Re: [PATCH bpf-next v2 3/3] bpf: treewide: Annotate BPF kfuncs in BTF

[Jiri Olsa <olsajiri@gmail.com>]

On Fri, Jan 05, 2024 at 09:55:43AM -0700, Daniel Xu wrote:
> On Fri, Jan 05, 2024 at 04:11:33PM +0100, Jiri Olsa wrote:
> > On Thu, Jan 04, 2024 at 07:45:49PM -0700, Daniel Xu wrote:
> > 
> > SNIP
> > 
> > > diff --git a/fs/verity/measure.c b/fs/verity/measure.c
> > > index bf7a5f4cccaf..3969d54158d1 100644
> > > --- a/fs/verity/measure.c
> > > +++ b/fs/verity/measure.c
> > > @@ -159,9 +159,9 @@ __bpf_kfunc int bpf_get_fsverity_digest(struct file *file, struct bpf_dynptr_ker
> > >  
> > >  __bpf_kfunc_end_defs();
> > >  
> > > -BTF_SET8_START(fsverity_set_ids)
> > > +BTF_KFUNCS_START(fsverity_set_ids)
> > >  BTF_ID_FLAGS(func, bpf_get_fsverity_digest, KF_TRUSTED_ARGS)
> > > -BTF_SET8_END(fsverity_set_ids)
> > > +BTF_KFUNCS_END(fsverity_set_ids)
> > >  
> > >  static int bpf_get_fsverity_digest_filter(const struct bpf_prog *prog, u32 kfunc_id)
> > >  {
> > > diff --git a/kernel/bpf/btf.c b/kernel/bpf/btf.c
> > > index 51e8b4bee0c8..8cc718f37a9d 100644
> > > --- a/kernel/bpf/btf.c
> > > +++ b/kernel/bpf/btf.c
> > > @@ -7802,6 +7802,10 @@ int register_btf_kfunc_id_set(enum bpf_prog_type prog_type,
> > >  {
> > >  	enum btf_kfunc_hook hook;
> > >  
> > > +	/* All kfuncs need to be tagged as such in BTF */
> > > +	if (WARN_ON(!(kset->set->flags & BTF_SET8_KFUNCS)))
> > > +		return -EINVAL;
> > 
> > having the warning for module with wrong set8 flags seems wrong to me,
> > I think we should trigger the warn only for kernel calls.. by adding
> > kset->owner check in the condition above
> 
> Just checking:
> 
> The reasoning is that =m and out-of-tree modules can and should check
> return code, right?
> 
> And =y modules or vmlinux-based registrations do not check return code,
> so WARN() is necessary?
> 
> If so, I'd agree.

right, I was also concerned we could flood console with loading module
that just uses wrong set8.. perhaps we could just use WARN_ON_ONCE with
no additional checks

jirka