Re: [PATCH bpf-next 3/7] bpf: Register and unregister a struct_ops by their bpf_links.

[Martin KaFai Lau <martin.lau@linux.dev>]

On 2/16/23 8:42 AM, Kui-Feng Lee wrote:
>>> @@ -638,6 +647,8 @@ static struct bpf_map *bpf_struct_ops_map_alloc(union 
>>> bpf_attr *attr)
>>>       set_vm_flush_reset_perms(st_map->image);
>>>       bpf_map_init_from_attr(map, attr);
>>> +    map->map_flags |= attr->map_flags & BPF_F_LINK;
>>
>> This should have already been done in bpf_map_init_from_attr().
> 
> bpf_map_init_from_attr() will filter out all flags except BPF_F_RDONLY & 
> BPF_F_WRONLY.

should be the opposite:

static u32 bpf_map_flags_retain_permanent(u32 flags)
{
	/* Some map creation flags are not tied to the map object but
          * rather to the map fd instead, so they have no meaning upon
          * map object inspection since multiple file descriptors with
          * different (access) properties can exist here. Thus, given
          * this has zero meaning for the map itself, lets clear these
          * from here.
          */
	return flags & ~(BPF_F_RDONLY | BPF_F_WRONLY);
}