On 2/19/21 5:00 AM, Ilya Leoshkevich wrote:

On Thu, 2021-02-18 at 21:41 -0800, Yonghong Song wrote:

quoted

On 2/18/21 6:25 PM, Ilya Leoshkevich wrote:

quoted

Also document the expansion of the kind bitfield.

Signed-off-by: Ilya Leoshkevich <iii@linux.ibm.com>
---
   Documentation/bpf/btf.rst | 17 +++++++++++++++--
   1 file changed, 15 insertions(+), 2 deletions(-)

diff --git a/Documentation/bpf/btf.rst b/Documentation/bpf/btf.rst
index 44dc789de2b4..4f25c992d442 100644
--- a/Documentation/bpf/btf.rst
+++ b/Documentation/bpf/btf.rst

@@ -84,6 +84,7 @@ sequentially and type id is assigned to each

recognized type starting from id
       #define BTF_KIND_FUNC_PROTO     13      /* Function
Proto       */
       #define BTF_KIND_VAR            14      /* Variable     */
       #define BTF_KIND_DATASEC        15      /* Section      */
+    #define BTF_KIND_FLOAT          16      /* Floating
point       */
   
   Note that the type section encodes debug info, not just pure
types.
   ``BTF_KIND_FUNC`` is not a type, and it represents a defined
subprogram.

@@ -95,8 +96,8 @@ Each type contains the following common data::
           /* "info" bits arrangement
            * bits  0-15: vlen (e.g. # of struct's members)
            * bits 16-23: unused
-         * bits 24-27: kind (e.g. int, ptr, array...etc)
-         * bits 28-30: unused
+         * bits 24-28: kind (e.g. int, ptr, array...etc)
+         * bits 29-30: unused
            * bit     31: kind_flag, currently used by
            *             struct, union and fwd
            */

@@ -452,6 +453,18 @@ map definition.
     * ``offset``: the in-section offset of the variable
     * ``size``: the size of the variable in bytes
   
+2.2.16 BTF_KIND_FLOAT
+~~~~~~~~~~~~~~~~~~~~~
+
+``struct btf_type`` encoding requirement:
+ * ``name_off``: any valid offset
+ * ``info.kind_flag``: 0
+ * ``info.kind``: BTF_KIND_FLOAT
+ * ``info.vlen``: 0
+ * ``size``: the size of the float type in bytes.

I would be good to specify the allowed size in bytes 2, multiple of
4.
currently we do not have a maximum value, maybe 128. have a float
type
something like 2^10 seems strange.

I tried to write this all down and realized it's simpler to enumerate
the allowed values: 2, 4, 8, 12 and 16. I don't think there are 32-byte
floats on any of the architectures supported by the kernel.

This make senses. My above 128 means 128bits (sorry!), which is 16 
bytes, align with what you suggested.