Re: [PATCH RFC 1/6] bpf: Add BTF_KIND_FLOAT to uapi

[Alexei Starovoitov <alexei.starovoitov@gmail.com>]

On Thu, Feb 11, 2021 at 1:26 PM Ilya Leoshkevich <iii@linux.ibm.com> wrote:
>
> On Wed, 2021-02-10 at 16:19 -0800, Andrii Nakryiko wrote:
> > On Tue, Feb 9, 2021 at 7:03 PM Ilya Leoshkevich <iii@linux.ibm.com>
> > wrote:
> > >
> > > Add a new kind value, expand the kind bitfield, add a macro for
> > > parsing the additional u32.
> > >
> > > Signed-off-by: Ilya Leoshkevich <iii@linux.ibm.com>
> > > ---
> > >  include/uapi/linux/btf.h       | 10 ++++++++--
> > >  tools/include/uapi/linux/btf.h | 10 ++++++++--
> > >  2 files changed, 16 insertions(+), 4 deletions(-)
> > >
> > > diff --git a/include/uapi/linux/btf.h b/include/uapi/linux/btf.h
> > > index 5a667107ad2c..e713430cb033 100644
> > > --- a/include/uapi/linux/btf.h
> > > +++ b/include/uapi/linux/btf.h
> > > @@ -52,7 +52,7 @@ struct btf_type {
> > >         };
> > >  };
> > >
> > > -#define BTF_INFO_KIND(info)    (((info) >> 24) & 0x0f)
> > > +#define BTF_INFO_KIND(info)    (((info) >> 24) & 0x1f)
> > >  #define BTF_INFO_VLEN(info)    ((info) & 0xffff)
> > >  #define BTF_INFO_KFLAG(info)   ((info) >> 31)
> > >
> > > @@ -72,7 +72,8 @@ struct btf_type {
> > >  #define BTF_KIND_FUNC_PROTO    13      /* Function Proto       */
> > >  #define BTF_KIND_VAR           14      /* Variable     */
> > >  #define BTF_KIND_DATASEC       15      /* Section      */
> > > -#define BTF_KIND_MAX           BTF_KIND_DATASEC
> > > +#define BTF_KIND_FLOAT         16      /* Floating point       */
> > > +#define BTF_KIND_MAX           BTF_KIND_FLOAT
> > >  #define NR_BTF_KINDS           (BTF_KIND_MAX + 1)
> > >
> > >  /* For some specific BTF_KIND, "struct btf_type" is immediately
> > > @@ -169,4 +170,9 @@ struct btf_var_secinfo {
> > >         __u32   size;
> > >  };
> > >
> > > +/* BTF_KIND_FLOAT is followed by a u32 and the following
> >
> >
> > what's the point of that u32, if BTF_FLOAT_BITS() is just t->size *
> > 8?
> > Why adding this complexity. BTF_KIND_INT has bits because we had an
> > inconvenient bitfield encoding as a special BTF_KIND_INT types, which
> > we since stopped using in favor of encoding bitfield sizes and
> > offsets
> > inside struct/union fields. I don't think there is any need for that
> > with FLOAT, so why waste space and add complexity and possibility for
> > inconsistencies?
>
> You are right, this is not necessary. I don't think something like a
> floating-point bitfield exists in the first place.
>
> > Disclaimer: I'm in a "just BTF_KIND_INT encoding bit for
> > floating-point numbers" camp.
>
> Despite me being the guy, who sent this series, I like such a simpler
> approach as well. In fact, my first attempt at this was even simpler -
> just a char[] - but this didn't let us distinguish floats from "real"
> byte arrays, which BTF_KIND_INT encoding does. But I think we need to
> convince Alexey that this would be OK? :-) If that helps, I can
> implement the BTF_KIND_INT encoding variant, so that we could compare
> both approaches. What do you think?

Sorry to crash your party :)
I'm very much against calling "float" a different kind of "int".
BTF is not equivalent to dwarf. It's not a traditional debug format
which purpose is to describe the types, line info, etc.
BTF is used in verification and its correctness is crucial.
Imagine a function with KIND_INT argument if there is no KIND_FLOAT
btf_func_model construction will silently consume 'int' which is actually
'float' which will cause kernel crash.
There is a wip on "unstable helpers". It needs accurate argument processing too.
The same thing will apply there. If KIND_FLOAT is mistaken for KIND_INT
the kernel will crash again, but in a different way.
The usage of floating point in the kernel is minimal, but the point stands.