to execute it in the same context. This reduce the > > NET_LOCK() contention and implicitly preempt the packet processing. > > > > Diff below improves the situation with PF disabled, I didn't test with > > PF enabled. > > Updated diff that includes sashan@ suggestions and do not stop the purge > task when PF is disabled. Otherwise some states are not purged until PF > is re-enabled. This can be optimized later. > thank Hrvoje for spotting the glitch. change looks good (actually better) to me. OK sashan@
Re: Reducing NET_LOCK() contention
On 18/07/17(Tue) 15:55, Martin Pieuchot wrote:
> When forwarding a lot of traffic with 10G interfaces contention on the
> NET_LOCK() is "visible". Each time you type "ifconfig" you can go grab
> a coffee...
>
> The problem has a name: pf_purge_thread(). This thread is created by
> default and run even if you don't have PF enabled. Every `hz' it wakes
> up, grab the lock and go to sleep.
>
> Since the execution of this thread is serialized with the `softnet' task,
> it makes more sense to execute it in the same context. This reduce the
> NET_LOCK() contention and implicitly preempt the packet processing.
>
> Diff below improves the situation with PF disabled, I didn't test with
> PF enabled.
Updated diff that includes sashan@ suggestions and do not stop the purge
task when PF is disabled. Otherwise some states are not purged until PF
is re-enabled. This can be optimized later.
ok?
Index: net/pf.c
===
RCS file: /cvs/src/sys/net/pf.c,v
retrieving revision 1.1037
diff -u -p -r1.1037 pf.c
--- net/pf.c4 Jul 2017 14:10:15 - 1.1037
+++ net/pf.c2 Aug 2017 08:08:47 -
@@ -121,6 +121,10 @@ u_char pf_tcp_secret[16];
int pf_tcp_secret_init;
int pf_tcp_iss_off;
+int pf_npurge;
+struct task pf_purge_task = TASK_INITIALIZER(pf_purge, _npurge);
+struct timeout pf_purge_to = TIMEOUT_INITIALIZER(pf_purge_timeout, NULL);
+
enum pf_test_status {
PF_TEST_FAIL = -1,
PF_TEST_OK,
@@ -1200,36 +1204,43 @@ pf_purge_expired_rules(void)
}
void
-pf_purge_thread(void *v)
+pf_purge_timeout(void *unused)
{
- int nloops = 0, s;
+ task_add(softnettq, _purge_task);
+}
- for (;;) {
- tsleep(pf_purge_thread, PWAIT, "pftm", 1 * hz);
+void
+pf_purge(void *xnloops)
+{
+ int *nloops = xnloops;
+ int s;
- NET_LOCK(s);
+ KERNEL_LOCK();
+ NET_LOCK(s);
- PF_LOCK();
- /* process a fraction of the state table every second */
- pf_purge_expired_states(1 + (pf_status.states
- / pf_default_rule.timeout[PFTM_INTERVAL]));
+ PF_LOCK();
+ /* process a fraction of the state table every second */
+ pf_purge_expired_states(1 + (pf_status.states
+ / pf_default_rule.timeout[PFTM_INTERVAL]));
- /* purge other expired types every PFTM_INTERVAL seconds */
- if (++nloops >= pf_default_rule.timeout[PFTM_INTERVAL]) {
- pf_purge_expired_src_nodes(0);
- pf_purge_expired_rules();
- }
- PF_UNLOCK();
- /*
-* Fragments don't require PF_LOCK(), they use their own lock.
-*/
- if (nloops >= pf_default_rule.timeout[PFTM_INTERVAL]) {
- pf_purge_expired_fragments();
- nloops = 0;
- }
+ /* purge other expired types every PFTM_INTERVAL seconds */
+ if (++(*nloops) >= pf_default_rule.timeout[PFTM_INTERVAL]) {
+ pf_purge_expired_src_nodes(0);
+ pf_purge_expired_rules();
+ }
+ PF_UNLOCK();
- NET_UNLOCK(s);
+ /*
+* Fragments don't require PF_LOCK(), they use their own lock.
+*/
+ if ((*nloops) >= pf_default_rule.timeout[PFTM_INTERVAL]) {
+ pf_purge_expired_fragments();
+ *nloops = 0;
}
+ NET_UNLOCK(s);
+ KERNEL_UNLOCK();
+
+ timeout_add(_purge_to, 1 * hz);
}
int32_t
Index: net/pf_ioctl.c
===
RCS file: /cvs/src/sys/net/pf_ioctl.c,v
retrieving revision 1.320
diff -u -p -r1.320 pf_ioctl.c
--- net/pf_ioctl.c 27 Jul 2017 12:09:51 - 1.320
+++ net/pf_ioctl.c 2 Aug 2017 08:01:53 -
@@ -231,16 +231,6 @@ pfattach(int num)
/* XXX do our best to avoid a conflict */
pf_status.hostid = arc4random();
-
- /* require process context to purge states, so perform in a thread */
- kthread_create_deferred(pf_thread_create, NULL);
-}
-
-void
-pf_thread_create(void *v)
-{
- if (kthread_create(pf_purge_thread, NULL, NULL, "pfpurge"))
- panic("pfpurge thread");
}
int
@@ -1027,6 +1017,7 @@ pfioctl(dev_t dev, u_long cmd, caddr_t a
pf_status.stateid = time_second;
pf_status.stateid = pf_status.stateid << 32;
}
+ timeout_add(_purge_to, 1 * hz);
pf_create_queues();
DPFPRINTF(LOG_NOTICE, "pf: started");
}
@@ -2320,7 +2311,7 @@ pfioctl(dev_t dev, u_long cmd, caddr_t a
pf_default_rule_new.timeout[i];
if (pf_default_rule.timeout[i] ==
Reducing NET_LOCK() contention
When forwarding a lot of traffic with 10G interfaces contention on the
NET_LOCK() is "visible". Each time you type "ifconfig" you can go grab
a coffee...
The problem has a name: pf_purge_thread(). This thread is created by
default and run even if you don't have PF enabled. Every `hz' it wakes
up, grab the lock and go to sleep.
Since the execution of this thread is serialized with the `softnet' task,
it makes more sense to execute it in the same context. This reduce the
NET_LOCK() contention and implicitly preempt the packet processing.
Diff below improves the situation with PF disabled, I didn't test with
PF enabled.
Index: net/pf.c
===
RCS file: /cvs/src/sys/net/pf.c,v
retrieving revision 1.1037
diff -u -p -r1.1037 pf.c
--- net/pf.c4 Jul 2017 14:10:15 - 1.1037
+++ net/pf.c18 Jul 2017 13:28:10 -
@@ -121,6 +121,10 @@ u_char pf_tcp_secret[16];
int pf_tcp_secret_init;
int pf_tcp_iss_off;
+int pf_npurge;
+struct task pf_purge_task = TASK_INITIALIZER(pf_purge, _npurge);
+struct timeout pf_purge_to = TIMEOUT_INITIALIZER(pf_purge_timeout, NULL);
+
enum pf_test_status {
PF_TEST_FAIL = -1,
PF_TEST_OK,
@@ -1200,36 +1204,44 @@ pf_purge_expired_rules(void)
}
void
-pf_purge_thread(void *v)
+pf_purge_timeout(void *unused)
{
- int nloops = 0, s;
-
- for (;;) {
- tsleep(pf_purge_thread, PWAIT, "pftm", 1 * hz);
-
- NET_LOCK(s);
+ task_add(softnettq, _purge_task);
+}
- PF_LOCK();
- /* process a fraction of the state table every second */
- pf_purge_expired_states(1 + (pf_status.states
- / pf_default_rule.timeout[PFTM_INTERVAL]));
+void
+pf_purge(void *xnloops)
+{
+ int *nloops = xnloops;
+ int s;
- /* purge other expired types every PFTM_INTERVAL seconds */
- if (++nloops >= pf_default_rule.timeout[PFTM_INTERVAL]) {
- pf_purge_expired_src_nodes(0);
- pf_purge_expired_rules();
- }
- PF_UNLOCK();
- /*
-* Fragments don't require PF_LOCK(), they use their own lock.
-*/
- if (nloops >= pf_default_rule.timeout[PFTM_INTERVAL]) {
- pf_purge_expired_fragments();
- nloops = 0;
- }
+ KERNEL_LOCK();
+ NET_LOCK(s);
- NET_UNLOCK(s);
+ PF_LOCK();
+ /* process a fraction of the state table every second */
+ pf_purge_expired_states(1 + (pf_status.states
+ / pf_default_rule.timeout[PFTM_INTERVAL]));
+
+ /* purge other expired types every PFTM_INTERVAL seconds */
+ if (++(*nloops) >= pf_default_rule.timeout[PFTM_INTERVAL]) {
+ pf_purge_expired_src_nodes(0);
+ pf_purge_expired_rules();
+ }
+ PF_UNLOCK();
+
+ /*
+* Fragments don't require PF_LOCK(), they use their own lock.
+*/
+ if ((*nloops) >= pf_default_rule.timeout[PFTM_INTERVAL]) {
+ pf_purge_expired_fragments();
+ *nloops = 0;
}
+ NET_UNLOCK(s);
+ KERNEL_UNLOCK();
+
+ if (pf_status.running)
+ timeout_add(_purge_to, 1 * hz);
}
int32_t
Index: net/pf_ioctl.c
===
RCS file: /cvs/src/sys/net/pf_ioctl.c,v
retrieving revision 1.318
diff -u -p -r1.318 pf_ioctl.c
--- net/pf_ioctl.c 5 Jul 2017 11:40:17 - 1.318
+++ net/pf_ioctl.c 18 Jul 2017 13:29:39 -
@@ -231,16 +231,6 @@ pfattach(int num)
/* XXX do our best to avoid a conflict */
pf_status.hostid = arc4random();
-
- /* require process context to purge states, so perform in a thread */
- kthread_create_deferred(pf_thread_create, NULL);
-}
-
-void
-pf_thread_create(void *v)
-{
- if (kthread_create(pf_purge_thread, NULL, NULL, "pfpurge"))
- panic("pfpurge thread");
}
int
@@ -1027,6 +1017,7 @@ pfioctl(dev_t dev, u_long cmd, caddr_t a
pf_status.stateid = time_second;
pf_status.stateid = pf_status.stateid << 32;
}
+ timeout_add(_purge_to, 1 * hz);
pf_create_queues();
DPFPRINTF(LOG_NOTICE, "pf: started");
}
@@ -2291,7 +2282,7 @@ pfioctl(dev_t dev, u_long cmd, caddr_t a
pf_default_rule_new.timeout[i];
if (pf_default_rule.timeout[i] == PFTM_INTERVAL &&
pf_default_rule.timeout[i] < old)
- wakeup(pf_purge_thread);
+ task_add(softnettq, _purge_task);
}
