[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <b5aa477d-a4b1-45cb-af44-bd737504734e@linux.alibaba.com>
Date: Fri, 11 Oct 2024 14:43:56 +0800
From: "D. Wythe" <alibuda@...ux.alibaba.com>
To: Alexei Starovoitov <alexei.starovoitov@...il.com>
Cc: kgraul@...ux.ibm.com, wenjia@...ux.ibm.com, jaka@...ux.ibm.com,
wintera@...ux.ibm.com, guwen@...ux.alibaba.com,
Alexei Starovoitov <ast@...nel.org>, Jakub Kicinski <kuba@...nel.org>,
"David S. Miller" <davem@...emloft.net>,
Network Development <netdev@...r.kernel.org>,
linux-s390 <linux-s390@...r.kernel.org>, linux-rdma@...r.kernel.org,
Tony Lu <tonylu@...ux.alibaba.com>, Paolo Abeni <pabeni@...hat.com>,
Eric Dumazet <edumazet@...gle.com>, bpf <bpf@...r.kernel.org>
Subject: Re: [PATCH net-next] net/smc: Introduce a hook to modify syn_smc at
runtime
On 10/11/24 12:21 AM, Alexei Starovoitov wrote:
> On Wed, Oct 9, 2024 at 8:58 PM D. Wythe <alibuda@...ux.alibaba.com> wrote:
>>
>>
>> +__bpf_hook_start();
>> +
>> +__weak noinline int select_syn_smc(const struct sock *sk, struct sockaddr *peer)
>> +{
>> + return 1;
>> +}
>> +
>> +__bpf_hook_end();
>> +
>> int smc_nl_dump_hs_limitation(struct sk_buff *skb, struct netlink_callback *cb)
>> {
>> struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
>> @@ -156,19 +165,43 @@ static struct sock *smc_tcp_syn_recv_sock(const struct sock *sk,
>> return NULL;
>> }
>>
>> -static bool smc_hs_congested(const struct sock *sk)
>> +static void smc_openreq_init(struct request_sock *req,
>> + const struct tcp_options_received *rx_opt,
>> + struct sk_buff *skb, const struct sock *sk)
>> {
>> + struct inet_request_sock *ireq = inet_rsk(req);
>> + struct sockaddr_storage rmt_sockaddr = {};
>> const struct smc_sock *smc;
>>
>> smc = smc_clcsock_user_data(sk);
>>
>> if (!smc)
>> - return true;
>> + return;
>>
>> - if (workqueue_congested(WORK_CPU_UNBOUND, smc_hs_wq))
>> - return true;
>> + if (smc->limit_smc_hs && workqueue_congested(WORK_CPU_UNBOUND, smc_hs_wq))
>> + goto out_no_smc;
>>
>> - return false;
>> + rmt_sockaddr.ss_family = sk->sk_family;
>> +
>> + if (rmt_sockaddr.ss_family == AF_INET) {
>> + struct sockaddr_in *rmt4_sockaddr = (struct sockaddr_in *)&rmt_sockaddr;
>> +
>> + rmt4_sockaddr->sin_addr.s_addr = ireq->ir_rmt_addr;
>> + rmt4_sockaddr->sin_port = ireq->ir_rmt_port;
>> +#if IS_ENABLED(CONFIG_IPV6)
>> + } else {
>> + struct sockaddr_in6 *rmt6_sockaddr = (struct sockaddr_in6 *)&rmt_sockaddr;
>> +
>> + rmt6_sockaddr->sin6_addr = ireq->ir_v6_rmt_addr;
>> + rmt6_sockaddr->sin6_port = ireq->ir_rmt_port;
>> +#endif /* CONFIG_IPV6 */
>> + }
>> +
>> + ireq->smc_ok = select_syn_smc(sk, (struct sockaddr *)&rmt_sockaddr);
>> + return;
>> +out_no_smc:
>> + ireq->smc_ok = 0;
>> + return;
>> }
>>
>> struct smc_hashinfo smc_v4_hashinfo = {
>> @@ -1671,7 +1704,7 @@ int smc_connect(struct socket *sock, struct sockaddr *addr,
>> }
>>
>> smc_copy_sock_settings_to_clc(smc);
>> - tcp_sk(smc->clcsock->sk)->syn_smc = 1;
>> + tcp_sk(smc->clcsock->sk)->syn_smc = select_syn_smc(sk, addr);
>> if (smc->connect_nonblock) {
>> rc = -EALREADY;
>> goto out;
>> @@ -2650,8 +2683,7 @@ int smc_listen(struct socket *sock, int backlog)
>>
>> inet_csk(smc->clcsock->sk)->icsk_af_ops = &smc->af_ops;
>>
>> - if (smc->limit_smc_hs)
>> - tcp_sk(smc->clcsock->sk)->smc_hs_congested = smc_hs_congested;
>> + tcp_sk(smc->clcsock->sk)->smc_openreq_init = smc_openreq_init;
>>
>> rc = kernel_listen(smc->clcsock, backlog);
>> if (rc) {
>> @@ -3475,6 +3507,24 @@ static void __net_exit smc_net_stat_exit(struct net *net)
>> .exit = smc_net_stat_exit,
>> };
>>
>> +#if IS_ENABLED(CONFIG_BPF_SYSCALL)
>> +BTF_SET8_START(bpf_smc_fmodret_ids)
>> +BTF_ID_FLAGS(func, select_syn_smc)
>> +BTF_SET8_END(bpf_smc_fmodret_ids)
>> +
>> +static const struct btf_kfunc_id_set bpf_smc_fmodret_set = {
>> + .owner = THIS_MODULE,
>> + .set = &bpf_smc_fmodret_ids,
>> +};
>> +
>> +static int bpf_smc_kfunc_init(void)
>> +{
>> + return register_btf_fmodret_id_set(&bpf_smc_fmodret_set);
>> +}
>
> fmodret was an approach that hid-bpf took initially,
> but eventually they removed it all and switched to struct-ops approach.
> Please learn that lesson.
> Use struct_ops from the beginning.
>
> I did a presentation recently explaining the motivation behind
> struct_ops and tips on how to extend the kernel.
> TLDR: the step one is to design the extension _without_ bpf.
> The interface should be usable for kernel modules.
> And then when you have *_ops style api in place
> the bpf progs will plug-in without extra work.
>
> Slides:
> https://github.com/4ast/docs/blob/main/BPF%20struct-ops.pdf
Hi Alexei,
Thanks very much for your suggestion.
In fact, I tried struct_ops in SMC about a year ago. Unfortunately, at that time struct_ops did not
support registration from modules, and I had to move some smc dependencies into bpf, which met with
community opposition. However, I noticed that this feature is now supported, so perhaps this is an
opportunity.
But on the other hand, given the current functionality, I wonder if struct_ops might be an overkill.
I haven't been able to come up with a suitable abstraction to define this ops, and in the future,
this ops might only contain the very one callback (select_syn_smc).
Looking forward for your advises.
Thanks,
D. Wythe
Powered by blists - more mailing lists