Date:   Sun, 16 Dec 2018 06:18:59 +0000
From:   Yonghong Song <yhs@...com>
To:     Alexei Starovoitov <alexei.starovoitov@...il.com>,
        Martin Lau <kafai@...com>
CC:     "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        Alexei Starovoitov <ast@...com>,
        Daniel Borkmann <daniel@...earbox.net>,
        Kernel Team <Kernel-team@...com>
Subject: Re: [PATCH bpf-next v2 2/8] bpf: btf: fix struct/union/fwd types with
 kind_flag



On 12/15/18 9:44 AM, Alexei Starovoitov wrote:
> On Sat, Dec 15, 2018 at 04:37:06PM +0000, Martin Lau wrote:
>> On Fri, Dec 14, 2018 at 03:34:27PM -0800, Yonghong Song wrote:
>>> This patch fixed two issues with BTF. One is related to
>>> struct/union bitfield encoding and the other is related to
>>> forward type.
>>>
>>> Issue #1 and solution:
>>> ======================
>>>
>>> Current btf encoding of bitfield follows what pahole generates.
>>> For each bitfield, pahole will duplicate the type chain and
>>> put the bitfield size at the final int or enum type.
>>> Since the BTF enum type cannot encode bit size,
>>> pahole workarounds the issue by generating
>>> an int type whenever the enum bit size is not 32.
>>>
>>> For example,
>>>    -bash-4.4$ cat t.c
>>>    typedef int ___int;
>>>    enum A { A1, A2, A3 };
>>>    struct t {
>>>      int a[5];
>>>      ___int b:4;
>>>      volatile enum A c:4;
>>>    } g;
>>>    -bash-4.4$ gcc -c -O2 -g t.c
>>> The current kernel supports the following BTF encoding:
>>>    $ pahole -JV t.o
>>>    [1] TYPEDEF ___int type_id=2
>>>    [2] INT int size=4 bit_offset=0 nr_bits=32 encoding=SIGNED
>>>    [3] ENUM A size=4 vlen=3
>>>          A1 val=0
>>>          A2 val=1
>>>          A3 val=2
>>>    [4] STRUCT t size=24 vlen=3
>>>          a type_id=5 bits_offset=0
>>>          b type_id=9 bits_offset=160
>>>          c type_id=11 bits_offset=164
>>>    [5] ARRAY (anon) type_id=2 index_type_id=2 nr_elems=5
>>>    [6] INT sizetype size=8 bit_offset=0 nr_bits=64 encoding=(none)
>>>    [7] VOLATILE (anon) type_id=3
>>>    [8] INT int size=1 bit_offset=0 nr_bits=4 encoding=(none)
>>>    [9] TYPEDEF ___int type_id=8
>>>    [10] INT (anon) size=1 bit_offset=0 nr_bits=4 encoding=SIGNED
>>>    [11] VOLATILE (anon) type_id=10
>>>
>>> Two issues are in the above:
>>>    . by changing enum type to int, we lost the original
>>>      type information and this will not be ideal later
>>>      when we try to convert BTF to a header file.
>>>    . the type duplication for bitfields will cause
>>>      BTF bloat. Duplicated types cannot be deduplicated
>>>      later if the bitfield size is different.
>>>
>>> To fix this issue, this patch implemented a compatible
>>> change for BTF struct type encoding:
>>>    . the bit 31 of struct_type->info, previously reserved,
>>>      now is used to indicate whether bitfield_size is
>>>      encoded in btf_member or not.
>>>    . if bit 31 of struct_type->info is set,
>>>      btf_member->offset will encode like:
>>>        bit 0 - 23: bit offset
>>>        bit 24 - 31: bitfield size
>>>      if bit 31 is not set, the old behavior is preserved:
>>>        bit 0 - 31: bit offset
>>>
>>> So if the struct contains a bit field, the maximum bit offset
>>> will be reduced to (2^24 - 1) instead of MAX_UINT. The maximum
>>> bitfield size will be 256 which is enough for today as maximum
>>> bitfield in compiler can be 128 where int128 type is supported.
>>>
>>> This kernel patch intends to support the new BTF encoding:
>>>    $ pahole -JV t.o
>>>    [1] TYPEDEF ___int type_id=2
>>>    [2] INT int size=4 bit_offset=0 nr_bits=32 encoding=SIGNED
>>>    [3] ENUM A size=4 vlen=3
>>>          A1 val=0
>>>          A2 val=1
>>>          A3 val=2
>>>    [4] STRUCT t kind_flag=1 size=24 vlen=3
>>>          a type_id=5 bitfield_size=0 bits_offset=0
>>>          b type_id=1 bitfield_size=4 bits_offset=160
>>>          c type_id=7 bitfield_size=4 bits_offset=164
>>>    [5] ARRAY (anon) type_id=2 index_type_id=2 nr_elems=5
>>>    [6] INT sizetype size=8 bit_offset=0 nr_bits=64 encoding=(none)
>>>    [7] VOLATILE (anon) type_id=3
>>>
>>> Issue #2 and solution:
>>> ======================
>>>
>>> Current forward type in BTF does not specify whether the original
>>> type is struct or union. This will not work for type pretty print
>>> and BTF-to-header-file conversion as struct/union must be specified.
>>>    $ cat tt.c
>>>    struct t;
>>>    union u;
>>>    int foo(struct t *t, union u *u) { return 0; }
>>>    $ gcc -c -g -O2 tt.c
>>>    $ pahole -JV tt.o
>>>    [1] INT int size=4 bit_offset=0 nr_bits=32 encoding=SIGNED
>>>    [2] FWD t type_id=0
>>>    [3] PTR (anon) type_id=2
>>>    [4] FWD u type_id=0
>>>    [5] PTR (anon) type_id=4
>>>
>>> To fix this issue, similar to issue #1, type->info bit 31
>>> is used. If the bit is set, it is union type. Otherwise, it is
>>> a struct type.
>>>
>>>    $ pahole -JV tt.o
>>>    [1] INT int size=4 bit_offset=0 nr_bits=32 encoding=SIGNED
>>>    [2] FWD t kind_flag=0 type_id=0
>>>    [3] PTR (anon) kind_flag=0 type_id=2
>>>    [4] FWD u kind_flag=1 type_id=0
>>>    [5] PTR (anon) kind_flag=0 type_id=4
>>>
>>> Pahole/LLVM change:
>>> ===================
>>>
>>> The new kind_flag functionality has been implemented in pahole
>>> and llvm:
>>>    https://github.com/yonghong-song/pahole/tree/bitfield
>>>    https://github.com/yonghong-song/llvm/tree/bitfield
>>>
>>> Note that pahole hasn't implemented func/func_proto kind
>>> and .BTF.ext. So to print function signature with bpftool,
>>> the llvm compiler should be used.
>>>
>>> Fixes: 69b693f0aefa ("bpf: btf: Introduce BPF Type Format (BTF)")
>>> Signed-off-by: Martin KaFai Lau <kafai@...com>
>>> Signed-off-by: Yonghong Song <yhs@...com>
>>> ---
>>>   include/uapi/linux/btf.h |  15 ++-
>>>   kernel/bpf/btf.c         | 274 ++++++++++++++++++++++++++++++++++++---
>>>   2 files changed, 267 insertions(+), 22 deletions(-)
>>>
>>> diff --git a/include/uapi/linux/btf.h b/include/uapi/linux/btf.h
>>> index 14f66948fc95..34aba40ed926 100644
>>> --- a/include/uapi/linux/btf.h
>>> +++ b/include/uapi/linux/btf.h
>>> @@ -34,7 +34,9 @@ struct btf_type {
>>>   	 * 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 28-31: unused
>>> +	 * bits 28-30: unused
>>> +	 * bit     31: kind_flag, currently used by
>>> +	 *             struct, union and fwd
>>>   	 */
>>>   	__u32 info;
>>>   	/* "size" is used by INT, ENUM, STRUCT and UNION.
>>> @@ -52,6 +54,7 @@ struct btf_type {
>>>   
>>>   #define BTF_INFO_KIND(info)	(((info) >> 24) & 0x0f)
>>>   #define BTF_INFO_VLEN(info)	((info) & 0xffff)
>>> +#define BTF_INFO_KFLAG(info)	((info) >> 31)
>>>   
>>>   #define BTF_KIND_UNKN		0	/* Unknown	*/
>>>   #define BTF_KIND_INT		1	/* Integer	*/
>>> @@ -110,9 +113,17 @@ struct btf_array {
>>>   struct btf_member {
>>>   	__u32	name_off;
>>>   	__u32	type;
>>> -	__u32	offset;	/* offset in bits */
>>> +	__u32	offset;	/* [bitfield_size and] offset in bits */
>>>   };
>>>   
>>> +/* If the type info kind_flag set, the btf_member.offset
>>> + * contains both member bit offset and bitfield size, and
>>> + * bitfield size will set for struct/union bitfield members.
>>> + * Otherwise, it contains only bit offset.
>>> + */
>> nit. It may be better to move this comment to the btf_member.offset
>> above.
>>
>>> +#define BTF_MEMBER_BITFIELD_SIZE(val)	((val) >> 24)
>>> +#define BTF_MEMBER_BIT_OFFSET(val)	((val) & 0xffffff)
>> After re-thinking this setup again, I still think
>> having these macros in btf.h to also do the kflag checking
>> would be nice.
>>
>> Unlike BTF_INFO_KIND() and BTF_INT_ENCODING() which don't
>> depend on other facts,  the btf.h raw user must check kflag
>> anyway before calling BTF_MEMBER_BIT*().
>> Forcing a kflag check before the user can access these convenient
>> 0xfffff and >>24 conversions may enforce this kflag check to
>> some extend.
>>
>> Since it is in uapi, it will not be easy to change later.
>> The above concern could be overkill ;), just want to ensure
>> it has been thought through a bit more here.
>>
>> It could be as easy as moving the new btf_member_bit*() from
>> btf.c to here and remove these two macros (or move them back to btf.c).
> 
> I think moving:
> +static u32 btf_member_bitfield_size(const struct btf_type *struct_type,
> +                                   const struct btf_member *member)
> +{
> +       return btf_type_kflag(struct_type) ? BTF_MEMBER_BITFIELD_SIZE(member->offset)
> +                                          : 0;
> +}
> 
> into uapi/btf.h may or may not be useful for btf uapi users.
> What are the chances that these static inline helpers will be
> reused by BTF logic in libbpf or other libs?
> At this point we don't know.
> So I would keep btf.h minimal.
> I agree that BTF_MEMBER_BIT_OFFSET() shouldn't be reused blindly.
> The users have to do BTF_INFO_KFLAG() check first.
> But this is the case for pretty much all of BTF data structures.
> In that sense BTF_MEMBER_BITFIELD_SIZE/BTF_MEMBER_BIT_OFFSET
> serve as a documentation and fit the style of btf.h header.
> imo having these two macros in uapi/btf.h is better than
> replacing them with comment only.
> 
> After this set the whole BTF really needs to be documented.

Yes，Martin and I will work on the documentation very soon.

Powered by blists - more mailing lists