Re: [RFC bpf-next v2 2/9] bpf: add register and unregister functions for struct_ops.

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

On 9/15/23 17:05, Martin KaFai Lau wrote:
> On 9/12/23 11:14 PM, thinker.li@gmail.com wrote:
>> +int register_bpf_struct_ops(struct bpf_struct_ops_mod *mod)
>> +{
>> +    struct bpf_struct_ops *st_ops = mod->st_ops;
>> +    struct bpf_verifier_log *log;
>> +    struct btf *btf;
>> +    int err;
>> +
>> +    if (mod->st_ops == NULL ||
>> +        mod->owner == NULL)
>> +        return -EINVAL;
>> +
>> +    log = kzalloc(sizeof(*log), GFP_KERNEL | __GFP_NOWARN);
>> +    if (!log) {
>> +        err = -ENOMEM;
>> +        goto errout;
>> +    }
>> +
>> +    log->level = BPF_LOG_KERNEL;
>> +
>> +    btf = btf_get_module_btf(mod->owner);
> 
> Where is btf_put called?
> 
> It is not stored anywhere in patch 2, so a bit confusing. I quickly 
> looked at the following patches but also don't see the bpf_put.

It is my fault to use it without calling btf_put().
I miss-understood the API, thought it doesn't increase refcount by
mistake.

> 
>> +    if (!btf) {
>> +        err = -EINVAL;
>> +        goto errout;
>> +    }
>> +
>> +    bpf_struct_ops_init_one(st_ops, btf, log);
>> +    err = add_struct_ops(st_ops);
>> +
>> +errout:
>> +    kfree(log);
>> +
>> +    return err;
>> +}
>> +EXPORT_SYMBOL(register_bpf_struct_ops);
>> +
>> +int unregister_bpf_struct_ops(struct bpf_struct_ops_mod *mod)
> 
> It is not clear to me why the subsystem needs to explicitly call 
> unregister_bpf_struct_ops(). Can it be done similar to the module kfunc 
> support (the kfunc_set_tab goes away with the btf)?

It could be. However, registering to module notifier
(register_module_notifier()) is more straight forward if we go
this way. What do you think?

> 
> Related to this, does it need to maintain a global struct_ops array for 
> all kernel module? Can the struct_ops be maintained under its 
> corresponding module btf itself?

What is the purpose?
We have a global struct_ops array already, although it is not
per-module. For now, the number of struct_ops is pretty small.
We have only one so far, and it is unlikely to grow fast in
near future. It is probably a bit overkill to have
per-module ones if this is what you mean.

> 
>> +{
>> +    struct bpf_struct_ops *st_ops = mod->st_ops;
>> +    int err;
>> +
>> +    err = remove_struct_ops(st_ops);
>> +    if (!err && st_ops->uninit)
>> +        err = st_ops->uninit();
>> +
>> +    return err;
>> +}
>> +EXPORT_SYMBOL(unregister_bpf_struct_ops);
> 
>