btf.rst

BPF Type Format (BTF)

1. Introduction

BTF (BPF Type Format) is the metadata format which encodes the debug info related to BPF program/map. The name BTF was used initially to describe data types. The BTF was later extended to include function info for defined subroutines, and line info for source/line information.

The debug info is used for map pretty print, function signature, etc. The function signature enables better bpf program/function kernel symbol. The line info helps generate source annotated translated byte code, jited code and verifier log.

The BTF specification contains two parts,

• BTF kernel API
• BTF ELF file format

The kernel API is the contract between user space and kernel. The kernel verifies the BTF info before using it. The ELF file format is a user space contract between ELF file and libbpf loader.

The type and string sections are part of the BTF kernel API, describing the debug info (mostly types related) referenced by the bpf program. These two sections are discussed in details in :ref:`BTF_Type_String`.

2. BTF Type and String Encoding

The file include/uapi/linux/btf.h provides high-level definition of how types/strings are encoded.

The beginning of data blob must be:

struct btf_header {
    __u16   magic;
    __u8    version;
    __u8    flags;
    __u32   hdr_len;

    /* All offsets are in bytes relative to the end of this header */
    __u32   type_off;       /* offset of type section       */
    __u32   type_len;       /* length of type section       */
    __u32   str_off;        /* offset of string section     */
    __u32   str_len;        /* length of string section     */
};

The magic is 0xeB9F, which has different encoding for big and little endian systems, and can be used to test whether BTF is generated for big- or little-endian target. The btf_header is designed to be extensible with hdr_len equal to sizeof(struct btf_header) when a data blob is generated.

2.1 String Encoding

The first string in the string section must be a null string. The rest of string table is a concatenation of other null-terminated strings.

2.2 Type Encoding

The type id 0 is reserved for void type. The type section is parsed sequentially and type id is assigned to each recognized type starting from id 1. Currently, the following types are supported:

#define BTF_KIND_INT            1       /* Integer      */
#define BTF_KIND_PTR            2       /* Pointer      */
#define BTF_KIND_ARRAY          3       /* Array        */
#define BTF_KIND_STRUCT         4       /* Struct       */
#define BTF_KIND_UNION          5       /* Union        */
#define BTF_KIND_ENUM           6       /* Enumeration  */
#define BTF_KIND_FWD            7       /* Forward      */
#define BTF_KIND_TYPEDEF        8       /* Typedef      */
#define BTF_KIND_VOLATILE       9       /* Volatile     */
#define BTF_KIND_CONST          10      /* Const        */
#define BTF_KIND_RESTRICT       11      /* Restrict     */
#define BTF_KIND_FUNC           12      /* Function     */
#define BTF_KIND_FUNC_PROTO     13      /* Function Proto       */

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.

Each type contains the following common data:

struct btf_type {
    __u32 name_off;
    /* "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
     * bit     31: kind_flag, currently used by
     *             struct, union and fwd
     */
    __u32 info;
    /* "size" is used by INT, ENUM, STRUCT and UNION.
     * "size" tells the size of the type it is describing.
     *
     * "type" is used by PTR, TYPEDEF, VOLATILE, CONST, RESTRICT,
     * FUNC and FUNC_PROTO.
     * "type" is a type_id referring to another type.
     */
    union {
            __u32 size;
            __u32 type;
    };
};

For certain kinds, the common data are followed by kind-specific data. The name_off in struct btf_type specifies the offset in the string table. The following sections detail encoding of each kind.

2.2.1 BTF_KIND_INT

struct btf_type encoding requirement:

• name_off: any valid offset
• info.kind_flag: 0
• info.kind: BTF_KIND_INT
• info.vlen: 0
• size: the size of the int type in bytes.

btf_type is followed by a u32 with the following bits arrangement:

#define BTF_INT_ENCODING(VAL)   (((VAL) & 0x0f000000) >> 24)
#define BTF_INT_OFFSET(VAL)     (((VAL  & 0x00ff0000)) >> 16)
#define BTF_INT_BITS(VAL)       ((VAL)  & 0x000000ff)

The BTF_INT_ENCODING has the following attributes:

#define BTF_INT_SIGNED  (1 << 0)
#define BTF_INT_CHAR    (1 << 1)
#define BTF_INT_BOOL    (1 << 2)

The BTF_INT_ENCODING() provides extra information: signedness, char, or bool, for the int type. The char and bool encoding are mostly useful for pretty print. At most one encoding can be specified for the int type.

The BTF_INT_BITS() specifies the number of actual bits held by this int type. For example, a 4-bit bitfield encodes BTF_INT_BITS() equals to 4. The btf_type.size * 8 must be equal to or greater than BTF_INT_BITS() for the type. The maximum value of BTF_INT_BITS() is 128.

The BTF_INT_OFFSET() specifies the starting bit offset to calculate values for this int. For example, a bitfield struct member has:

• btf member bit offset 100 from the start of the structure,
• btf member pointing to an int type,
• the int type has BTF_INT_OFFSET() = 2 and BTF_INT_BITS() = 4

Then in the struct memory layout, this member will occupy 4 bits starting from bits 100 + 2 = 102.

Alternatively, the bitfield struct member can be the following to access the same bits as the above:

• btf member bit offset 102,
• btf member pointing to an int type,
• the int type has BTF_INT_OFFSET() = 0 and BTF_INT_BITS() = 4

The original intention of BTF_INT_OFFSET() is to provide flexibility of bitfield encoding. Currently, both llvm and pahole generate BTF_INT_OFFSET() = 0 for all int types.

2.2.2 BTF_KIND_PTR

struct btf_type encoding requirement:

• name_off: 0
• info.kind_flag: 0
• info.kind: BTF_KIND_PTR
• info.vlen: 0
• type: the pointee type of the pointer

No additional type data follow btf_type.

2.2.3 BTF_KIND_ARRAY

struct btf_type encoding requirement:

• name_off: 0
• info.kind_flag: 0
• info.kind: BTF_KIND_ARRAY
• info.vlen: 0
• size/type: 0, not used

btf_type is followed by one struct btf_array:

struct btf_array {
    __u32   type;
    __u32   index_type;
    __u32   nelems;
};

The struct btf_array encoding:

• type: the element type
• index_type: the index type
• nelems: the number of elements for this array (0 is also allowed).

The index_type can be any regular int type (u8, u16, u32, u64, unsigned __int128). The original design of including index_type follows DWARF, which has an index_type for its array type. Currently in BTF, beyond type verification, the index_type is not used.

The struct btf_array allows chaining through element type to represent multidimensional arrays. For example, for int a[5][6], the following type information illustrates the chaining:

• [1]: int
• [2]: array, btf_array.type = [1], btf_array.nelems = 6
• [3]: array, btf_array.type = [2], btf_array.nelems = 5

Currently, both pahole and llvm collapse multidimensional array into one-dimensional array, e.g., for a[5][6], the btf_array.nelems is equal to 30. This is because the original use case is map pretty print where the whole array is dumped out so one-dimensional array is enough. As more BTF usage is explored, pahole and llvm can be changed to generate proper chained representation for multidimensional arrays.

2.2.4 BTF_KIND_STRUCT

2.2.5 BTF_KIND_UNION

struct btf_type encoding requirement:

• name_off: 0 or offset to a valid C identifier
• info.kind_flag: 0 or 1
• info.kind: BTF_KIND_STRUCT or BTF_KIND_UNION
• info.vlen: the number of struct/union members
• info.size: the size of the struct/union in bytes

btf_type is followed by info.vlen number of struct btf_member.:

struct btf_member {
    __u32   name_off;
    __u32   type;
    __u32   offset;
};

struct btf_member encoding:

• name_off: offset to a valid C identifier
• type: the member type
• offset: <see below>

If the type info kind_flag is not set, the offset contains only bit offset of the member. Note that the base type of the bitfield can only be int or enum type. If the bitfield size is 32, the base type can be either int or enum type. If the bitfield size is not 32, the base type must be int, and int type BTF_INT_BITS() encodes the bitfield size.

If the kind_flag is set, the btf_member.offset contains both member bitfield size and bit offset. The bitfield size and bit offset are calculated as below.:

#define BTF_MEMBER_BITFIELD_SIZE(val)   ((val) >> 24)
#define BTF_MEMBER_BIT_OFFSET(val)      ((val) & 0xffffff)

In this case, if the base type is an int type, it must be a regular int type:

• BTF_INT_OFFSET() must be 0.
• BTF_INT_BITS() must be equal to {1,2,4,8,16} * 8.

The following kernel patch introduced kind_flag and explained why both modes exist:

https://github.com/torvalds/linux/commit/9d5f9f701b1891466fb3dbb1806ad97716f95cc3#diff-fa650a64fdd3968396883d2fe8215ff3

2.2.6 BTF_KIND_ENUM

struct btf_type encoding requirement:

• name_off: 0 or offset to a valid C identifier
• info.kind_flag: 0
• info.kind: BTF_KIND_ENUM
• info.vlen: number of enum values
• size: 4

btf_type is followed by info.vlen number of struct btf_enum.:

struct btf_enum {
    __u32   name_off;
    __s32   val;
};

The btf_enum encoding:

• name_off: offset to a valid C identifier
• val: any value

2.2.7 BTF_KIND_FWD

struct btf_type encoding requirement:

• name_off: offset to a valid C identifier
• info.kind_flag: 0 for struct, 1 for union
• info.kind: BTF_KIND_FWD
• info.vlen: 0
• type: 0

No additional type data follow btf_type.

2.2.8 BTF_KIND_TYPEDEF

struct btf_type encoding requirement:

• name_off: offset to a valid C identifier
• info.kind_flag: 0
• info.kind: BTF_KIND_TYPEDEF
• info.vlen: 0
• type: the type which can be referred by name at name_off

No additional type data follow btf_type.

2.2.9 BTF_KIND_VOLATILE

struct btf_type encoding requirement:

• name_off: 0
• info.kind_flag: 0
• info.kind: BTF_KIND_VOLATILE
• info.vlen: 0
• type: the type with volatile qualifier

No additional type data follow btf_type.

2.2.10 BTF_KIND_CONST

struct btf_type encoding requirement:

• name_off: 0
• info.kind_flag: 0
• info.kind: BTF_KIND_CONST
• info.vlen: 0
• type: the type with const qualifier

No additional type data follow btf_type.

2.2.11 BTF_KIND_RESTRICT

struct btf_type encoding requirement:

• name_off: 0
• info.kind_flag: 0
• info.kind: BTF_KIND_RESTRICT
• info.vlen: 0
• type: the type with restrict qualifier

No additional type data follow btf_type.

2.2.12 BTF_KIND_FUNC

struct btf_type encoding requirement:

• name_off: offset to a valid C identifier
• info.kind_flag: 0
• info.kind: BTF_KIND_FUNC
• info.vlen: 0
• type: a BTF_KIND_FUNC_PROTO type

No additional type data follow btf_type.

A BTF_KIND_FUNC defines not a type, but a subprogram (function) whose signature is defined by type. The subprogram is thus an instance of that type. The BTF_KIND_FUNC may in turn be referenced by a func_info in the :ref:`BTF_Ext_Section` (ELF) or in the arguments to :ref:`BPF_Prog_Load` (ABI).

2.2.13 BTF_KIND_FUNC_PROTO

struct btf_type encoding requirement:

• name_off: 0
• info.kind_flag: 0
• info.kind: BTF_KIND_FUNC_PROTO
• info.vlen: # of parameters
• type: the return type

btf_type is followed by info.vlen number of struct btf_param.:

struct btf_param {
    __u32   name_off;
    __u32   type;
};

If a BTF_KIND_FUNC_PROTO type is referred by a BTF_KIND_FUNC type, then btf_param.name_off must point to a valid C identifier except for the possible last argument representing the variable argument. The btf_param.type refers to parameter type.

If the function has variable arguments, the last parameter is encoded with name_off = 0 and type = 0.

3. BTF Kernel API

The following bpf syscall command involves BTF:

• BPF_BTF_LOAD: load a blob of BTF data into kernel
• BPF_MAP_CREATE: map creation with btf key and value type info.
• BPF_PROG_LOAD: prog load with btf function and line info.
• BPF_BTF_GET_FD_BY_ID: get a btf fd
• BPF_OBJ_GET_INFO_BY_FD: btf, func_info, line_info and other btf related info are returned.

The workflow typically looks like:

Application:
    BPF_BTF_LOAD
        |
        v
    BPF_MAP_CREATE and BPF_PROG_LOAD
        |
        V
    ......

Introspection tool:
    ......
    BPF_{PROG,MAP}_GET_NEXT_ID (get prog/map id's)
        |
        V
    BPF_{PROG,MAP}_GET_FD_BY_ID (get a prog/map fd)
        |
        V
    BPF_OBJ_GET_INFO_BY_FD (get bpf_prog_info/bpf_map_info with btf_id)
        |                                     |
        V                                     |
    BPF_BTF_GET_FD_BY_ID (get btf_fd)         |
        |                                     |
        V                                     |
    BPF_OBJ_GET_INFO_BY_FD (get btf)          |
        |                                     |
        V                                     V
    pretty print types, dump func signatures and line info, etc.

3.1 BPF_BTF_LOAD

Load a blob of BTF data into kernel. A blob of data, described in :ref:`BTF_Type_String`, can be directly loaded into the kernel. A btf_fd is returned to a userspace.

3.2 BPF_MAP_CREATE

A map can be created with btf_fd and specified key/value type id.:

__u32   btf_fd;         /* fd pointing to a BTF type data */
__u32   btf_key_type_id;        /* BTF type_id of the key */
__u32   btf_value_type_id;      /* BTF type_id of the value */

In libbpf, the map can be defined with extra annotation like below:

struct bpf_map_def SEC("maps") btf_map = {
    .type = BPF_MAP_TYPE_ARRAY,
    .key_size = sizeof(int),
    .value_size = sizeof(struct ipv_counts),
    .max_entries = 4,
};
BPF_ANNOTATE_KV_PAIR(btf_map, int, struct ipv_counts);

Here, the parameters for macro BPF_ANNOTATE_KV_PAIR are map name, key and value types for the map. During ELF parsing, libbpf is able to extract key/value type_id's and assign them to BPF_MAP_CREATE attributes automatically.

3.3 BPF_PROG_LOAD

During prog_load, func_info and line_info can be passed to kernel with proper values for the following attributes:

__u32           insn_cnt;
__aligned_u64   insns;
......
__u32           prog_btf_fd;    /* fd pointing to BTF type data */
__u32           func_info_rec_size;     /* userspace bpf_func_info size */
__aligned_u64   func_info;      /* func info */
__u32           func_info_cnt;  /* number of bpf_func_info records */
__u32           line_info_rec_size;     /* userspace bpf_line_info size */
__aligned_u64   line_info;      /* line info */
__u32           line_info_cnt;  /* number of bpf_line_info records */

The func_info and line_info are an array of below, respectively.:

struct bpf_func_info {
    __u32   insn_off; /* [0, insn_cnt - 1] */
    __u32   type_id;  /* pointing to a BTF_KIND_FUNC type */
};
struct bpf_line_info {
    __u32   insn_off; /* [0, insn_cnt - 1] */
    __u32   file_name_off; /* offset to string table for the filename */
    __u32   line_off; /* offset to string table for the source line */
    __u32   line_col; /* line number and column number */
};

func_info_rec_size is the size of each func_info record, and line_info_rec_size is the size of each line_info record. Passing the record size to kernel make it possible to extend the record itself in the future.

Below are requirements for func_info:

• func_info[0].insn_off must be 0.
• the func_info insn_off is in strictly increasing order and matches bpf func boundaries.

Below are requirements for line_info:

• the first insn in each func must have a line_info record pointing to it.
• the line_info insn_off is in strictly increasing order.

For line_info, the line number and column number are defined as below:

#define BPF_LINE_INFO_LINE_NUM(line_col)        ((line_col) >> 10)
#define BPF_LINE_INFO_LINE_COL(line_col)        ((line_col) & 0x3ff)

3.4 BPF_{PROG,MAP}_GET_NEXT_ID

In kernel, every loaded program, map or btf has a unique id. The id won't change during the lifetime of a program, map, or btf.

The bpf syscall command BPF_{PROG,MAP}_GET_NEXT_ID returns all id's, one for each command, to user space, for bpf program or maps, respectively, so an inspection tool can inspect all programs and maps.

3.5 BPF_{PROG,MAP}_GET_FD_BY_ID

An introspection tool cannot use id to get details about program or maps. A file descriptor needs to be obtained first for reference-counting purpose.

3.6 BPF_OBJ_GET_INFO_BY_FD

Once a program/map fd is acquired, an introspection tool can get the detailed information from kernel about this fd, some of which are BTF-related. For example, bpf_map_info returns btf_id and key/value type ids. bpf_prog_info returns btf_id, func_info, and line info for translated bpf byte codes, and jited_line_info.

3.7 BPF_BTF_GET_FD_BY_ID

With btf_id obtained in bpf_map_info and bpf_prog_info, bpf syscall command BPF_BTF_GET_FD_BY_ID can retrieve a btf fd. Then, with command BPF_OBJ_GET_INFO_BY_FD, the btf blob, originally loaded into the kernel with BPF_BTF_LOAD, can be retrieved.

With the btf blob, bpf_map_info, and bpf_prog_info, an introspection tool has full btf knowledge and is able to pretty print map key/values, dump func signatures and line info, along with byte/jit codes.

4. ELF File Format Interface

4.1 .BTF section

The .BTF section contains type and string data. The format of this section is same as the one describe in :ref:`BTF_Type_String`.

4.2 .BTF.ext section

The .BTF.ext section encodes func_info and line_info which needs loader manipulation before loading into the kernel.

The specification for .BTF.ext section is defined at tools/lib/bpf/btf.h and tools/lib/bpf/btf.c.

The current header of .BTF.ext section:

struct btf_ext_header {
    __u16   magic;
    __u8    version;
    __u8    flags;
    __u32   hdr_len;

    /* All offsets are in bytes relative to the end of this header */
    __u32   func_info_off;
    __u32   func_info_len;
    __u32   line_info_off;
    __u32   line_info_len;
};

It is very similar to .BTF section. Instead of type/string section, it contains func_info and line_info section. See :ref:`BPF_Prog_Load` for details about func_info and line_info record format.

The func_info is organized as below.:

func_info_rec_size
btf_ext_info_sec for section #1 /* func_info for section #1 */
btf_ext_info_sec for section #2 /* func_info for section #2 */
...

func_info_rec_size specifies the size of bpf_func_info structure when .BTF.ext is generated. btf_ext_info_sec, defined below, is a collection of func_info for each specific ELF section.:

struct btf_ext_info_sec {
   __u32   sec_name_off; /* offset to section name */
   __u32   num_info;
   /* Followed by num_info * record_size number of bytes */
   __u8    data[0];
};

Here, num_info must be greater than 0.

The line_info is organized as below.:

line_info_rec_size
btf_ext_info_sec for section #1 /* line_info for section #1 */
btf_ext_info_sec for section #2 /* line_info for section #2 */
...

line_info_rec_size specifies the size of bpf_line_info structure when .BTF.ext is generated.

The interpretation of bpf_func_info->insn_off and bpf_line_info->insn_off is different between kernel API and ELF API. For kernel API, the insn_off is the instruction offset in the unit of struct bpf_insn. For ELF API, the insn_off is the byte offset from the beginning of section (btf_ext_info_sec->sec_name_off).

5. Using BTF

5.1 bpftool map pretty print

With BTF, the map key/value can be printed based on fields rather than simply raw bytes. This is especially valuable for large structure or if your data structure has bitfields. For example, for the following map,: