Re: [PATCH] make net_gso_ok return false when gso_type is zero(invalid)
> Note that TCP stack now works with GSO being always on.
> 0a6b2a1dc2a2 ("tcp: switch to GSO being always on")
I've tested on the latest net-next branch
17dec0a949153d9ac00760ba2f5b78cb583e995f. The problem still exists. My
patch won't work. Reverting commit 0a6b2a1dc2a2 won't help.
Re: [PATCH] make net_gso_ok return false when gso_type is zero(invalid)
2018-04-10 18:32 GMT+02:00 Marcelo Ricardo Leitner :
> On Sun, Apr 08, 2018 at 08:41:21PM +0200, Wenhua Shi wrote:
>> 2018-04-08 18:51 GMT+02:00 David Miller :
>> >
>> > From: Wenhua Shi
>> > Date: Fri, 6 Apr 2018 03:43:39 +0200
>> >
>> > > Signed-off-by: Wenhua Shi
>> >
>> > This precondition should be made impossible instead of having to do
>> > an extra check everywhere that this helper is invoked, many of which
>> > are in fast paths.
>>
>> I believe the precondition you said is quite true. In my situation, I
>> have to disable GSO for some packet and I notice that it leads to a
>> worse performance (slower than 1Mbps, was almost 800Mbps).
>>
>> Here's the hook I use on debian 9.4, kernel version 4.9:
>
> There is quite a distance between 4.9 and net/net-next. Did you test
> on a more recent kernel too?
>
> Note that TCP stack now works with GSO being always on.
> 0a6b2a1dc2a2 ("tcp: switch to GSO being always on")
>
I've tried testing on the Fedora rawhide channel. The kernel version
is 4.17.0. Detail information is attached.
Without the hook
[root@fedora-s-1vcpu-1gb-sfo1-01 testing]# iperf -c
myanothernormalmachine -d
Server listening on TCP port 5001
TCP window size: 85.3 KByte (default)
Client connecting to myanothernormalmachine, TCP port 5001
TCP window size: 85.0 KByte (default)
[ 3] local 107.170.240.XXX port 44692 connected with
104.131.148.XXX port 5001
[ 5] local 107.170.240.XXX port 5001 connected with
104.131.148.XXX port 53978
[ ID] Interval Transfer Bandwidth
[ 3] 0.0-10.0 sec 1.04 GBytes 892 Mbits/sec
[ 5] 0.0-10.0 sec 757 MBytes 638 Mbits/sec
With the hook
[root@fedora-s-1vcpu-1gb-sfo1-01 testing]# iperf -c
myanothernormalmachine -d
Server listening on TCP port 5001
TCP window size: 85.3 KByte (default)
Client connecting to myanothernormalmachine, TCP port 5001
TCP window size: 85.0 KByte (default)
[ 3] local 107.170.240.XXX port 44694 connected with
104.131.148.XXX port 5001
[ 5] local 107.170.240.XXX port 5001 connected with
104.131.148.XXX port 53980
[ ID] Interval Transfer Bandwidth
[ 5] 0.0-10.0 sec 1.04 GBytes 894 Mbits/sec
[ 3] 0.0-13.5 sec 170 KBytes 103 Kbits/sec
Kernel
[root@fedora-s-1vcpu-1gb-sfo1-01 testing]# uname -a
Linux fedora-s-1vcpu-1gb-sfo1-01.localdomain
4.17.0-0.rc0.git5.2.fc29.x86_64 #1 SMP Mon Apr 9 17:16:30 UTC 2018
x86_64 x86_64 x86_64 GNU/Linux
Hook Source Code
[root@fedora-s-1vcpu-1gb-sfo1-01 testing]# cat testing.c
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
unsigned int hook_outgoing(
void * priv,
struct sk_buff * skb,
const struct nf_hook_state * state)
{
printk(KERN_INFO "Hook working...\n");
/* for some reason I have to disable GSO */
skb_gso_reset(skb);
/* The following won't work any more. */
// skb->sk->sk_gso_type = ~0;
return NF_ACCEPT;
}
static struct nf_hook_ops hook =
{
.hook = hook_outgoing,
.pf = PF_INET,
.hooknum = NF_INET_POST_ROUTING,
.priority = NF_IP_PRI_LAST,
};
static int __init init_testing(void)
{
nf_register_net_hook(&init_net, &hook);
return 0;
}
static void __exit exit_testing(void)
{
nf_unregister_net_hook(&init_net, &hook);
}
MODULE_LICENSE("GPL");
module_init(init_testing);
module_exit(exit_testing);
It turns out the problem exists and my previous bypassing trick is not
working any more. I'm now testing whether the patch is working for the
latest net-next branch.
Re: [PATCH] make net_gso_ok return false when gso_type is zero(invalid)
2018-04-08 18:51 GMT+02:00 David Miller :
>
> From: Wenhua Shi
> Date: Fri, 6 Apr 2018 03:43:39 +0200
>
> > Signed-off-by: Wenhua Shi
>
> This precondition should be made impossible instead of having to do
> an extra check everywhere that this helper is invoked, many of which
> are in fast paths.
I believe the precondition you said is quite true. In my situation, I
have to disable GSO for some packet and I notice that it leads to a
worse performance (slower than 1Mbps, was almost 800Mbps).
Here's the hook I use on debian 9.4, kernel version 4.9:
#include
#include
#include
#include
#include
#include
#include
#include
#include
unsigned int hook_outgoing (
void * priv,
struct sk_buff * skb,
const struct nf_hook_state * state)
{
/* for some reason I have to disable GSO */
skb_gso_reset(skb);
/* After I force sk_can_gso to return false here, the
performance comes back normal. */
// skb->sk->sk_gso_type = ~0;
return NF_ACCEPT;
}
static struct nf_hook_ops hook =
{
.hook = hook_outgoing,
.pf = PF_INET,
.hooknum = NF_INET_POST_ROUTING,
.priority = NF_IP_PRI_LAST,
};
static int __init init_testing(void)
{
nf_register_hook(&hook);
return 0;
}
static void __exit exit_testing(void)
{
nf_unregister_hook(&hook);
}
module_init(init_testing);
module_exit(exit_testing);
Here are the performance measurements.
Without the previous hook:
root@debian-s-1vcpu-1gb-sfo1-01:~/test# iperf -c myanothernormaldebian -d
Server listening on TCP port 5001
TCP window size: 85.3 KByte (default)
Client connecting to myanothernormaldebian, TCP port 5001
TCP window size: 255 KByte (default)
[ 3] local 192.241.204.XXX port 60528 connected with
104.131.148.XXX port 5001
[ 5] local 192.241.204.XXX port 5001 connected with
104.131.148.XXX port 58576
[ ID] Interval Transfer Bandwidth
[ 3] 0.0-10.0 sec 922 MBytes 773 Mbits/sec
[ 5] 0.0-10.1 sec 1.00 GBytes 849 Mbits/sec
And with the previous hook:
root@debian-s-1vcpu-1gb-sfo1-01:~/test# iperf -c myanothernormaldebian -d
Server listening on TCP port 5001
TCP window size: 85.3 KByte (default)
Client connecting to myanothernormaldebian, TCP port 5001
TCP window size: 85.0 KByte (default)
[ 3] local 192.241.204.XXX port 60530 connected with
104.131.148.XXX port 5001
[ 5] local 192.241.204.XXX port 5001 connected with
104.131.148.XXX port 58578
[ ID] Interval Transfer Bandwidth
[ 5] 0.0-10.2 sec 1.02 GBytes 864 Mbits/sec
[ 3] 0.0-13.5 sec 170 KBytes 103 Kbits/sec
Or it's just because of that I'm disabling the GSO in a wrong way?
[PATCH] make net_gso_ok return false when gso_type is zero(invalid)
Signed-off-by: Wenhua Shi --- include/linux/netdevice.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h index cf44503e..1f26cbcf 100644 --- a/include/linux/netdevice.h +++ b/include/linux/netdevice.h @@ -4187,7 +4187,7 @@ static inline bool net_gso_ok(netdev_features_t features, int gso_type) BUILD_BUG_ON(SKB_GSO_ESP != (NETIF_F_GSO_ESP >> NETIF_F_GSO_SHIFT)); BUILD_BUG_ON(SKB_GSO_UDP != (NETIF_F_GSO_UDP >> NETIF_F_GSO_SHIFT)); - return (features & feature) == feature; + return feature && (features & feature) == feature; } static inline bool skb_gso_ok(struct sk_buff *skb, netdev_features_t features) -- 2.11.0
[PATCH] fix typo in skbuff.c
--- net/core/skbuff.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/net/core/skbuff.c b/net/core/skbuff.c index 16982de6..e62476be 100644 --- a/net/core/skbuff.c +++ b/net/core/skbuff.c @@ -1896,7 +1896,7 @@ void *__pskb_pull_tail(struct sk_buff *skb, int delta) } /* If we need update frag list, we are in troubles. -* Certainly, it possible to add an offset to skb data, +* Certainly, it is possible to add an offset to skb data, * but taking into account that pulling is expected to * be very rare operation, it is worth to fight against * further bloating skb head and crucify ourselves here instead. -- 2.11.0
