Re: [PATCH bpf-next v3 2/7] bpf: enable detaching links of struct_ops objects.

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

On 5/20/24 18:22, Martin KaFai Lau wrote:
> On 5/9/24 5:29 PM, Kui-Feng Lee wrote:
>> +static int bpf_struct_ops_map_link_detach(struct bpf_link *link)
>> +{
>> +    struct bpf_struct_ops_link *st_link = container_of(link, struct 
>> bpf_struct_ops_link, link);
>> +    struct bpf_struct_ops_map *st_map;
>> +    struct bpf_map *map;
>> +
>> +    mutex_lock(&update_mutex);
>> +
>> +    map = rcu_dereference_protected(st_link->map, 
>> lockdep_is_held(&update_mutex));
>> +    if (!map) {
>> +        mutex_unlock(&update_mutex);
>> +        return -EINVAL;
>> +    }
>> +    st_map = container_of(map, struct bpf_struct_ops_map, map);
>> +
>> +    st_map->st_ops_desc->st_ops->unreg(&st_map->kvalue.data, link);
>> +
>> +    rcu_assign_pointer(st_link->map, NULL);
>> +    /* Pair with bpf_map_get() in bpf_struct_ops_link_create() or
>> +     * bpf_map_inc() in bpf_struct_ops_map_link_update().
>> +     */
>> +    bpf_map_put(&st_map->map);
>> +
>> +    mutex_unlock(&update_mutex);
>> +
>> +    return 0;
>> +}
>> +
>>   static const struct bpf_link_ops bpf_struct_ops_map_lops = {
>>       .dealloc = bpf_struct_ops_map_link_dealloc,
>> +    .detach = bpf_struct_ops_map_link_detach,
>>       .show_fdinfo = bpf_struct_ops_map_link_show_fdinfo,
>>       .fill_link_info = bpf_struct_ops_map_link_fill_link_info,
>>       .update_map = bpf_struct_ops_map_link_update,
>> @@ -1176,13 +1208,19 @@ int bpf_struct_ops_link_create(union bpf_attr 
>> *attr)
>>       if (err)
>>           goto err_out;
>> +    /* Init link->map before calling reg() in case being detached
>> +     * immediately.
>> +     */
> 
> It is not obvious in the patch how this (immediate detach by subsystem 
> after reg) may work without race, so I think it is easier to ask.
> 
> [ I put back the err_out context at the end ]
> 
>> +    RCU_INIT_POINTER(link->map, map);
>> +
>>       err = st_map->st_ops_desc->st_ops->reg(st_map->kvalue.data, 
>> &link->link);
>>       if (err) {
>> +        RCU_INIT_POINTER(link->map, NULL);
> 
> In the bpf_struct_ops_map_link_detach() above, the update to link->map 
> is protected by the update_mutex. Could you explain how the link->map 
> update to NULL is safe here without holding the update_mutex?

If err is not zero, it means the subsystem rejects the pair of the
object and the link passing in. So, it has no reasonable to call
bpf_struct_ops_map_link_detach() for this link.

Does it make sense to you?

> 
>>           bpf_link_cleanup(&link_primer);
>> +        /* The link has been free by bpf_link_cleanup() */
>>           link = NULL;
>>           goto err_out;
>>       }

At this point, we don't change the content of the link anymore except
changing link->fd in bpf_link_settle(). So, it should be safe to call
bpf_struct_ops_map_link_detach() from the subsystem.

Should I explain it in a comment if you think it makes sense to you?

>> -    RCU_INIT_POINTER(link->map, map);
>>       return bpf_link_settle(&link_primer);
>>
> 
> err_out:
>      bpf_map_put(map);
>      kfree(link);
>      return err;
> }
> 
>