On Wed, Feb 14, 2018 at 1:19 AM, David Miller <da...@davemloft.net> wrote:
> From: Tonghao Zhang <xiangxia.m....@gmail.com>
> Date: Mon, 12 Feb 2018 18:44:00 -0800
>
>> Sometimes, we want to know how many tcp sockets are in use
>> different _net_ namespaces. It's a key resource metric. With
>> this patch, we can get it via /proc/net/sockstat.
>>
>> The 'alloc' show in /proc/net/sockstat is the total tcp
>> sockets in the kernel. This patch moves it to namespace,
>> via adding a counter because the previous counter is used
>> in proto(e.g tcp, udp and sctp) memory management.
>>
>> Signed-off-by: Tonghao Zhang <xiangxia.m....@gmail.com>
> ...
>> @@ -453,6 +453,8 @@ void tcp_init_sock(struct sock *sk)
>> sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
>>
>> sk_sockets_allocated_inc(sk);
>> + if (likely(sk->sk_net_refcnt))
>> + tcp_sock_allocated_add(sock_net(sk), 1);
>> }
>> EXPORT_SYMBOL(tcp_init_sock);
>>
> ...
>> @@ -1928,6 +1928,8 @@ void tcp_v4_destroy_sock(struct sock *sk)
>> tcp_saved_syn_free(tp);
>>
>> sk_sockets_allocated_dec(sk);
>> + if (likely(sk->sk_net_refcnt))
>> + tcp_sock_allocated_add(sock_net(sk), -1);
>> }
>> EXPORT_SYMBOL(tcp_v4_destroy_sock);
>>
> ...
>> @@ -559,6 +559,9 @@ struct sock *tcp_create_openreq_child(const struct sock
>> *sk,
>> newtp->rack.reo_wnd_persist = 0;
>> newtp->rack.dsack_seen = 0;
>>
>> + if (likely(newsk->sk_net_refcnt))
>> + tcp_sock_allocated_add(sock_net(newsk), 1);
>> +
>> __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS);
>
> The amount of new conditional tests in these fast paths are not
> justified for this new counter which is of debatable usefullness.
sorry, too late for reply. Did you mean this counter will affect performance ?
I tested the patch witch netperf.
Before patch:
# netperf -H 192.168.3.5 -t TCP_CRR -l 60 -- -O "THROUGHPUT,
THROUGHPUT_UNITS, MIN_LATENCY, MAX_LATENCY, MEAN_LATENCY"
MIGRATED TCP Connect/Request/Response TEST from 0.0.0.0 (0.0.0.0) port
0 AF_INET to 192.168.3.5 () port 0 AF_INET
Throughput Throughput Minimum Maximum Mean
Units Latency Latency Latency
Microseconds Microseconds Microseconds
7719.86 Trans/s 70 660 129.32
# netperf -H 192.168.3.5 -t TCP_CRR -l 60 -- -O "THROUGHPUT,
THROUGHPUT_UNITS, MIN_LATENCY, MAX_LATENCY, MEAN_LATENCY"
MIGRATED TCP Connect/Request/Response TEST from 0.0.0.0 (0.0.0.0) port
0 AF_INET to 192.168.3.5 () port 0 AF_INET
Throughput Throughput Minimum Maximum Mean
Units Latency Latency Latency
Microseconds Microseconds Microseconds
7574.00 Trans/s 86 13016 131.82
# netperf -H 192.168.3.5 -t TCP_CRR -l 60 -- -O "THROUGHPUT,
THROUGHPUT_UNITS, MIN_LATENCY, MAX_LATENCY, MEAN_LATENCY"
MIGRATED TCP Connect/Request/Response TEST from 0.0.0.0 (0.0.0.0) port
0 AF_INET to 192.168.3.5 () port 0 AF_INET
Throughput Throughput Minimum Maximum Mean
Units Latency Latency Latency
Microseconds Microseconds Microseconds
7556.66 Trans/s 87 15152 132.06
After patch:
# netperf -H 192.168.3.5 -t TCP_CRR -l 60 -- -O "THROUGHPUT,
THROUGHPUT_UNITS, MIN_LATENCY, MAX_LATENCY, MEAN_LATENCY"
MIGRATED TCP Connect/Request/Response TEST from 0.0.0.0 (0.0.0.0) port
0 AF_INET to 192.168.3.5 () port 0 AF_INET
Throughput Throughput Minimum Maximum Mean
Units Latency Latency Latency
Microseconds Microseconds Microseconds
7277.86 Trans/s 86 14884 137.19
# netperf -H 192.168.3.5 -t TCP_CRR -l 60 -- -O "THROUGHPUT,
THROUGHPUT_UNITS, MIN_LATENCY, MAX_LATENCY, MEAN_LATENCY"
MIGRATED TCP Connect/Request/Response TEST from 0.0.0.0 (0.0.0.0) port
0 AF_INET to 192.168.3.5 () port 0 AF_INET
Throughput Throughput Minimum Maximum Mean
Units Latency Latency Latency
Microseconds Microseconds Microseconds
7534.73 Trans/s 92 7531 132.50
# netperf -H 192.168.3.5 -t TCP_CRR -l 60 -- -O "THROUGHPUT,
THROUGHPUT_UNITS, MIN_LATENCY, MAX_LATENCY, MEAN_LATENCY"
MIGRATED TCP Connect/Request/Response TEST from 0.0.0.0 (0.0.0.0) port
0 AF_INET to 192.168.3.5 () port 0 AF_INET
Throughput Throughput Minimum Maximum Mean
Units Latency Latency Latency
Microseconds Microseconds Microseconds
7622.17 Trans/s 89 2370 130.98
> I'm not applying this, sorry.
