xref: /linux/Documentation/bpf/standardization/instruction-set.rst (revision 39dc8b8ea387ce7f4fe2d2d6d550ed52aa9aa040)
1.. contents::
2.. sectnum::
3
4======================================
5BPF Instruction Set Architecture (ISA)
6======================================
7
8eBPF (which is no longer an acronym for anything), also commonly
9referred to as BPF, is a technology with origins in the Linux kernel
10that can run untrusted programs in a privileged context such as an
11operating system kernel. This document specifies the BPF instruction
12set architecture (ISA).
13
14Documentation conventions
15=========================
16
17For brevity and consistency, this document refers to families
18of types using a shorthand syntax and refers to several expository,
19mnemonic functions when describing the semantics of instructions.
20The range of valid values for those types and the semantics of those
21functions are defined in the following subsections.
22
23Types
24-----
25This document refers to integer types with the notation `SN` to specify
26a type's signedness (`S`) and bit width (`N`), respectively.
27
28.. table:: Meaning of signedness notation.
29
30  ==== =========
31  S    Meaning
32  ==== =========
33  u    unsigned
34  s    signed
35  ==== =========
36
37.. table:: Meaning of bit-width notation.
38
39  ===== =========
40  N     Bit width
41  ===== =========
42  8     8 bits
43  16    16 bits
44  32    32 bits
45  64    64 bits
46  128   128 bits
47  ===== =========
48
49For example, `u32` is a type whose valid values are all the 32-bit unsigned
50numbers and `s16` is a type whose valid values are all the 16-bit signed
51numbers.
52
53Functions
54---------
55* htobe16: Takes an unsigned 16-bit number in host-endian format and
56  returns the equivalent number as an unsigned 16-bit number in big-endian
57  format.
58* htobe32: Takes an unsigned 32-bit number in host-endian format and
59  returns the equivalent number as an unsigned 32-bit number in big-endian
60  format.
61* htobe64: Takes an unsigned 64-bit number in host-endian format and
62  returns the equivalent number as an unsigned 64-bit number in big-endian
63  format.
64* htole16: Takes an unsigned 16-bit number in host-endian format and
65  returns the equivalent number as an unsigned 16-bit number in little-endian
66  format.
67* htole32: Takes an unsigned 32-bit number in host-endian format and
68  returns the equivalent number as an unsigned 32-bit number in little-endian
69  format.
70* htole64: Takes an unsigned 64-bit number in host-endian format and
71  returns the equivalent number as an unsigned 64-bit number in little-endian
72  format.
73* bswap16: Takes an unsigned 16-bit number in either big- or little-endian
74  format and returns the equivalent number with the same bit width but
75  opposite endianness.
76* bswap32: Takes an unsigned 32-bit number in either big- or little-endian
77  format and returns the equivalent number with the same bit width but
78  opposite endianness.
79* bswap64: Takes an unsigned 64-bit number in either big- or little-endian
80  format and returns the equivalent number with the same bit width but
81  opposite endianness.
82
83
84Definitions
85-----------
86
87.. glossary::
88
89  Sign Extend
90    To `sign extend an` ``X`` `-bit number, A, to a` ``Y`` `-bit number, B  ,` means to
91
92    #. Copy all ``X`` bits from `A` to the lower ``X`` bits of `B`.
93    #. Set the value of the remaining ``Y`` - ``X`` bits of `B` to the value of
94       the  most-significant bit of `A`.
95
96.. admonition:: Example
97
98  Sign extend an 8-bit number ``A`` to a 16-bit number ``B`` on a big-endian platform:
99  ::
100
101    A:          10000110
102    B: 11111111 10000110
103
104Conformance groups
105------------------
106
107An implementation does not need to support all instructions specified in this
108document (e.g., deprecated instructions).  Instead, a number of conformance
109groups are specified.  An implementation must support the base32 conformance
110group and may support additional conformance groups, where supporting a
111conformance group means it must support all instructions in that conformance
112group.
113
114The use of named conformance groups enables interoperability between a runtime
115that executes instructions, and tools such as compilers that generate
116instructions for the runtime.  Thus, capability discovery in terms of
117conformance groups might be done manually by users or automatically by tools.
118
119Each conformance group has a short ASCII label (e.g., "base32") that
120corresponds to a set of instructions that are mandatory.  That is, each
121instruction has one or more conformance groups of which it is a member.
122
123This document defines the following conformance groups:
124
125* base32: includes all instructions defined in this
126  specification unless otherwise noted.
127* base64: includes base32, plus instructions explicitly noted
128  as being in the base64 conformance group.
129* atomic32: includes 32-bit atomic operation instructions (see `Atomic operations`_).
130* atomic64: includes atomic32, plus 64-bit atomic operation instructions.
131* divmul32: includes 32-bit division, multiplication, and modulo instructions.
132* divmul64: includes divmul32, plus 64-bit division, multiplication,
133  and modulo instructions.
134* packet: deprecated packet access instructions.
135
136Instruction encoding
137====================
138
139BPF has two instruction encodings:
140
141* the basic instruction encoding, which uses 64 bits to encode an instruction
142* the wide instruction encoding, which appends a second 64 bits
143  after the basic instruction for a total of 128 bits.
144
145Basic instruction encoding
146--------------------------
147
148A basic instruction is encoded as follows::
149
150  +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
151  |    opcode     |     regs      |            offset             |
152  +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
153  |                              imm                              |
154  +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
155
156**opcode**
157  operation to perform, encoded as follows::
158
159    +-+-+-+-+-+-+-+-+
160    |specific |class|
161    +-+-+-+-+-+-+-+-+
162
163  **specific**
164    The format of these bits varies by instruction class
165
166  **class**
167    The instruction class (see `Instruction classes`_)
168
169**regs**
170  The source and destination register numbers, encoded as follows
171  on a little-endian host::
172
173    +-+-+-+-+-+-+-+-+
174    |src_reg|dst_reg|
175    +-+-+-+-+-+-+-+-+
176
177  and as follows on a big-endian host::
178
179    +-+-+-+-+-+-+-+-+
180    |dst_reg|src_reg|
181    +-+-+-+-+-+-+-+-+
182
183  **src_reg**
184    the source register number (0-10), except where otherwise specified
185    (`64-bit immediate instructions`_ reuse this field for other purposes)
186
187  **dst_reg**
188    destination register number (0-10), unless otherwise specified
189    (future instructions might reuse this field for other purposes)
190
191**offset**
192  signed integer offset used with pointer arithmetic, except where
193  otherwise specified (some arithmetic instructions reuse this field
194  for other purposes)
195
196**imm**
197  signed integer immediate value
198
199Note that the contents of multi-byte fields ('offset' and 'imm') are
200stored using big-endian byte ordering on big-endian hosts and
201little-endian byte ordering on little-endian hosts.
202
203For example::
204
205  opcode                  offset imm          assembly
206         src_reg dst_reg
207  07     0       1        00 00  44 33 22 11  r1 += 0x11223344 // little
208         dst_reg src_reg
209  07     1       0        00 00  11 22 33 44  r1 += 0x11223344 // big
210
211Note that most instructions do not use all of the fields.
212Unused fields shall be cleared to zero.
213
214Wide instruction encoding
215--------------------------
216
217Some instructions are defined to use the wide instruction encoding,
218which uses two 32-bit immediate values.  The 64 bits following
219the basic instruction format contain a pseudo instruction
220with 'opcode', 'dst_reg', 'src_reg', and 'offset' all set to zero.
221
222This is depicted in the following figure::
223
224  +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
225  |    opcode     |     regs      |            offset             |
226  +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
227  |                              imm                              |
228  +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
229  |                           reserved                            |
230  +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
231  |                           next_imm                            |
232  +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
233
234**opcode**
235  operation to perform, encoded as explained above
236
237**regs**
238  The source and destination register numbers (unless otherwise
239  specified), encoded as explained above
240
241**offset**
242  signed integer offset used with pointer arithmetic, unless
243  otherwise specified
244
245**imm**
246  signed integer immediate value
247
248**reserved**
249  unused, set to zero
250
251**next_imm**
252  second signed integer immediate value
253
254Instruction classes
255-------------------
256
257The three least significant bits of the 'opcode' field store the instruction class:
258
259=====  =====  ===============================  ===================================
260class  value  description                      reference
261=====  =====  ===============================  ===================================
262LD     0x0    non-standard load operations     `Load and store instructions`_
263LDX    0x1    load into register operations    `Load and store instructions`_
264ST     0x2    store from immediate operations  `Load and store instructions`_
265STX    0x3    store from register operations   `Load and store instructions`_
266ALU    0x4    32-bit arithmetic operations     `Arithmetic and jump instructions`_
267JMP    0x5    64-bit jump operations           `Arithmetic and jump instructions`_
268JMP32  0x6    32-bit jump operations           `Arithmetic and jump instructions`_
269ALU64  0x7    64-bit arithmetic operations     `Arithmetic and jump instructions`_
270=====  =====  ===============================  ===================================
271
272Arithmetic and jump instructions
273================================
274
275For arithmetic and jump instructions (``ALU``, ``ALU64``, ``JMP`` and
276``JMP32``), the 8-bit 'opcode' field is divided into three parts::
277
278  +-+-+-+-+-+-+-+-+
279  |  code |s|class|
280  +-+-+-+-+-+-+-+-+
281
282**code**
283  the operation code, whose meaning varies by instruction class
284
285**s (source)**
286  the source operand location, which unless otherwise specified is one of:
287
288  ======  =====  ==============================================
289  source  value  description
290  ======  =====  ==============================================
291  K       0      use 32-bit 'imm' value as source operand
292  X       1      use 'src_reg' register value as source operand
293  ======  =====  ==============================================
294
295**instruction class**
296  the instruction class (see `Instruction classes`_)
297
298Arithmetic instructions
299-----------------------
300
301``ALU`` uses 32-bit wide operands while ``ALU64`` uses 64-bit wide operands for
302otherwise identical operations. ``ALU64`` instructions belong to the
303base64 conformance group unless noted otherwise.
304The 'code' field encodes the operation as below, where 'src' and 'dst' refer
305to the values of the source and destination registers, respectively.
306
307=====  =====  =======  ==========================================================
308name   code   offset   description
309=====  =====  =======  ==========================================================
310ADD    0x0    0        dst += src
311SUB    0x1    0        dst -= src
312MUL    0x2    0        dst \*= src
313DIV    0x3    0        dst = (src != 0) ? (dst / src) : 0
314SDIV   0x3    1        dst = (src != 0) ? (dst s/ src) : 0
315OR     0x4    0        dst \|= src
316AND    0x5    0        dst &= src
317LSH    0x6    0        dst <<= (src & mask)
318RSH    0x7    0        dst >>= (src & mask)
319NEG    0x8    0        dst = -dst
320MOD    0x9    0        dst = (src != 0) ? (dst % src) : dst
321SMOD   0x9    1        dst = (src != 0) ? (dst s% src) : dst
322XOR    0xa    0        dst ^= src
323MOV    0xb    0        dst = src
324MOVSX  0xb    8/16/32  dst = (s8,s16,s32)src
325ARSH   0xc    0        :term:`sign extending<Sign Extend>` dst >>= (src & mask)
326END    0xd    0        byte swap operations (see `Byte swap instructions`_ below)
327=====  =====  =======  ==========================================================
328
329Underflow and overflow are allowed during arithmetic operations, meaning
330the 64-bit or 32-bit value will wrap. If BPF program execution would
331result in division by zero, the destination register is instead set to zero.
332If execution would result in modulo by zero, for ``ALU64`` the value of
333the destination register is unchanged whereas for ``ALU`` the upper
33432 bits of the destination register are zeroed.
335
336``{ADD, X, ALU}``, where 'code' = ``ADD``, 'source' = ``X``, and 'class' = ``ALU``, means::
337
338  dst = (u32) ((u32) dst + (u32) src)
339
340where '(u32)' indicates that the upper 32 bits are zeroed.
341
342``{ADD, X, ALU64}`` means::
343
344  dst = dst + src
345
346``{XOR, K, ALU}`` means::
347
348  dst = (u32) dst ^ (u32) imm
349
350``{XOR, K, ALU64}`` means::
351
352  dst = dst ^ imm
353
354Note that most arithmetic instructions have 'offset' set to 0. Only three instructions
355(``SDIV``, ``SMOD``, ``MOVSX``) have a non-zero 'offset'.
356
357Division, multiplication, and modulo operations for ``ALU`` are part
358of the "divmul32" conformance group, and division, multiplication, and
359modulo operations for ``ALU64`` are part of the "divmul64" conformance
360group.
361The division and modulo operations support both unsigned and signed flavors.
362
363For unsigned operations (``DIV`` and ``MOD``), for ``ALU``,
364'imm' is interpreted as a 32-bit unsigned value. For ``ALU64``,
365'imm' is first :term:`sign extended<Sign Extend>` from 32 to 64 bits, and then
366interpreted as a 64-bit unsigned value.
367
368For signed operations (``SDIV`` and ``SMOD``), for ``ALU``,
369'imm' is interpreted as a 32-bit signed value. For ``ALU64``, 'imm'
370is first :term:`sign extended<Sign Extend>` from 32 to 64 bits, and then
371interpreted as a 64-bit signed value.
372
373Note that there are varying definitions of the signed modulo operation
374when the dividend or divisor are negative, where implementations often
375vary by language such that Python, Ruby, etc.  differ from C, Go, Java,
376etc. This specification requires that signed modulo use truncated division
377(where -13 % 3 == -1) as implemented in C, Go, etc.::
378
379   a % n = a - n * trunc(a / n)
380
381The ``MOVSX`` instruction does a move operation with sign extension.
382``{MOVSX, X, ALU}`` :term:`sign extends<Sign Extend>` 8-bit and 16-bit operands into
38332-bit operands, and zeroes the remaining upper 32 bits.
384``{MOVSX, X, ALU64}`` :term:`sign extends<Sign Extend>` 8-bit, 16-bit, and 32-bit
385operands into 64-bit operands.  Unlike other arithmetic instructions,
386``MOVSX`` is only defined for register source operands (``X``).
387
388The ``NEG`` instruction is only defined when the source bit is clear
389(``K``).
390
391Shift operations use a mask of 0x3F (63) for 64-bit operations and 0x1F (31)
392for 32-bit operations.
393
394Byte swap instructions
395----------------------
396
397The byte swap instructions use instruction classes of ``ALU`` and ``ALU64``
398and a 4-bit 'code' field of ``END``.
399
400The byte swap instructions operate on the destination register
401only and do not use a separate source register or immediate value.
402
403For ``ALU``, the 1-bit source operand field in the opcode is used to
404select what byte order the operation converts from or to. For
405``ALU64``, the 1-bit source operand field in the opcode is reserved
406and must be set to 0.
407
408=====  ========  =====  =================================================
409class  source    value  description
410=====  ========  =====  =================================================
411ALU    TO_LE     0      convert between host byte order and little endian
412ALU    TO_BE     1      convert between host byte order and big endian
413ALU64  Reserved  0      do byte swap unconditionally
414=====  ========  =====  =================================================
415
416The 'imm' field encodes the width of the swap operations.  The following widths
417are supported: 16, 32 and 64.  Width 64 operations belong to the base64
418conformance group and other swap operations belong to the base32
419conformance group.
420
421Examples:
422
423``{END, TO_LE, ALU}`` with 'imm' = 16/32/64 means::
424
425  dst = htole16(dst)
426  dst = htole32(dst)
427  dst = htole64(dst)
428
429``{END, TO_BE, ALU}`` with 'imm' = 16/32/64 means::
430
431  dst = htobe16(dst)
432  dst = htobe32(dst)
433  dst = htobe64(dst)
434
435``{END, TO_LE, ALU64}`` with 'imm' = 16/32/64 means::
436
437  dst = bswap16(dst)
438  dst = bswap32(dst)
439  dst = bswap64(dst)
440
441Jump instructions
442-----------------
443
444``JMP32`` uses 32-bit wide operands and indicates the base32
445conformance group, while ``JMP`` uses 64-bit wide operands for
446otherwise identical operations, and indicates the base64 conformance
447group unless otherwise specified.
448The 'code' field encodes the operation as below:
449
450========  =====  =======  =================================  ===================================================
451code      value  src_reg  description                        notes
452========  =====  =======  =================================  ===================================================
453JA        0x0    0x0      PC += offset                       {JA, K, JMP} only
454JA        0x0    0x0      PC += imm                          {JA, K, JMP32} only
455JEQ       0x1    any      PC += offset if dst == src
456JGT       0x2    any      PC += offset if dst > src          unsigned
457JGE       0x3    any      PC += offset if dst >= src         unsigned
458JSET      0x4    any      PC += offset if dst & src
459JNE       0x5    any      PC += offset if dst != src
460JSGT      0x6    any      PC += offset if dst > src          signed
461JSGE      0x7    any      PC += offset if dst >= src         signed
462CALL      0x8    0x0      call helper function by static ID  {CALL, K, JMP} only, see `Helper functions`_
463CALL      0x8    0x1      call PC += imm                     {CALL, K, JMP} only, see `Program-local functions`_
464CALL      0x8    0x2      call helper function by BTF ID     {CALL, K, JMP} only, see `Helper functions`_
465EXIT      0x9    0x0      return                             {CALL, K, JMP} only
466JLT       0xa    any      PC += offset if dst < src          unsigned
467JLE       0xb    any      PC += offset if dst <= src         unsigned
468JSLT      0xc    any      PC += offset if dst < src          signed
469JSLE      0xd    any      PC += offset if dst <= src         signed
470========  =====  =======  =================================  ===================================================
471
472where 'PC' denotes the program counter, and the offset to increment by
473is in units of 64-bit instructions relative to the instruction following
474the jump instruction.  Thus 'PC += 1' skips execution of the next
475instruction if it's a basic instruction or results in undefined behavior
476if the next instruction is a 128-bit wide instruction.
477
478The BPF program needs to store the return value into register R0 before doing an
479``EXIT``.
480
481Example:
482
483``{JSGE, X, JMP32}`` means::
484
485  if (s32)dst s>= (s32)src goto +offset
486
487where 's>=' indicates a signed '>=' comparison.
488
489``{JA, K, JMP32}`` means::
490
491  gotol +imm
492
493where 'imm' means the branch offset comes from the 'imm' field.
494
495Note that there are two flavors of ``JA`` instructions. The
496``JMP`` class permits a 16-bit jump offset specified by the 'offset'
497field, whereas the ``JMP32`` class permits a 32-bit jump offset
498specified by the 'imm' field. A > 16-bit conditional jump may be
499converted to a < 16-bit conditional jump plus a 32-bit unconditional
500jump.
501
502All ``CALL`` and ``JA`` instructions belong to the
503base32 conformance group.
504
505Helper functions
506~~~~~~~~~~~~~~~~
507
508Helper functions are a concept whereby BPF programs can call into a
509set of function calls exposed by the underlying platform.
510
511Historically, each helper function was identified by a static ID
512encoded in the 'imm' field.  The available helper functions may differ
513for each program type, but static IDs are unique across all program types.
514
515Platforms that support the BPF Type Format (BTF) support identifying
516a helper function by a BTF ID encoded in the 'imm' field, where the BTF ID
517identifies the helper name and type.
518
519Program-local functions
520~~~~~~~~~~~~~~~~~~~~~~~
521Program-local functions are functions exposed by the same BPF program as the
522caller, and are referenced by offset from the call instruction, similar to
523``JA``.  The offset is encoded in the 'imm' field of the call instruction.
524An ``EXIT`` within the program-local function will return to the caller.
525
526Load and store instructions
527===========================
528
529For load and store instructions (``LD``, ``LDX``, ``ST``, and ``STX``), the
5308-bit 'opcode' field is divided as follows::
531
532  +-+-+-+-+-+-+-+-+
533  |mode |sz |class|
534  +-+-+-+-+-+-+-+-+
535
536**mode**
537  The mode modifier is one of:
538
539    =============  =====  ====================================  =============
540    mode modifier  value  description                           reference
541    =============  =====  ====================================  =============
542    IMM            0      64-bit immediate instructions         `64-bit immediate instructions`_
543    ABS            1      legacy BPF packet access (absolute)   `Legacy BPF Packet access instructions`_
544    IND            2      legacy BPF packet access (indirect)   `Legacy BPF Packet access instructions`_
545    MEM            3      regular load and store operations     `Regular load and store operations`_
546    MEMSX          4      sign-extension load operations        `Sign-extension load operations`_
547    ATOMIC         6      atomic operations                     `Atomic operations`_
548    =============  =====  ====================================  =============
549
550**sz (size)**
551  The size modifier is one of:
552
553    ====  =====  =====================
554    size  value  description
555    ====  =====  =====================
556    W     0      word        (4 bytes)
557    H     1      half word   (2 bytes)
558    B     2      byte
559    DW    3      double word (8 bytes)
560    ====  =====  =====================
561
562  Instructions using ``DW`` belong to the base64 conformance group.
563
564**class**
565  The instruction class (see `Instruction classes`_)
566
567Regular load and store operations
568---------------------------------
569
570The ``MEM`` mode modifier is used to encode regular load and store
571instructions that transfer data between a register and memory.
572
573``{MEM, <size>, STX}`` means::
574
575  *(size *) (dst + offset) = src
576
577``{MEM, <size>, ST}`` means::
578
579  *(size *) (dst + offset) = imm
580
581``{MEM, <size>, LDX}`` means::
582
583  dst = *(unsigned size *) (src + offset)
584
585Where '<size>' is one of: ``B``, ``H``, ``W``, or ``DW``, and
586'unsigned size' is one of: u8, u16, u32, or u64.
587
588Sign-extension load operations
589------------------------------
590
591The ``MEMSX`` mode modifier is used to encode :term:`sign-extension<Sign Extend>` load
592instructions that transfer data between a register and memory.
593
594``{MEMSX, <size>, LDX}`` means::
595
596  dst = *(signed size *) (src + offset)
597
598Where '<size>' is one of: ``B``, ``H``, or ``W``, and
599'signed size' is one of: s8, s16, or s32.
600
601Atomic operations
602-----------------
603
604Atomic operations are operations that operate on memory and can not be
605interrupted or corrupted by other access to the same memory region
606by other BPF programs or means outside of this specification.
607
608All atomic operations supported by BPF are encoded as store operations
609that use the ``ATOMIC`` mode modifier as follows:
610
611* ``{ATOMIC, W, STX}`` for 32-bit operations, which are
612  part of the "atomic32" conformance group.
613* ``{ATOMIC, DW, STX}`` for 64-bit operations, which are
614  part of the "atomic64" conformance group.
615* 8-bit and 16-bit wide atomic operations are not supported.
616
617The 'imm' field is used to encode the actual atomic operation.
618Simple atomic operation use a subset of the values defined to encode
619arithmetic operations in the 'imm' field to encode the atomic operation:
620
621========  =====  ===========
622imm       value  description
623========  =====  ===========
624ADD       0x00   atomic add
625OR        0x40   atomic or
626AND       0x50   atomic and
627XOR       0xa0   atomic xor
628========  =====  ===========
629
630
631``{ATOMIC, W, STX}`` with 'imm' = ADD means::
632
633  *(u32 *)(dst + offset) += src
634
635``{ATOMIC, DW, STX}`` with 'imm' = ADD means::
636
637  *(u64 *)(dst + offset) += src
638
639In addition to the simple atomic operations, there also is a modifier and
640two complex atomic operations:
641
642===========  ================  ===========================
643imm          value             description
644===========  ================  ===========================
645FETCH        0x01              modifier: return old value
646XCHG         0xe0 | FETCH      atomic exchange
647CMPXCHG      0xf0 | FETCH      atomic compare and exchange
648===========  ================  ===========================
649
650The ``FETCH`` modifier is optional for simple atomic operations, and
651always set for the complex atomic operations.  If the ``FETCH`` flag
652is set, then the operation also overwrites ``src`` with the value that
653was in memory before it was modified.
654
655The ``XCHG`` operation atomically exchanges ``src`` with the value
656addressed by ``dst + offset``.
657
658The ``CMPXCHG`` operation atomically compares the value addressed by
659``dst + offset`` with ``R0``. If they match, the value addressed by
660``dst + offset`` is replaced with ``src``. In either case, the
661value that was at ``dst + offset`` before the operation is zero-extended
662and loaded back to ``R0``.
663
66464-bit immediate instructions
665-----------------------------
666
667Instructions with the ``IMM`` 'mode' modifier use the wide instruction
668encoding defined in `Instruction encoding`_, and use the 'src_reg' field of the
669basic instruction to hold an opcode subtype.
670
671The following table defines a set of ``{IMM, DW, LD}`` instructions
672with opcode subtypes in the 'src_reg' field, using new terms such as "map"
673defined further below:
674
675=======  =========================================  ===========  ==============
676src_reg  pseudocode                                 imm type     dst type
677=======  =========================================  ===========  ==============
6780x0      dst = (next_imm << 32) | imm               integer      integer
6790x1      dst = map_by_fd(imm)                       map fd       map
6800x2      dst = map_val(map_by_fd(imm)) + next_imm   map fd       data address
6810x3      dst = var_addr(imm)                        variable id  data address
6820x4      dst = code_addr(imm)                       integer      code address
6830x5      dst = map_by_idx(imm)                      map index    map
6840x6      dst = map_val(map_by_idx(imm)) + next_imm  map index    data address
685=======  =========================================  ===========  ==============
686
687where
688
689* map_by_fd(imm) means to convert a 32-bit file descriptor into an address of a map (see `Maps`_)
690* map_by_idx(imm) means to convert a 32-bit index into an address of a map
691* map_val(map) gets the address of the first value in a given map
692* var_addr(imm) gets the address of a platform variable (see `Platform Variables`_) with a given id
693* code_addr(imm) gets the address of the instruction at a specified relative offset in number of (64-bit) instructions
694* the 'imm type' can be used by disassemblers for display
695* the 'dst type' can be used for verification and JIT compilation purposes
696
697Maps
698~~~~
699
700Maps are shared memory regions accessible by BPF programs on some platforms.
701A map can have various semantics as defined in a separate document, and may or
702may not have a single contiguous memory region, but the 'map_val(map)' is
703currently only defined for maps that do have a single contiguous memory region.
704
705Each map can have a file descriptor (fd) if supported by the platform, where
706'map_by_fd(imm)' means to get the map with the specified file descriptor. Each
707BPF program can also be defined to use a set of maps associated with the
708program at load time, and 'map_by_idx(imm)' means to get the map with the given
709index in the set associated with the BPF program containing the instruction.
710
711Platform Variables
712~~~~~~~~~~~~~~~~~~
713
714Platform variables are memory regions, identified by integer ids, exposed by
715the runtime and accessible by BPF programs on some platforms.  The
716'var_addr(imm)' operation means to get the address of the memory region
717identified by the given id.
718
719Legacy BPF Packet access instructions
720-------------------------------------
721
722BPF previously introduced special instructions for access to packet data that were
723carried over from classic BPF. These instructions used an instruction
724class of ``LD``, a size modifier of ``W``, ``H``, or ``B``, and a
725mode modifier of ``ABS`` or ``IND``.  The 'dst_reg' and 'offset' fields were
726set to zero, and 'src_reg' was set to zero for ``ABS``.  However, these
727instructions are deprecated and should no longer be used.  All legacy packet
728access instructions belong to the "packet" conformance group.
729