Re: [PATCH bpf-next 6/6] selftests/bpf: test detaching struct_ops links.

[Kui-Feng Lee <sinquersw@gmail.com>]

On 5/2/24 11:15, Martin KaFai Lau wrote:
> On 4/29/24 2:36 PM, Kui-Feng Lee wrote:
>> @@ -572,6 +576,12 @@ static int bpf_dummy_reg(void *kdata)
>>       if (ops->test_2)
>>           ops->test_2(4, ops->data);
>> +    if (ops->do_unreg) {
>> +        rcu_read_lock();
>> +        bpf_struct_ops_kvalue_unreg(kdata);
> 
> Instead of unreg() immediately before the reg() has returned, the test 
> should reflect more on how the subsystem can use it in practice. The 
> subsystem does not do unreg() during reg().
> 
> It also needs to test a case when the link is created and successfully 
> registered to the subsystem. The user space does BPF_LINK_DETACH first 
> and then the subsystem does link->ops->detach() by itself later.

agree

> 
> It can create a kfunc in bpf_testmod.c to trigger the subsystem to do 
> link->ops->detach(). The kfunc can be called by a SEC("syscall") bpf 
> prog which is run by bpf_prog_test_run_opts(). The test_progs can then 
> decide on the timing when to do link->ops->detach() to test different 
> cases.

What is the purpose of this part?
If it goes through link->ops->detach(), it should work just like to call
bpf_link_detach() twice on the same link from the user space. Do you
want to make sure detaching a link twice work?

> 
>> +        rcu_read_unlock();
>> +    }
>> +
>>       return 0;
>>   }
>