Documentation/bpf/btf.rst

   1 =====================
   2 BPF Type Format (BTF)
   3 =====================
   4
   5 1. Introduction
   6 ***************
   7
   8 BTF (BPF Type Format) is the metadata format which encodes the debug info
   9 related to BPF program/map. The name BTF was used initially to describe data
  10 types. The BTF was later extended to include function info for defined
  11 subroutines, and line info for source/line information.
  12
  13 The debug info is used for map pretty print, function signature, etc. The
  14 function signature enables better bpf program/function kernel symbol. The line
  15 info helps generate source annotated translated byte code, jited code and
  16 verifier log.
  17
  18 The BTF specification contains two parts,
  19   * BTF kernel API
  20   * BTF ELF file format
  21
  22 The kernel API is the contract between user space and kernel. The kernel
  23 verifies the BTF info before using it. The ELF file format is a user space
  24 contract between ELF file and libbpf loader.
  25
  26 The type and string sections are part of the BTF kernel API, describing the
  27 debug info (mostly types related) referenced by the bpf program. These two
  28 sections are discussed in details in :ref:`BTF_Type_String`.
  29
  30 .. _BTF_Type_String:
  31
  32 2. BTF Type and String Encoding
  33 *******************************
  34
  35 The file ``include/uapi/linux/btf.h`` provides high-level definition of how
  36 types/strings are encoded.
  37
  38 The beginning of data blob must be::
  39
  40     struct btf_header {
  41         __u16   magic;
  42         __u8    version;
  43         __u8    flags;
  44         __u32   hdr_len;
  45
  46         /* All offsets are in bytes relative to the end of this header */
  47         __u32   type_off;       /* offset of type section       */
  48         __u32   type_len;       /* length of type section       */
  49         __u32   str_off;        /* offset of string section     */
  50         __u32   str_len;        /* length of string section     */
  51     };
  52
  53 The magic is ``0xeB9F``, which has different encoding for big and little
  54 endian systems, and can be used to test whether BTF is generated for big- or
  55 little-endian target. The ``btf_header`` is designed to be extensible with
  56 ``hdr_len`` equal to ``sizeof(struct btf_header)`` when a data blob is
  57 generated.
  58
  59 2.1 String Encoding
  60 ===================
  61
  62 The first string in the string section must be a null string. The rest of
  63 string table is a concatenation of other null-terminated strings.
  64
  65 2.2 Type Encoding
  66 =================
  67
  68 The type id ``0`` is reserved for ``void`` type. The type section is parsed
  69 sequentially and type id is assigned to each recognized type starting from id
  70 ``1``. Currently, the following types are supported::
  71
  72     #define BTF_KIND_INT            1       /* Integer      */
  73     #define BTF_KIND_PTR            2       /* Pointer      */
  74     #define BTF_KIND_ARRAY          3       /* Array        */
  75     #define BTF_KIND_STRUCT         4       /* Struct       */
  76     #define BTF_KIND_UNION          5       /* Union        */
  77     #define BTF_KIND_ENUM           6       /* Enumeration  */
  78     #define BTF_KIND_FWD            7       /* Forward      */
  79     #define BTF_KIND_TYPEDEF        8       /* Typedef      */
  80     #define BTF_KIND_VOLATILE       9       /* Volatile     */
  81     #define BTF_KIND_CONST          10      /* Const        */
  82     #define BTF_KIND_RESTRICT       11      /* Restrict     */
  83     #define BTF_KIND_FUNC           12      /* Function     */
  84     #define BTF_KIND_FUNC_PROTO     13      /* Function Proto       */
  85
  86 Note that the type section encodes debug info, not just pure types.
  87 ``BTF_KIND_FUNC`` is not a type, and it represents a defined subprogram.
  88
  89 Each type contains the following common data::
  90
  91     struct btf_type {
  92         __u32 name_off;
  93         /* "info" bits arrangement
  94          * bits  0-15: vlen (e.g. # of struct's members)
  95          * bits 16-23: unused
  96          * bits 24-27: kind (e.g. int, ptr, array...etc)
  97          * bits 28-30: unused
  98          * bit     31: kind_flag, currently used by
  99          *             struct, union and fwd
 100          */
 101         __u32 info;
 102         /* "size" is used by INT, ENUM, STRUCT and UNION.
 103          * "size" tells the size of the type it is describing.
 104          *
 105          * "type" is used by PTR, TYPEDEF, VOLATILE, CONST, RESTRICT,
 106          * FUNC and FUNC_PROTO.
 107          * "type" is a type_id referring to another type.
 108          */
 109         union {
 110                 __u32 size;
 111                 __u32 type;
 112         };
 113     };
 114
 115 For certain kinds, the common data are followed by kind-specific data. The
 116 ``name_off`` in ``struct btf_type`` specifies the offset in the string table.
 117 The following sections detail encoding of each kind.
 118
 119 2.2.1 BTF_KIND_INT
 120 ~~~~~~~~~~~~~~~~~~
 121
 122 ``struct btf_type`` encoding requirement:
 123  * ``name_off``: any valid offset
 124  * ``info.kind_flag``: 0
 125  * ``info.kind``: BTF_KIND_INT
 126  * ``info.vlen``: 0
 127  * ``size``: the size of the int type in bytes.
 128
 129 ``btf_type`` is followed by a ``u32`` with the following bits arrangement::
 130
 131   #define BTF_INT_ENCODING(VAL)   (((VAL) & 0x0f000000) >> 24)
 132   #define BTF_INT_OFFSET(VAL)     (((VAL  & 0x00ff0000)) >> 16)
 133   #define BTF_INT_BITS(VAL)       ((VAL)  & 0x000000ff)
 134
 135 The ``BTF_INT_ENCODING`` has the following attributes::
 136
 137   #define BTF_INT_SIGNED  (1 << 0)
 138   #define BTF_INT_CHAR    (1 << 1)
 139   #define BTF_INT_BOOL    (1 << 2)
 140
 141 The ``BTF_INT_ENCODING()`` provides extra information: signedness, char, or
 142 bool, for the int type. The char and bool encoding are mostly useful for
 143 pretty print. At most one encoding can be specified for the int type.
 144
 145 The ``BTF_INT_BITS()`` specifies the number of actual bits held by this int
 146 type. For example, a 4-bit bitfield encodes ``BTF_INT_BITS()`` equals to 4.
 147 The ``btf_type.size * 8`` must be equal to or greater than ``BTF_INT_BITS()``
 148 for the type. The maximum value of ``BTF_INT_BITS()`` is 128.
 149
 150 The ``BTF_INT_OFFSET()`` specifies the starting bit offset to calculate values
 151 for this int. For example, a bitfield struct member has: * btf member bit
 152 offset 100 from the start of the structure, * btf member pointing to an int
 153 type, * the int type has ``BTF_INT_OFFSET() = 2`` and ``BTF_INT_BITS() = 4``
 154
 155 Then in the struct memory layout, this member will occupy ``4`` bits starting
 156 from bits ``100 + 2 = 102``.
 157
 158 Alternatively, the bitfield struct member can be the following to access the
 159 same bits as the above:
 160
 161  * btf member bit offset 102,
 162  * btf member pointing to an int type,
 163  * the int type has ``BTF_INT_OFFSET() = 0`` and ``BTF_INT_BITS() = 4``
 164
 165 The original intention of ``BTF_INT_OFFSET()`` is to provide flexibility of
 166 bitfield encoding. Currently, both llvm and pahole generate
 167 ``BTF_INT_OFFSET() = 0`` for all int types.
 168
 169 2.2.2 BTF_KIND_PTR
 170 ~~~~~~~~~~~~~~~~~~
 171
 172 ``struct btf_type`` encoding requirement:
 173   * ``name_off``: 0
 174   * ``info.kind_flag``: 0
 175   * ``info.kind``: BTF_KIND_PTR
 176   * ``info.vlen``: 0
 177   * ``type``: the pointee type of the pointer
 178
 179 No additional type data follow ``btf_type``.
 180
 181 2.2.3 BTF_KIND_ARRAY
 182 ~~~~~~~~~~~~~~~~~~~~
 183
 184 ``struct btf_type`` encoding requirement:
 185   * ``name_off``: 0
 186   * ``info.kind_flag``: 0
 187   * ``info.kind``: BTF_KIND_ARRAY
 188   * ``info.vlen``: 0
 189   * ``size/type``: 0, not used
 190
 191 ``btf_type`` is followed by one ``struct btf_array``::
 192
 193     struct btf_array {
 194         __u32   type;
 195         __u32   index_type;
 196         __u32   nelems;
 197     };
 198
 199 The ``struct btf_array`` encoding:
 200   * ``type``: the element type
 201   * ``index_type``: the index type
 202   * ``nelems``: the number of elements for this array (``0`` is also allowed).
 203
 204 The ``index_type`` can be any regular int type (``u8``, ``u16``, ``u32``,
 205 ``u64``, ``unsigned __int128``). The original design of including
 206 ``index_type`` follows DWARF, which has an ``index_type`` for its array type.
 207 Currently in BTF, beyond type verification, the ``index_type`` is not used.
 208
 209 The ``struct btf_array`` allows chaining through element type to represent
 210 multidimensional arrays. For example, for ``int a[5][6]``, the following type
 211 information illustrates the chaining:
 212
 213   * [1]: int
 214   * [2]: array, ``btf_array.type = [1]``, ``btf_array.nelems = 6``
 215   * [3]: array, ``btf_array.type = [2]``, ``btf_array.nelems = 5``
 216
 217 Currently, both pahole and llvm collapse multidimensional array into
 218 one-dimensional array, e.g., for ``a[5][6]``, the ``btf_array.nelems`` is
 219 equal to ``30``. This is because the original use case is map pretty print
 220 where the whole array is dumped out so one-dimensional array is enough. As
 221 more BTF usage is explored, pahole and llvm can be changed to generate proper
 222 chained representation for multidimensional arrays.
 223
 224 2.2.4 BTF_KIND_STRUCT
 225 ~~~~~~~~~~~~~~~~~~~~~
 226 2.2.5 BTF_KIND_UNION
 227 ~~~~~~~~~~~~~~~~~~~~
 228
 229 ``struct btf_type`` encoding requirement:
 230   * ``name_off``: 0 or offset to a valid C identifier
 231   * ``info.kind_flag``: 0 or 1
 232   * ``info.kind``: BTF_KIND_STRUCT or BTF_KIND_UNION
 233   * ``info.vlen``: the number of struct/union members
 234   * ``info.size``: the size of the struct/union in bytes
 235
 236 ``btf_type`` is followed by ``info.vlen`` number of ``struct btf_member``.::
 237
 238     struct btf_member {
 239         __u32   name_off;
 240         __u32   type;
 241         __u32   offset;
 242     };
 243
 244 ``struct btf_member`` encoding:
 245   * ``name_off``: offset to a valid C identifier
 246   * ``type``: the member type
 247   * ``offset``: <see below>
 248
 249 If the type info ``kind_flag`` is not set, the offset contains only bit offset
 250 of the member. Note that the base type of the bitfield can only be int or enum
 251 type. If the bitfield size is 32, the base type can be either int or enum
 252 type. If the bitfield size is not 32, the base type must be int, and int type
 253 ``BTF_INT_BITS()`` encodes the bitfield size.
 254
 255 If the ``kind_flag`` is set, the ``btf_member.offset`` contains both member
 256 bitfield size and bit offset. The bitfield size and bit offset are calculated
 257 as below.::
 258
 259   #define BTF_MEMBER_BITFIELD_SIZE(val)   ((val) >> 24)
 260   #define BTF_MEMBER_BIT_OFFSET(val)      ((val) & 0xffffff)
 261
 262 In this case, if the base type is an int type, it must be a regular int type:
 263
 264   * ``BTF_INT_OFFSET()`` must be 0.
 265   * ``BTF_INT_BITS()`` must be equal to ``{1,2,4,8,16} * 8``.
 266
 267 The following kernel patch introduced ``kind_flag`` and explained why both
 268 modes exist:
 269
 270   https://github.com/torvalds/linux/commit/9d5f9f701b1891466fb3dbb1806ad97716f95cc3#diff-fa650a64fdd3968396883d2fe8215ff3
 271
 272 2.2.6 BTF_KIND_ENUM
 273 ~~~~~~~~~~~~~~~~~~~
 274
 275 ``struct btf_type`` encoding requirement:
 276   * ``name_off``: 0 or offset to a valid C identifier
 277   * ``info.kind_flag``: 0
 278   * ``info.kind``: BTF_KIND_ENUM
 279   * ``info.vlen``: number of enum values
 280   * ``size``: 4
 281
 282 ``btf_type`` is followed by ``info.vlen`` number of ``struct btf_enum``.::
 283
 284     struct btf_enum {
 285         __u32   name_off;
 286         __s32   val;
 287     };
 288
 289 The ``btf_enum`` encoding:
 290   * ``name_off``: offset to a valid C identifier
 291   * ``val``: any value
 292
 293 2.2.7 BTF_KIND_FWD
 294 ~~~~~~~~~~~~~~~~~~
 295
 296 ``struct btf_type`` encoding requirement:
 297   * ``name_off``: offset to a valid C identifier
 298   * ``info.kind_flag``: 0 for struct, 1 for union
 299   * ``info.kind``: BTF_KIND_FWD
 300   * ``info.vlen``: 0
 301   * ``type``: 0
 302
 303 No additional type data follow ``btf_type``.
 304
 305 2.2.8 BTF_KIND_TYPEDEF
 306 ~~~~~~~~~~~~~~~~~~~~~~
 307
 308 ``struct btf_type`` encoding requirement:
 309   * ``name_off``: offset to a valid C identifier
 310   * ``info.kind_flag``: 0
 311   * ``info.kind``: BTF_KIND_TYPEDEF
 312   * ``info.vlen``: 0
 313   * ``type``: the type which can be referred by name at ``name_off``
 314
 315 No additional type data follow ``btf_type``.
 316
 317 2.2.9 BTF_KIND_VOLATILE
 318 ~~~~~~~~~~~~~~~~~~~~~~~
 319
 320 ``struct btf_type`` encoding requirement:
 321   * ``name_off``: 0
 322   * ``info.kind_flag``: 0
 323   * ``info.kind``: BTF_KIND_VOLATILE
 324   * ``info.vlen``: 0
 325   * ``type``: the type with ``volatile`` qualifier
 326
 327 No additional type data follow ``btf_type``.
 328
 329 2.2.10 BTF_KIND_CONST
 330 ~~~~~~~~~~~~~~~~~~~~~
 331
 332 ``struct btf_type`` encoding requirement:
 333   * ``name_off``: 0
 334   * ``info.kind_flag``: 0
 335   * ``info.kind``: BTF_KIND_CONST
 336   * ``info.vlen``: 0
 337   * ``type``: the type with ``const`` qualifier
 338
 339 No additional type data follow ``btf_type``.
 340
 341 2.2.11 BTF_KIND_RESTRICT
 342 ~~~~~~~~~~~~~~~~~~~~~~~~
 343
 344 ``struct btf_type`` encoding requirement:
 345   * ``name_off``: 0
 346   * ``info.kind_flag``: 0
 347   * ``info.kind``: BTF_KIND_RESTRICT
 348   * ``info.vlen``: 0
 349   * ``type``: the type with ``restrict`` qualifier
 350
 351 No additional type data follow ``btf_type``.
 352
 353 2.2.12 BTF_KIND_FUNC
 354 ~~~~~~~~~~~~~~~~~~~~
 355
 356 ``struct btf_type`` encoding requirement:
 357   * ``name_off``: offset to a valid C identifier
 358   * ``info.kind_flag``: 0
 359   * ``info.kind``: BTF_KIND_FUNC
 360   * ``info.vlen``: 0
 361   * ``type``: a BTF_KIND_FUNC_PROTO type
 362
 363 No additional type data follow ``btf_type``.
 364
 365 A BTF_KIND_FUNC defines not a type, but a subprogram (function) whose
 366 signature is defined by ``type``. The subprogram is thus an instance of that
 367 type. The BTF_KIND_FUNC may in turn be referenced by a func_info in the
 368 :ref:`BTF_Ext_Section` (ELF) or in the arguments to :ref:`BPF_Prog_Load`
 369 (ABI).
 370
 371 2.2.13 BTF_KIND_FUNC_PROTO
 372 ~~~~~~~~~~~~~~~~~~~~~~~~~~
 373
 374 ``struct btf_type`` encoding requirement:
 375   * ``name_off``: 0
 376   * ``info.kind_flag``: 0
 377   * ``info.kind``: BTF_KIND_FUNC_PROTO
 378   * ``info.vlen``: # of parameters
 379   * ``type``: the return type
 380
 381 ``btf_type`` is followed by ``info.vlen`` number of ``struct btf_param``.::
 382
 383     struct btf_param {
 384         __u32   name_off;
 385         __u32   type;
 386     };
 387
 388 If a BTF_KIND_FUNC_PROTO type is referred by a BTF_KIND_FUNC type, then
 389 ``btf_param.name_off`` must point to a valid C identifier except for the
 390 possible last argument representing the variable argument. The btf_param.type
 391 refers to parameter type.
 392
 393 If the function has variable arguments, the last parameter is encoded with
 394 ``name_off = 0`` and ``type = 0``.
 395
 396 3. BTF Kernel API
 397 *****************
 398
 399 The following bpf syscall command involves BTF:
 400    * BPF_BTF_LOAD: load a blob of BTF data into kernel
 401    * BPF_MAP_CREATE: map creation with btf key and value type info.
 402    * BPF_PROG_LOAD: prog load with btf function and line info.
 403    * BPF_BTF_GET_FD_BY_ID: get a btf fd
 404    * BPF_OBJ_GET_INFO_BY_FD: btf, func_info, line_info
 405      and other btf related info are returned.
 406
 407 The workflow typically looks like:
 408 ::
 409
 410   Application:
 411       BPF_BTF_LOAD
 412           |
 413           v
 414       BPF_MAP_CREATE and BPF_PROG_LOAD
 415           |
 416           V
 417       ......
 418
 419   Introspection tool:
 420       ......
 421       BPF_{PROG,MAP}_GET_NEXT_ID (get prog/map id's)
 422           |
 423           V
 424       BPF_{PROG,MAP}_GET_FD_BY_ID (get a prog/map fd)
 425           |
 426           V
 427       BPF_OBJ_GET_INFO_BY_FD (get bpf_prog_info/bpf_map_info with btf_id)
 428           |                                     |
 429           V                                     |
 430       BPF_BTF_GET_FD_BY_ID (get btf_fd)         |
 431           |                                     |
 432           V                                     |
 433       BPF_OBJ_GET_INFO_BY_FD (get btf)          |
 434           |                                     |
 435           V                                     V
 436       pretty print types, dump func signatures and line info, etc.
 437
 438
 439 3.1 BPF_BTF_LOAD
 440 ================
 441
 442 Load a blob of BTF data into kernel. A blob of data, described in
 443 :ref:`BTF_Type_String`, can be directly loaded into the kernel. A ``btf_fd``
 444 is returned to a userspace.
 445
 446 3.2 BPF_MAP_CREATE
 447 ==================
 448
 449 A map can be created with ``btf_fd`` and specified key/value type id.::
 450
 451     __u32   btf_fd;         /* fd pointing to a BTF type data */
 452     __u32   btf_key_type_id;        /* BTF type_id of the key */
 453     __u32   btf_value_type_id;      /* BTF type_id of the value */
 454
 455 In libbpf, the map can be defined with extra annotation like below:
 456 ::
 457
 458     struct bpf_map_def SEC("maps") btf_map = {
 459         .type = BPF_MAP_TYPE_ARRAY,
 460         .key_size = sizeof(int),
 461         .value_size = sizeof(struct ipv_counts),
 462         .max_entries = 4,
 463     };
 464     BPF_ANNOTATE_KV_PAIR(btf_map, int, struct ipv_counts);
 465
 466 Here, the parameters for macro BPF_ANNOTATE_KV_PAIR are map name, key and
 467 value types for the map. During ELF parsing, libbpf is able to extract
 468 key/value type_id's and assign them to BPF_MAP_CREATE attributes
 469 automatically.
 470
 471 .. _BPF_Prog_Load:
 472
 473 3.3 BPF_PROG_LOAD
 474 =================
 475
 476 During prog_load, func_info and line_info can be passed to kernel with proper
 477 values for the following attributes:
 478 ::
 479
 480     __u32           insn_cnt;
 481     __aligned_u64   insns;
 482     ......
 483     __u32           prog_btf_fd;    /* fd pointing to BTF type data */
 484     __u32           func_info_rec_size;     /* userspace bpf_func_info size */
 485     __aligned_u64   func_info;      /* func info */
 486     __u32           func_info_cnt;  /* number of bpf_func_info records */
 487     __u32           line_info_rec_size;     /* userspace bpf_line_info size */
 488     __aligned_u64   line_info;      /* line info */
 489     __u32           line_info_cnt;  /* number of bpf_line_info records */
 490
 491 The func_info and line_info are an array of below, respectively.::
 492
 493     struct bpf_func_info {
 494         __u32   insn_off; /* [0, insn_cnt - 1] */
 495         __u32   type_id;  /* pointing to a BTF_KIND_FUNC type */
 496     };
 497     struct bpf_line_info {
 498         __u32   insn_off; /* [0, insn_cnt - 1] */
 499         __u32   file_name_off; /* offset to string table for the filename */
 500         __u32   line_off; /* offset to string table for the source line */
 501         __u32   line_col; /* line number and column number */
 502     };
 503
 504 func_info_rec_size is the size of each func_info record, and
 505 line_info_rec_size is the size of each line_info record. Passing the record
 506 size to kernel make it possible to extend the record itself in the future.
 507
 508 Below are requirements for func_info:
 509   * func_info[0].insn_off must be 0.
 510   * the func_info insn_off is in strictly increasing order and matches
 511     bpf func boundaries.
 512
 513 Below are requirements for line_info:
 514   * the first insn in each func must have a line_info record pointing to it.
 515   * the line_info insn_off is in strictly increasing order.
 516
 517 For line_info, the line number and column number are defined as below:
 518 ::
 519
 520     #define BPF_LINE_INFO_LINE_NUM(line_col)        ((line_col) >> 10)
 521     #define BPF_LINE_INFO_LINE_COL(line_col)        ((line_col) & 0x3ff)
 522
 523 3.4 BPF_{PROG,MAP}_GET_NEXT_ID
 524
 525 In kernel, every loaded program, map or btf has a unique id. The id won't
 526 change during the lifetime of a program, map, or btf.
 527
 528 The bpf syscall command BPF_{PROG,MAP}_GET_NEXT_ID returns all id's, one for
 529 each command, to user space, for bpf program or maps, respectively, so an
 530 inspection tool can inspect all programs and maps.
 531
 532 3.5 BPF_{PROG,MAP}_GET_FD_BY_ID
 533
 534 An introspection tool cannot use id to get details about program or maps.
 535 A file descriptor needs to be obtained first for reference-counting purpose.
 536
 537 3.6 BPF_OBJ_GET_INFO_BY_FD
 538 ==========================
 539
 540 Once a program/map fd is acquired, an introspection tool can get the detailed
 541 information from kernel about this fd, some of which are BTF-related. For
 542 example, ``bpf_map_info`` returns ``btf_id`` and key/value type ids.
 543 ``bpf_prog_info`` returns ``btf_id``, func_info, and line info for translated
 544 bpf byte codes, and jited_line_info.
 545
 546 3.7 BPF_BTF_GET_FD_BY_ID
 547 ========================
 548
 549 With ``btf_id`` obtained in ``bpf_map_info`` and ``bpf_prog_info``, bpf
 550 syscall command BPF_BTF_GET_FD_BY_ID can retrieve a btf fd. Then, with
 551 command BPF_OBJ_GET_INFO_BY_FD, the btf blob, originally loaded into the
 552 kernel with BPF_BTF_LOAD, can be retrieved.
 553
 554 With the btf blob, ``bpf_map_info``, and ``bpf_prog_info``, an introspection
 555 tool has full btf knowledge and is able to pretty print map key/values, dump
 556 func signatures and line info, along with byte/jit codes.
 557
 558 4. ELF File Format Interface
 559 ****************************
 560
 561 4.1 .BTF section
 562 ================
 563
 564 The .BTF section contains type and string data. The format of this section is
 565 same as the one describe in :ref:`BTF_Type_String`.
 566
 567 .. _BTF_Ext_Section:
 568
 569 4.2 .BTF.ext section
 570 ====================
 571
 572 The .BTF.ext section encodes func_info and line_info which needs loader
 573 manipulation before loading into the kernel.
 574
 575 The specification for .BTF.ext section is defined at ``tools/lib/bpf/btf.h``
 576 and ``tools/lib/bpf/btf.c``.
 577
 578 The current header of .BTF.ext section::
 579
 580     struct btf_ext_header {
 581         __u16   magic;
 582         __u8    version;
 583         __u8    flags;
 584         __u32   hdr_len;
 585
 586         /* All offsets are in bytes relative to the end of this header */
 587         __u32   func_info_off;
 588         __u32   func_info_len;
 589         __u32   line_info_off;
 590         __u32   line_info_len;
 591     };
 592
 593 It is very similar to .BTF section. Instead of type/string section, it
 594 contains func_info and line_info section. See :ref:`BPF_Prog_Load` for details
 595 about func_info and line_info record format.
 596
 597 The func_info is organized as below.::
 598
 599      func_info_rec_size
 600      btf_ext_info_sec for section #1 /* func_info for section #1 */
 601      btf_ext_info_sec for section #2 /* func_info for section #2 */
 602      ...
 603
 604 ``func_info_rec_size`` specifies the size of ``bpf_func_info`` structure when
 605 .BTF.ext is generated. ``btf_ext_info_sec``, defined below, is a collection of
 606 func_info for each specific ELF section.::
 607
 608      struct btf_ext_info_sec {
 609         __u32   sec_name_off; /* offset to section name */
 610         __u32   num_info;
 611         /* Followed by num_info * record_size number of bytes */
 612         __u8    data[0];
 613      };
 614
 615 Here, num_info must be greater than 0.
 616
 617 The line_info is organized as below.::
 618
 619      line_info_rec_size
 620      btf_ext_info_sec for section #1 /* line_info for section #1 */
 621      btf_ext_info_sec for section #2 /* line_info for section #2 */
 622      ...
 623
 624 ``line_info_rec_size`` specifies the size of ``bpf_line_info`` structure when
 625 .BTF.ext is generated.
 626
 627 The interpretation of ``bpf_func_info->insn_off`` and
 628 ``bpf_line_info->insn_off`` is different between kernel API and ELF API. For
 629 kernel API, the ``insn_off`` is the instruction offset in the unit of ``struct
 630 bpf_insn``. For ELF API, the ``insn_off`` is the byte offset from the
 631 beginning of section (``btf_ext_info_sec->sec_name_off``).
 632
 633 5. Using BTF
 634 ************
 635
 636 5.1 bpftool map pretty print
 637 ============================
 638
 639 With BTF, the map key/value can be printed based on fields rather than simply
 640 raw bytes. This is especially valuable for large structure or if your data
 641 structure has bitfields. For example, for the following map,::
 642
 643       enum A { A1, A2, A3, A4, A5 };
 644       typedef enum A ___A;
 645       struct tmp_t {
 646            char a1:4;
 647            int  a2:4;
 648            int  :4;
 649            __u32 a3:4;
 650            int b;
 651            ___A b1:4;
 652            enum A b2:4;
 653       };
 654       struct bpf_map_def SEC("maps") tmpmap = {
 655            .type = BPF_MAP_TYPE_ARRAY,
 656            .key_size = sizeof(__u32),
 657            .value_size = sizeof(struct tmp_t),
 658            .max_entries = 1,
 659       };
 660       BPF_ANNOTATE_KV_PAIR(tmpmap, int, struct tmp_t);
 661
 662 bpftool is able to pretty print like below:
 663 ::
 664
 665       [{
 666             "key": 0,
 667             "value": {
 668                 "a1": 0x2,
 669                 "a2": 0x4,
 670                 "a3": 0x6,
 671                 "b": 7,
 672                 "b1": 0x8,
 673                 "b2": 0xa
 674             }
 675         }
 676       ]
 677
 678 5.2 bpftool prog dump
 679 =====================
 680
 681 The following is an example showing how func_info and line_info can help prog
 682 dump with better kernel symbol names, function prototypes and line
 683 information.::
 684
 685     $ bpftool prog dump jited pinned /sys/fs/bpf/test_btf_haskv
 686     [...]
 687     int test_long_fname_2(struct dummy_tracepoint_args * arg):
 688     bpf_prog_44a040bf25481309_test_long_fname_2:
 689     ; static int test_long_fname_2(struct dummy_tracepoint_args *arg)
 690        0:   push   %rbp
 691        1:   mov    %rsp,%rbp
 692        4:   sub    $0x30,%rsp
 693        b:   sub    $0x28,%rbp
 694        f:   mov    %rbx,0x0(%rbp)
 695       13:   mov    %r13,0x8(%rbp)
 696       17:   mov    %r14,0x10(%rbp)
 697       1b:   mov    %r15,0x18(%rbp)
 698       1f:   xor    %eax,%eax
 699       21:   mov    %rax,0x20(%rbp)
 700       25:   xor    %esi,%esi
 701     ; int key = 0;
 702       27:   mov    %esi,-0x4(%rbp)
 703     ; if (!arg->sock)
 704       2a:   mov    0x8(%rdi),%rdi
 705     ; if (!arg->sock)
 706       2e:   cmp    $0x0,%rdi
 707       32:   je     0x0000000000000070
 708       34:   mov    %rbp,%rsi
 709     ; counts = bpf_map_lookup_elem(&btf_map, &key);
 710     [...]
 711
 712 5.3 Verifier Log
 713 ================
 714
 715 The following is an example of how line_info can help debugging verification
 716 failure.::
 717
 718        /* The code at tools/testing/selftests/bpf/test_xdp_noinline.c
 719         * is modified as below.
 720         */
 721        data = (void *)(long)xdp->data;
 722        data_end = (void *)(long)xdp->data_end;
 723        /*
 724        if (data + 4 > data_end)
 725                return XDP_DROP;
 726        */
 727        *(u32 *)data = dst->dst;
 728
 729     $ bpftool prog load ./test_xdp_noinline.o /sys/fs/bpf/test_xdp_noinline type xdp
 730         ; data = (void *)(long)xdp->data;
 731         224: (79) r2 = *(u64 *)(r10 -112)
 732         225: (61) r2 = *(u32 *)(r2 +0)
 733         ; *(u32 *)data = dst->dst;
 734         226: (63) *(u32 *)(r2 +0) = r1
 735         invalid access to packet, off=0 size=4, R2(id=0,off=0,r=0)
 736         R2 offset is outside of the packet
 737
 738 6. BTF Generation
 739 *****************
 740
 741 You need latest pahole
 742
 743   https://git.kernel.org/pub/scm/devel/pahole/pahole.git/
 744
 745 or llvm (8.0 or later). The pahole acts as a dwarf2btf converter. It doesn't
 746 support .BTF.ext and btf BTF_KIND_FUNC type yet. For example,::
 747
 748       -bash-4.4$ cat t.c
 749       struct t {
 750         int a:2;
 751         int b:3;
 752         int c:2;
 753       } g;
 754       -bash-4.4$ gcc -c -O2 -g t.c
 755       -bash-4.4$ pahole -JV t.o
 756       File t.o:
 757       [1] STRUCT t kind_flag=1 size=4 vlen=3
 758               a type_id=2 bitfield_size=2 bits_offset=0
 759               b type_id=2 bitfield_size=3 bits_offset=2
 760               c type_id=2 bitfield_size=2 bits_offset=5
 761       [2] INT int size=4 bit_offset=0 nr_bits=32 encoding=SIGNED
 762
 763 The llvm is able to generate .BTF and .BTF.ext directly with -g for bpf target
 764 only. The assembly code (-S) is able to show the BTF encoding in assembly
 765 format.::
 766
 767     -bash-4.4$ cat t2.c
 768     typedef int __int32;
 769     struct t2 {
 770       int a2;
 771       int (*f2)(char q1, __int32 q2, ...);
 772       int (*f3)();
 773     } g2;
 774     int main() { return 0; }
 775     int test() { return 0; }
 776     -bash-4.4$ clang -c -g -O2 -target bpf t2.c
 777     -bash-4.4$ readelf -S t2.o
 778       ......
 779       [ 8] .BTF              PROGBITS         0000000000000000  00000247
 780            000000000000016e  0000000000000000           0     0     1
 781       [ 9] .BTF.ext          PROGBITS         0000000000000000  000003b5
 782            0000000000000060  0000000000000000           0     0     1
 783       [10] .rel.BTF.ext      REL              0000000000000000  000007e0
 784            0000000000000040  0000000000000010          16     9     8
 785       ......
 786     -bash-4.4$ clang -S -g -O2 -target bpf t2.c
 787     -bash-4.4$ cat t2.s
 788       ......
 789             .section        .BTF,"",@progbits
 790             .short  60319                   # 0xeb9f
 791             .byte   1
 792             .byte   0
 793             .long   24
 794             .long   0
 795             .long   220
 796             .long   220
 797             .long   122
 798             .long   0                       # BTF_KIND_FUNC_PROTO(id = 1)
 799             .long   218103808               # 0xd000000
 800             .long   2
 801             .long   83                      # BTF_KIND_INT(id = 2)
 802             .long   16777216                # 0x1000000
 803             .long   4
 804             .long   16777248                # 0x1000020
 805       ......
 806             .byte   0                       # string offset=0
 807             .ascii  ".text"                 # string offset=1
 808             .byte   0
 809             .ascii  "/home/yhs/tmp-pahole/t2.c" # string offset=7
 810             .byte   0
 811             .ascii  "int main() { return 0; }" # string offset=33
 812             .byte   0
 813             .ascii  "int test() { return 0; }" # string offset=58
 814             .byte   0
 815             .ascii  "int"                   # string offset=83
 816       ......
 817             .section        .BTF.ext,"",@progbits
 818             .short  60319                   # 0xeb9f
 819             .byte   1
 820             .byte   0
 821             .long   24
 822             .long   0
 823             .long   28
 824             .long   28
 825             .long   44
 826             .long   8                       # FuncInfo
 827             .long   1                       # FuncInfo section string offset=1
 828             .long   2
 829             .long   .Lfunc_begin0
 830             .long   3
 831             .long   .Lfunc_begin1
 832             .long   5
 833             .long   16                      # LineInfo
 834             .long   1                       # LineInfo section string offset=1
 835             .long   2
 836             .long   .Ltmp0
 837             .long   7
 838             .long   33
 839             .long   7182                    # Line 7 Col 14
 840             .long   .Ltmp3
 841             .long   7
 842             .long   58
 843             .long   8206                    # Line 8 Col 14
 844
 845 7. Testing
 846 **********
 847
 848 Kernel bpf selftest `test_btf.c` provides extensive set of BTF-related tests.