Re: [PATCH bpf-next v5 2/6] bpf/verifier: add bpf_timer as a kfunc capable type

[Benjamin Tissoires]

On Mar 24 2024, Alexei Starovoitov wrote:
> On Sat, Mar 23, 2024 at 9:57 PM Kumar Kartikeya Dwivedi
> <memxor@xxxxxxxxx> wrote:
> 
> > > > >
> > > > > Observation is correct. The patch is buggy,
> > > > > but the suggestion to follow process_dynptr_func() will lead
> > > > > to unnecessary complexity.
> > > > > dynptr-s are on stack with plenty of extra checks.
> > > >
> > > > The suggestion was to call process_timer_func, not process_dynptr_func.
> > > >
> > > > > In this case bpf_timer is in map_value.
> > > > > Much simpler is to follow KF_ARG_PTR_TO_MAP approach.
> > > >
> > > > What I meant by the example was that dynptr handling does the same
> > > > thing for kfuncs and helpers (using the same function), so timer
> > > > arguments should do the same (i.e. use process_timer_func), which will
> > > > do all checks for constant offset (ensuring var_off is tnum_is_const)
> > > > and match it against btf_record->timer_off.
> > >
> > > I don't follow. Please elaborate with a patch.
> > > The var_off and off is a part of the bug, but it's not the biggest part of it.
> >
> > Not compile tested.

Compiles just fine :)

> 
> I see. All makes sense to me.
> 
> Benjamin,
> pls incorporate it in your set.
> 

OK, done!

I just had to revert to the following or KF_ARG_TIMER_ID was not
recognized by the verifier:

---
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index 7ee20e9d14bd..a5e147468ac8 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -10848,7 +10848,7 @@ BTF_ID(struct, bpf_list_head)
 BTF_ID(struct, bpf_list_node)
 BTF_ID(struct, bpf_rb_root)
 BTF_ID(struct, bpf_rb_node)
-BTF_ID(struct, bpf_timer)
+BTF_ID(struct, bpf_timer_kern)
 
 static bool __is_kfunc_ptr_arg_type(const struct btf *btf,
 				    const struct btf_param *arg, int type)
---

Cheers,
Benjamin