Re: [PATCH v3 bpf-next 3/4] bpf: Support pointers in global func args

[Dmitrii Banshchikov <me@ubique.spb.ru>]

On Fri, Feb 12, 2021 at 06:09:37PM -0800, Alexei Starovoitov wrote:
> On Sat, Feb 13, 2021 at 12:56:41AM +0400, Dmitrii Banshchikov wrote:
> > Add an ability to pass a pointer to a type with known size in arguments
> > of a global function. Such pointers may be used to overcome the limit on
> > the maximum number of arguments, avoid expensive and tricky workarounds
> > and to have multiple output arguments.
> 
> Thanks a lot for adding this feature and exhaustive tests.
> It's a massive improvement in function-by-function verification.
> Hopefully it will increase its adoption.
> I've applied the set to bpf-next.
> 
> > @@ -5349,10 +5352,6 @@ int btf_check_func_arg_match(struct bpf_verifier_env *env, int subprog,
> >  			goto out;
> >  		}
> >  		if (btf_type_is_ptr(t)) {
> > -			if (reg->type == SCALAR_VALUE) {
> > -				bpf_log(log, "R%d is not a pointer\n", i + 1);
> > -				goto out;
> > -			}
> 
> Thanks for nuking this annoying warning along the way.
> People complained that the verification log for normal static functions
> contains above inexplicable message.
> 
> >  			/* If function expects ctx type in BTF check that caller
> >  			 * is passing PTR_TO_CTX.
> >  			 */
> > @@ -5367,6 +5366,25 @@ int btf_check_func_arg_match(struct bpf_verifier_env *env, int subprog,
> >  					goto out;
> >  				continue;
> >  			}
> > +
> > +			if (!is_global)
> > +				goto out;
> > +
> > +			t = btf_type_skip_modifiers(btf, t->type, NULL);
> > +
> > +			ref_t = btf_resolve_size(btf, t, &type_size);
> > +			if (IS_ERR(ref_t)) {
> > +				bpf_log(log,
> > +				    "arg#%d reference type('%s %s') size cannot be determined: %ld\n",
> > +				    i, btf_type_str(t), btf_name_by_offset(btf, t->name_off),
> > +					PTR_ERR(ref_t));
> 
> Hopefully one annoying message won't get replaced with this annoying message :)
> I think the type size should be known most of the time. So it should be fine.
> 
> > +		if (btf_type_is_ptr(t)) {
> > +			if (btf_get_prog_ctx_type(log, btf, t, prog_type, i)) {
> > +				reg->type = PTR_TO_CTX;
> > +				continue;
> > +			}
> 
> Do you think it would make sense to nuke another message in btf_get_prog_ctx_type ?
> With this newly gained usability of global function the message
> "arg#0 type is not a struct"
> is not useful.
> It was marginally useful in the past. Because global funcs supported
> ptr_to_ctx only it wasn't seen as often.
> Now this message probably can simply be removed. wdyt?

Yes, I hit this log message while was working on the patch and it
looked confusing but forgot to adjust/remove it.
I will prepare patch.
Thank you.



-- 

Dmitrii Banshchikov