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

[Ilya Leoshkevich <iii@linux.ibm.com>]

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?

> > + * is the 32 bits arrangement:
> > + */
> > +#define BTF_FLOAT_BITS(VAL)    ((VAL)  & 0x000000ff)
> > +
> >  #endif /* _UAPI__LINUX_BTF_H__ */
> > diff --git a/tools/include/uapi/linux/btf.h
> > b/tools/include/uapi/linux/btf.h
> > index 5a667107ad2c..e713430cb033 100644
> > --- a/tools/include/uapi/linux/btf.h
> > +++ b/tools/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
> > + * is the 32 bits arrangement:
> > + */
> > +#define BTF_FLOAT_BITS(VAL)    ((VAL)  & 0x000000ff)
> > +
> >  #endif /* _UAPI__LINUX_BTF_H__ */
> > --
> > 2.29.2
> >