[PATCH] net: sched: check tc_skip_classify as far as possible

Subsystems: networking [general], tc subsystem, the rest

STALE1793d

4 messages, 2 authors, 2021-10-29 · open the first message on its own page

[PATCH] net: sched: check tc_skip_classify as far as possible

From: <hidden>
Date: 2021-10-28 13:56:55

From: Tonghao Zhang <redacted>

We look up and then check tc_skip_classify flag in net
sched layer, even though skb don't want to be classified.
That case may consume a lot of cpu cycles.

Install the rules as below:
$ for id in $(seq 1 100); do
$ 	tc filter add ... egress prio $id ... action mirred egress redirect dev ifb0
$ done

netperf:
$ taskset -c 1 netperf -t TCP_RR -H ip -- -r 32,32
$ taskset -c 1 netperf -t TCP_STREAM -H ip -- -m 32

Without this patch:
10662.33 tps
108.95 Mbit/s

With this patch:
12434.48 tps
145.89 Mbit/s

For TCP_RR, there are 16.6% improvement, TCP_STREAM 33.9%.

Cc: Willem de Bruijn <willemb@google.com>
Cc: Cong Wang <redacted>
Cc: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Tonghao Zhang <redacted>
---
 net/core/dev.c      | 3 ++-
 net/sched/act_api.c | 3 ---
 2 files changed, 2 insertions(+), 4 deletions(-)
diff --git a/net/core/dev.c b/net/core/dev.c
index eb61a8821b3a..856ac1fb75b4 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -4155,7 +4155,8 @@ static int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
 #ifdef CONFIG_NET_CLS_ACT
 	skb->tc_at_ingress = 0;
 # ifdef CONFIG_NET_EGRESS
-	if (static_branch_unlikely(&egress_needed_key)) {
+	if (static_branch_unlikely(&egress_needed_key) &&
+	    !skb_skip_tc_classify(skb)) {
 		skb = sch_handle_egress(skb, &rc, dev);
 		if (!skb)
 			goto out;
diff --git a/net/sched/act_api.c b/net/sched/act_api.c
index 7dd3a2dc5fa4..bd66f27178be 100644
--- a/net/sched/act_api.c
+++ b/net/sched/act_api.c
@@ -722,9 +722,6 @@ int tcf_action_exec(struct sk_buff *skb, struct tc_action **actions,
 	int i;
 	int ret = TC_ACT_OK;
 
-	if (skb_skip_tc_classify(skb))
-		return TC_ACT_OK;
-
 restart_act_graph:
 	for (i = 0; i < nr_actions; i++) {
 		const struct tc_action *a = actions[i];
-- 
2.27.0

Re: [PATCH] net: sched: check tc_skip_classify as far as possible

From: Daniel Borkmann <daniel@iogearbox.net>
Date: 2021-10-28 14:28:23

On 10/28/21 3:56 PM, xiangxia.m.yue@gmail.com wrote:
From: Tonghao Zhang <redacted>

We look up and then check tc_skip_classify flag in net
sched layer, even though skb don't want to be classified.
That case may consume a lot of cpu cycles.

Install the rules as below:
$ for id in $(seq 1 100); do
$ 	tc filter add ... egress prio $id ... action mirred egress redirect dev ifb0
$ done
Do you actually have such a case in practice or is this just hypothetical?
Asking as this feels rather broken to begin with.
quoted hunk
netperf:
$ taskset -c 1 netperf -t TCP_RR -H ip -- -r 32,32
$ taskset -c 1 netperf -t TCP_STREAM -H ip -- -m 32

Without this patch:
10662.33 tps
108.95 Mbit/s

With this patch:
12434.48 tps
145.89 Mbit/s

For TCP_RR, there are 16.6% improvement, TCP_STREAM 33.9%.

Cc: Willem de Bruijn <willemb@google.com>
Cc: Cong Wang <redacted>
Cc: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Tonghao Zhang <redacted>
---
  net/core/dev.c      | 3 ++-
  net/sched/act_api.c | 3 ---
  2 files changed, 2 insertions(+), 4 deletions(-)
diff --git a/net/core/dev.c b/net/core/dev.c
index eb61a8821b3a..856ac1fb75b4 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -4155,7 +4155,8 @@ static int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
  #ifdef CONFIG_NET_CLS_ACT
  	skb->tc_at_ingress = 0;
  # ifdef CONFIG_NET_EGRESS
-	if (static_branch_unlikely(&egress_needed_key)) {
+	if (static_branch_unlikely(&egress_needed_key) &&
+	    !skb_skip_tc_classify(skb)) {
  		skb = sch_handle_egress(skb, &rc, dev);
  		if (!skb)
  			goto out;
diff --git a/net/sched/act_api.c b/net/sched/act_api.c
index 7dd3a2dc5fa4..bd66f27178be 100644
--- a/net/sched/act_api.c
+++ b/net/sched/act_api.c
@@ -722,9 +722,6 @@ int tcf_action_exec(struct sk_buff *skb, struct tc_action **actions,
  	int i;
  	int ret = TC_ACT_OK;
  
-	if (skb_skip_tc_classify(skb))
-		return TC_ACT_OK;
-
I think this might imply a change in behavior which could have the potential
to break setups in the wild.
  restart_act_graph:
  	for (i = 0; i < nr_actions; i++) {
  		const struct tc_action *a = actions[i];

Re: [PATCH] net: sched: check tc_skip_classify as far as possible

From: Tonghao Zhang <hidden>
Date: 2021-10-29 00:04:53

On Thu, Oct 28, 2021 at 10:28 PM Daniel Borkmann [off-list ref] wrote:
On 10/28/21 3:56 PM, xiangxia.m.yue@gmail.com wrote:
quoted
From: Tonghao Zhang <redacted>

We look up and then check tc_skip_classify flag in net
sched layer, even though skb don't want to be classified.
That case may consume a lot of cpu cycles.

Install the rules as below:
$ for id in $(seq 1 100); do
$     tc filter add ... egress prio $id ... action mirred egress redirect dev ifb0
$ done
Do you actually have such a case in practice or is this just hypothetical?
Hi Daniel, I did some research about this for k8s in production. There
are not so many tc prio(~5 different prio).
butg in this test, I use the 100 prio.

I reviewed the code, for the tx path, I think we check the
tc_skip_classify too later. In the rx path, we check it
in __netif_receive_skb_core.
Asking as this feels rather broken to begin with.
quoted
netperf:
$ taskset -c 1 netperf -t TCP_RR -H ip -- -r 32,32
$ taskset -c 1 netperf -t TCP_STREAM -H ip -- -m 32

Without this patch:
10662.33 tps
108.95 Mbit/s

With this patch:
12434.48 tps
145.89 Mbit/s

For TCP_RR, there are 16.6% improvement, TCP_STREAM 33.9%.

Cc: Willem de Bruijn <willemb@google.com>
Cc: Cong Wang <redacted>
Cc: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Tonghao Zhang <redacted>
---
  net/core/dev.c      | 3 ++-
  net/sched/act_api.c | 3 ---
  2 files changed, 2 insertions(+), 4 deletions(-)
diff --git a/net/core/dev.c b/net/core/dev.c
index eb61a8821b3a..856ac1fb75b4 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -4155,7 +4155,8 @@ static int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
  #ifdef CONFIG_NET_CLS_ACT
      skb->tc_at_ingress = 0;
  # ifdef CONFIG_NET_EGRESS
-     if (static_branch_unlikely(&egress_needed_key)) {
+     if (static_branch_unlikely(&egress_needed_key) &&
+         !skb_skip_tc_classify(skb)) {
              skb = sch_handle_egress(skb, &rc, dev);
              if (!skb)
                      goto out;
diff --git a/net/sched/act_api.c b/net/sched/act_api.c
index 7dd3a2dc5fa4..bd66f27178be 100644
--- a/net/sched/act_api.c
+++ b/net/sched/act_api.c
@@ -722,9 +722,6 @@ int tcf_action_exec(struct sk_buff *skb, struct tc_action **actions,
      int i;
      int ret = TC_ACT_OK;

-     if (skb_skip_tc_classify(skb))
-             return TC_ACT_OK;
-
I think this might imply a change in behavior which could have the potential
to break setups in the wild.
we may not change this code, i will send v2, if not comment.
quoted
  restart_act_graph:
      for (i = 0; i < nr_actions; i++) {
              const struct tc_action *a = actions[i];

-- 
Best regards, Tonghao

Re: [PATCH] net: sched: check tc_skip_classify as far as possible

From: Daniel Borkmann <daniel@iogearbox.net>
Date: 2021-10-29 15:35:13

On 10/29/21 2:04 AM, Tonghao Zhang wrote:
On Thu, Oct 28, 2021 at 10:28 PM Daniel Borkmann [off-list ref] wrote:
quoted
On 10/28/21 3:56 PM, xiangxia.m.yue@gmail.com wrote:
quoted
From: Tonghao Zhang <redacted>

We look up and then check tc_skip_classify flag in net
sched layer, even though skb don't want to be classified.
That case may consume a lot of cpu cycles.

Install the rules as below:
$ for id in $(seq 1 100); do
$     tc filter add ... egress prio $id ... action mirred egress redirect dev ifb0
$ done
Do you actually have such a case in practice or is this just hypothetical?
Hi Daniel, I did some research about this for k8s in production. There
are not so many tc prio(~5 different prio).
butg in this test, I use the 100 prio.

I reviewed the code, for the tx path, I think we check the
tc_skip_classify too later. In the rx path, we check it
in __netif_receive_skb_core.
quoted
Asking as this feels rather broken to begin with.
quoted
netperf:
$ taskset -c 1 netperf -t TCP_RR -H ip -- -r 32,32
$ taskset -c 1 netperf -t TCP_STREAM -H ip -- -m 32

Without this patch:
10662.33 tps
108.95 Mbit/s

With this patch:
12434.48 tps
145.89 Mbit/s

For TCP_RR, there are 16.6% improvement, TCP_STREAM 33.9%.

Cc: Willem de Bruijn <willemb@google.com>
Cc: Cong Wang <redacted>
Cc: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Tonghao Zhang <redacted>
---
   net/core/dev.c      | 3 ++-
   net/sched/act_api.c | 3 ---
   2 files changed, 2 insertions(+), 4 deletions(-)
diff --git a/net/core/dev.c b/net/core/dev.c
index eb61a8821b3a..856ac1fb75b4 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -4155,7 +4155,8 @@ static int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
   #ifdef CONFIG_NET_CLS_ACT
       skb->tc_at_ingress = 0;
   # ifdef CONFIG_NET_EGRESS
-     if (static_branch_unlikely(&egress_needed_key)) {
+     if (static_branch_unlikely(&egress_needed_key) &&
+         !skb_skip_tc_classify(skb)) {
               skb = sch_handle_egress(skb, &rc, dev);
               if (!skb)
                       goto out;
diff --git a/net/sched/act_api.c b/net/sched/act_api.c
index 7dd3a2dc5fa4..bd66f27178be 100644
--- a/net/sched/act_api.c
+++ b/net/sched/act_api.c
@@ -722,9 +722,6 @@ int tcf_action_exec(struct sk_buff *skb, struct tc_action **actions,
       int i;
       int ret = TC_ACT_OK;

-     if (skb_skip_tc_classify(skb))
-             return TC_ACT_OK;
-
I think this might imply a change in behavior which could have the potential
to break setups in the wild.
we may not change this code, i will send v2, if not comment.
Well none of it I'm afraid, the sch_handle_egress() is out for a very long time by
now and your change could have the potential to break setups in the wild.
quoted
quoted
   restart_act_graph:
       for (i = 0; i < nr_actions; i++) {
               const struct tc_action *a = actions[i];

Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help