kernel / Documentation / networking / filter.txt

Linux Socket Filtering aka Berkeley Packet Filter (BPF)
=======================================================

Introduction
------------

Linux Socket Filtering (LSF) is derived from the Berkeley Packet Filter.
Though there are some distinct differences between the BSD and Linux
Kernel filtering, but when we speak of BPF or LSF in Linux context, we
mean the very same mechanism of filtering in the Linux kernel.

BPF allows a user-space program to attach a filter onto any socket and
allow or disallow certain types of data to come through the socket. LSF
follows exactly the same filter code structure as BSD's BPF, so referring
to the BSD bpf.4 manpage is very helpful in creating filters.

On Linux, BPF is much simpler than on BSD. One does not have to worry
about devices or anything like that. You simply create your filter code,
send it to the kernel via the SO_ATTACH_FILTER option and if your filter
code passes the kernel check on it, you then immediately begin filtering
data on that socket.

You can also detach filters from your socket via the SO_DETACH_FILTER
option. This will probably not be used much since when you close a socket
that has a filter on it the filter is automagically removed. The other
less common case may be adding a different filter on the same socket where
you had another filter that is still running: the kernel takes care of
removing the old one and placing your new one in its place, assuming your
filter has passed the checks, otherwise if it fails the old filter will
remain on that socket.

SO_LOCK_FILTER option allows to lock the filter attached to a socket. Once
set, a filter cannot be removed or changed. This allows one process to
setup a socket, attach a filter, lock it then drop privileges and be
assured that the filter will be kept until the socket is closed.

The biggest user of this construct might be libpcap. Issuing a high-level
filter command like `tcpdump -i em1 port 22` passes through the libpcap
internal compiler that generates a structure that can eventually be loaded
via SO_ATTACH_FILTER to the kernel. `tcpdump -i em1 port 22 -ddd`
displays what is being placed into this structure.

Although we were only speaking about sockets here, BPF in Linux is used
in many more places. There's xt_bpf for netfilter, cls_bpf in the kernel
qdisc layer, SECCOMP-BPF (SECure COMPuting [1]), and lots of other places
such as team driver, PTP code, etc where BPF is being used.

 [1] Documentation/prctl/seccomp_filter.txt

Original BPF paper:

Steven McCanne and Van Jacobson. 1993. The BSD packet filter: a new
architecture for user-level packet capture. In Proceedings of the
USENIX Winter 1993 Conference Proceedings on USENIX Winter 1993
Conference Proceedings (USENIX'93). USENIX Association, Berkeley,
CA, USA, 2-2. [http://www.tcpdump.org/papers/bpf-usenix93.pdf]

Structure
---------

User space applications include <linux/filter.h> which contains the
following relevant structures:

struct sock_filter {	/* Filter block */
	__u16	code;   /* Actual filter code */
	__u8	jt;	/* Jump true */
	__u8	jf;	/* Jump false */
	__u32	k;      /* Generic multiuse field */
};

Such a structure is assembled as an array of 4-tuples, that contains
a code, jt, jf and k value. jt and jf are jump offsets and k a generic
value to be used for a provided code.

struct sock_fprog {			/* Required for SO_ATTACH_FILTER. */
	unsigned short		   len;	/* Number of filter blocks */
	struct sock_filter __user *filter;
};

For socket filtering, a pointer to this structure (as shown in
follow-up example) is being passed to the kernel through setsockopt(2).

Example
-------

#include <sys/socket.h>
#include <sys/types.h>
#include <arpa/inet.h>
#include <linux/if_ether.h>
/* ... */

/* From the example above: tcpdump -i em1 port 22 -dd */
struct sock_filter code[] = {
	{ 0x28,  0,  0, 0x0000000c },
	{ 0x15,  0,  8, 0x000086dd },
	{ 0x30,  0,  0, 0x00000014 },
	{ 0x15,  2,  0, 0x00000084 },
	{ 0x15,  1,  0, 0x00000006 },
	{ 0x15,  0, 17, 0x00000011 },
	{ 0x28,  0,  0, 0x00000036 },
	{ 0x15, 14,  0, 0x00000016 },
	{ 0x28,  0,  0, 0x00000038 },
	{ 0x15, 12, 13, 0x00000016 },
	{ 0x15,  0, 12, 0x00000800 },
	{ 0x30,  0,  0, 0x00000017 },
	{ 0x15,  2,  0, 0x00000084 },
	{ 0x15,  1,  0, 0x00000006 },
	{ 0x15,  0,  8, 0x00000011 },
	{ 0x28,  0,  0, 0x00000014 },
	{ 0x45,  6,  0, 0x00001fff },
	{ 0xb1,  0,  0, 0x0000000e },
	{ 0x48,  0,  0, 0x0000000e },
	{ 0x15,  2,  0, 0x00000016 },
	{ 0x48,  0,  0, 0x00000010 },
	{ 0x15,  0,  1, 0x00000016 },
	{ 0x06,  0,  0, 0x0000ffff },
	{ 0x06,  0,  0, 0x00000000 },
};

struct sock_fprog bpf = {
	.len = ARRAY_SIZE(code),
	.filter = code,
};

sock = socket(PF_PACKET, SOCK_RAW, htons(ETH_P_ALL));
if (sock < 0)
	/* ... bail out ... */

ret = setsockopt(sock, SOL_SOCKET, SO_ATTACH_FILTER, &bpf, sizeof(bpf));
if (ret < 0)
	/* ... bail out ... */

/* ... */
close(sock);

The above example code attaches a socket filter for a PF_PACKET socket
in order to let all IPv4/IPv6 packets with port 22 pass. The rest will
be dropped for this socket.

The setsockopt(2) call to SO_DETACH_FILTER doesn't need any arguments
and SO_LOCK_FILTER for preventing the filter to be detached, takes an
integer value with 0 or 1.

Note that socket filters are not restricted to PF_PACKET sockets only,
but can also be used on other socket families.

Summary of system calls:

 * setsockopt(sockfd, SOL_SOCKET, SO_ATTACH_FILTER, &val, sizeof(val));
 * setsockopt(sockfd, SOL_SOCKET, SO_DETACH_FILTER, &val, sizeof(val));
 * setsockopt(sockfd, SOL_SOCKET, SO_LOCK_FILTER,   &val, sizeof(val));

Normally, most use cases for socket filtering on packet sockets will be
covered by libpcap in high-level syntax, so as an application developer
you should stick to that. libpcap wraps its own layer around all that.

Unless i) using/linking to libpcap is not an option, ii) the required BPF
filters use Linux extensions that are not supported by libpcap's compiler,
iii) a filter might be more complex and not cleanly implementable with
libpcap's compiler, or iv) particular filter codes should be optimized
differently than libpcap's internal compiler does; then in such cases
writing such a filter "by hand" can be of an alternative. For example,
xt_bpf and cls_bpf users might have requirements that could result in
more complex filter code, or one that cannot be expressed with libpcap
(e.g. different return codes for various code paths). Moreover, BPF JIT
implementors may wish to manually write test cases and thus need low-level
access to BPF code as well.

BPF engine and instruction set
------------------------------

Under tools/net/ there's a small helper tool called bpf_asm which can
be used to write low-level filters for example scenarios mentioned in the
previous section. Asm-like syntax mentioned here has been implemented in
bpf_asm and will be used for further explanations (instead of dealing with
less readable opcodes directly, principles are the same). The syntax is
closely modelled after Steven McCanne's and Van Jacobson's BPF paper.

The BPF architecture consists of the following basic elements:

  Element          Description

  A                32 bit wide accumulator
  X                32 bit wide X register
  M[]              16 x 32 bit wide misc registers aka "scratch memory
                   store", addressable from 0 to 15

A program, that is translated by bpf_asm into "opcodes" is an array that
consists of the following elements (as already mentioned):

  op:16, jt:8, jf:8, k:32

The element op is a 16 bit wide opcode that has a particular instruction
encoded. jt and jf are two 8 bit wide jump targets, one for condition
"jump if true", the other one "jump if false". Eventually, element k
contains a miscellaneous argument that can be interpreted in different
ways depending on the given instruction in op.

The instruction set consists of load, store, branch, alu, miscellaneous
and return instructions that are also represented in bpf_asm syntax. This
table lists all bpf_asm instructions available resp. what their underlying
opcodes as defined in linux/filter.h stand for:

  Instruction      Addressing mode      Description

  ld               1, 2, 3, 4, 10       Load word into A
  ldi              4                    Load word into A
  ldh              1, 2                 Load half-word into A
  ldb              1, 2                 Load byte into A
  ldx              3, 4, 5, 10          Load word into X
  ldxi             4                    Load word into X
  ldxb             5                    Load byte into X

  st               3                    Store A into M[]
  stx              3                    Store X into M[]

  jmp              6                    Jump to label
  ja               6                    Jump to label
  jeq              7, 8                 Jump on k == A
  jneq             8                    Jump on k != A
  jne              8                    Jump on k != A
  jlt              8                    Jump on k < A
  jle              8                    Jump on k <= A
  jgt              7, 8                 Jump on k > A
  jge              7, 8                 Jump on k >= A
  jset             7, 8                 Jump on k & A

  add              0, 4                 A + <x>
  sub              0, 4                 A - <x>
  mul              0, 4                 A * <x>
  div              0, 4                 A / <x>
  mod              0, 4                 A % <x>
  neg              0, 4                 !A
  and              0, 4                 A & <x>
  or               0, 4                 A | <x>
  xor              0, 4                 A ^ <x>
  lsh              0, 4                 A << <x>
  rsh              0, 4                 A >> <x>

  tax                                   Copy A into X
  txa                                   Copy X into A

  ret              4, 9                 Return

The next table shows addressing formats from the 2nd column:

  Addressing mode  Syntax               Description

   0               x/%x                 Register X
   1               [k]                  BHW at byte offset k in the packet
   2               [x + k]              BHW at the offset X + k in the packet
   3               M[k]                 Word at offset k in M[]
   4               #k                   Literal value stored in k
   5               4*([k]&0xf)          Lower nibble * 4 at byte offset k in the packet
   6               L                    Jump label L
   7               #k,Lt,Lf             Jump to Lt if true, otherwise jump to Lf
   8               #k,Lt                Jump to Lt if predicate is true
   9               a/%a                 Accumulator A
  10               extension            BPF extension

The Linux kernel also has a couple of BPF extensions that are used along
with the class of load instructions by "overloading" the k argument with
a negative offset + a particular extension offset. The result of such BPF
extensions are loaded into A.

Possible BPF extensions are shown in the following table:

  Extension                             Description

  len                                   skb->len
  proto                                 skb->protocol
  type                                  skb->pkt_type
  poff                                  Payload start offset
  ifidx                                 skb->dev->ifindex
  nla                                   Netlink attribute of type X with offset A
  nlan                                  Nested Netlink attribute of type X with offset A
  mark                                  skb->mark
  queue                                 skb->queue_mapping
  hatype                                skb->dev->type
  rxhash                                skb->hash
  cpu                                   raw_smp_processor_id()
  vlan_tci                              vlan_tx_tag_get(skb)
  vlan_pr                               vlan_tx_tag_present(skb)
  rand                                  prandom_u32()

These extensions can also be prefixed with '#'.
Examples for low-level BPF:

** ARP packets:

  ldh [12]
  jne #0x806, drop
  ret #-1
  drop: ret #0

** IPv4 TCP packets:

  ldh [12]
  jne #0x800, drop
  ldb [23]
  jneq #6, drop
  ret #-1
  drop: ret #0

** (Accelerated) VLAN w/ id 10:

  ld vlan_tci
  jneq #10, drop
  ret #-1
  drop: ret #0

** icmp random packet sampling, 1 in 4
  ldh [12]
  jne #0x800, drop
  ldb [23]
  jneq #1, drop
  # get a random uint32 number
  ld rand
  mod #4
  jneq #1, drop
  ret #-1
  drop: ret #0

** SECCOMP filter example:

  ld [4]                  /* offsetof(struct seccomp_data, arch) */
  jne #0xc000003e, bad    /* AUDIT_ARCH_X86_64 */
  ld [0]                  /* offsetof(struct seccomp_data, nr) */
  jeq #15, good           /* __NR_rt_sigreturn */
  jeq #231, good          /* __NR_exit_group */
  jeq #60, good           /* __NR_exit */
  jeq #0, good            /* __NR_read */
  jeq #1, good            /* __NR_write */
  jeq #5, good            /* __NR_fstat */
  jeq #9, good            /* __NR_mmap */
  jeq #14, good           /* __NR_rt_sigprocmask */
  jeq #13, good           /* __NR_rt_sigaction */
  jeq #35, good           /* __NR_nanosleep */
  bad: ret #0             /* SECCOMP_RET_KILL */
  good: ret #0x7fff0000   /* SECCOMP_RET_ALLOW */

The above example code can be placed into a file (here called "foo"), and
then be passed to the bpf_asm tool for generating opcodes, output that xt_bpf
and cls_bpf understands and can directly be loaded with. Example with above
ARP code:

$ ./bpf_asm foo
4,40 0 0 12,21 0 1 2054,6 0 0 4294967295,6 0 0 0,

In copy and paste C-like output:

$ ./bpf_asm -c foo
{ 0x28,  0,  0, 0x0000000c },
{ 0x15,  0,  1, 0x00000806 },
{ 0x06,  0,  0, 0xffffffff },
{ 0x06,  0,  0, 0000000000 },

In particular, as usage with xt_bpf or cls_bpf can result in more complex BPF
filters that might not be obvious at first, it's good to test filters before
attaching to a live system. For that purpose, there's a small tool called
bpf_dbg under tools/net/ in the kernel source directory. This debugger allows
for testing BPF filters against given pcap files, single stepping through the
BPF code on the pcap's packets and to do BPF machine register dumps.

Starting bpf_dbg is trivial and just requires issuing:

# ./bpf_dbg

In case input and output do not equal stdin/stdout, bpf_dbg takes an
alternative stdin source as a first argument, and an alternative stdout
sink as a second one, e.g. `./bpf_dbg test_in.txt test_out.txt`.

Other than that, a particular libreadline configuration can be set via
file "~/.bpf_dbg_init" and the command history is stored in the file
"~/.bpf_dbg_history".

Interaction in bpf_dbg happens through a shell that also has auto-completion
support (follow-up example commands starting with '>' denote bpf_dbg shell).
The usual workflow would be to ...

> load bpf 6,40 0 0 12,21 0 3 2048,48 0 0 23,21 0 1 1,6 0 0 65535,6 0 0 0
  Loads a BPF filter from standard output of bpf_asm, or transformed via
  e.g. `tcpdump -iem1 -ddd port 22 | tr '\n' ','`. Note that for JIT
  debugging (next section), this command creates a temporary socket and
  loads the BPF code into the kernel. Thus, this will also be useful for
  JIT developers.

> load pcap foo.pcap
  Loads standard tcpdump pcap file.

> run [<n>]
bpf passes:1 fails:9
  Runs through all packets from a pcap to account how many passes and fails
  the filter will generate. A limit of packets to traverse can be given.

> disassemble
l0:	ldh [12]
l1:	jeq #0x800, l2, l5
l2:	ldb [23]
l3:	jeq #0x1, l4, l5
l4:	ret #0xffff
l5:	ret #0
  Prints out BPF code disassembly.

> dump
/* { op, jt, jf, k }, */
{ 0x28,  0,  0, 0x0000000c },
{ 0x15,  0,  3, 0x00000800 },
{ 0x30,  0,  0, 0x00000017 },
{ 0x15,  0,  1, 0x00000001 },
{ 0x06,  0,  0, 0x0000ffff },
{ 0x06,  0,  0, 0000000000 },
  Prints out C-style BPF code dump.

> breakpoint 0
breakpoint at: l0:	ldh [12]
> breakpoint 1
breakpoint at: l1:	jeq #0x800, l2, l5
  ...
  Sets breakpoints at particular BPF instructions. Issuing a `run` command
  will walk through the pcap file continuing from the current packet and
  break when a breakpoint is being hit (another `run` will continue from
  the currently active breakpoint executing next instructions):

  > run
  -- register dump --
  pc:       [0]                       <-- program counter
  code:     [40] jt[0] jf[0] k[12]    <-- plain BPF code of current instruction
  curr:     l0:	ldh [12]              <-- disassembly of current instruction
  A:        [00000000][0]             <-- content of A (hex, decimal)
  X:        [00000000][0]             <-- content of X (hex, decimal)
  M[0,15]:  [00000000][0]             <-- folded content of M (hex, decimal)
  -- packet dump --                   <-- Current packet from pcap (hex)
  len: 42
    0: 00 19 cb 55 55 a4 00 14 a4 43 78 69 08 06 00 01
   16: 08 00 06 04 00 01 00 14 a4 43 78 69 0a 3b 01 26
   32: 00 00 00 00 00 00 0a 3b 01 01
  (breakpoint)
  >

> breakpoint
breakpoints: 0 1
  Prints currently set breakpoints.

> step [-<n>, +<n>]
  Performs single stepping through the BPF program from the current pc
  offset. Thus, on each step invocation, above register dump is issued.
  This can go forwards and backwards in time, a plain `step` will break
  on the next BPF instruction, thus +1. (No `run` needs to be issued here.)

> select <n>
  Selects a given packet from the pcap file to continue from. Thus, on
  the next `run` or `step`, the BPF program is being evaluated against
  the user pre-selected packet. Numbering starts just as in Wireshark
  with index 1.

> quit
#
  Exits bpf_dbg.

JIT compiler
------------

The Linux kernel has a built-in BPF JIT compiler for x86_64, SPARC, PowerPC,
ARM and s390 and can be enabled through CONFIG_BPF_JIT. The JIT compiler is
transparently invoked for each attached filter from user space or for internal
kernel users if it has been previously enabled by root:

  echo 1 > /proc/sys/net/core/bpf_jit_enable

For JIT developers, doing audits etc, each compile run can output the generated
opcode image into the kernel log via:

  echo 2 > /proc/sys/net/core/bpf_jit_enable

Example output from dmesg:

[ 3389.935842] flen=6 proglen=70 pass=3 image=ffffffffa0069c8f
[ 3389.935847] JIT code: 00000000: 55 48 89 e5 48 83 ec 60 48 89 5d f8 44 8b 4f 68
[ 3389.935849] JIT code: 00000010: 44 2b 4f 6c 4c 8b 87 d8 00 00 00 be 0c 00 00 00
[ 3389.935850] JIT code: 00000020: e8 1d 94 ff e0 3d 00 08 00 00 75 16 be 17 00 00
[ 3389.935851] JIT code: 00000030: 00 e8 28 94 ff e0 83 f8 01 75 07 b8 ff ff 00 00
[ 3389.935852] JIT code: 00000040: eb 02 31 c0 c9 c3

In the kernel source tree under tools/net/, there's bpf_jit_disasm for
generating disassembly out of the kernel log's hexdump:

# ./bpf_jit_disasm
70 bytes emitted from JIT compiler (pass:3, flen:6)
ffffffffa0069c8f + <x>:
   0:	push   %rbp
   1:	mov    %rsp,%rbp
   4:	sub    $0x60,%rsp
   8:	mov    %rbx,-0x8(%rbp)
   c:	mov    0x68(%rdi),%r9d
  10:	sub    0x6c(%rdi),%r9d
  14:	mov    0xd8(%rdi),%r8
  1b:	mov    $0xc,%esi
  20:	callq  0xffffffffe0ff9442
  25:	cmp    $0x800,%eax
  2a:	jne    0x0000000000000042
  2c:	mov    $0x17,%esi
  31:	callq  0xffffffffe0ff945e
  36:	cmp    $0x1,%eax
  39:	jne    0x0000000000000042
  3b:	mov    $0xffff,%eax
  40:	jmp    0x0000000000000044
  42:	xor    %eax,%eax
  44:	leaveq
  45:	retq

Issuing option `-o` will "annotate" opcodes to resulting assembler
instructions, which can be very useful for JIT developers:

# ./bpf_jit_disasm -o
70 bytes emitted from JIT compiler (pass:3, flen:6)
ffffffffa0069c8f + <x>:
   0:	push   %rbp
	55
   1:	mov    %rsp,%rbp
	48 89 e5
   4:	sub    $0x60,%rsp
	48 83 ec 60
   8:	mov    %rbx,-0x8(%rbp)
	48 89 5d f8
   c:	mov    0x68(%rdi),%r9d
	44 8b 4f 68
  10:	sub    0x6c(%rdi),%r9d
	44 2b 4f 6c
  14:	mov    0xd8(%rdi),%r8
	4c 8b 87 d8 00 00 00
  1b:	mov    $0xc,%esi
	be 0c 00 00 00
  20:	callq  0xffffffffe0ff9442
	e8 1d 94 ff e0
  25:	cmp    $0x800,%eax
	3d 00 08 00 00
  2a:	jne    0x0000000000000042
	75 16
  2c:	mov    $0x17,%esi
	be 17 00 00 00
  31:	callq  0xffffffffe0ff945e
	e8 28 94 ff e0
  36:	cmp    $0x1,%eax
	83 f8 01
  39:	jne    0x0000000000000042
	75 07
  3b:	mov    $0xffff,%eax
	b8 ff ff 00 00
  40:	jmp    0x0000000000000044
	eb 02
  42:	xor    %eax,%eax
	31 c0
  44:	leaveq
	c9
  45:	retq
	c3

For BPF JIT developers, bpf_jit_disasm, bpf_asm and bpf_dbg provides a useful
toolchain for developing and testing the kernel's JIT compiler.

BPF kernel internals
--------------------
Internally, for the kernel interpreter, a different instruction set
format with similar underlying principles from BPF described in previous
paragraphs is being used. However, the instruction set format is modelled
closer to the underlying architecture to mimic native instruction sets, so
that a better performance can be achieved (more details later). This new
ISA is called 'eBPF' or 'internal BPF' interchangeably. (Note: eBPF which
originates from [e]xtended BPF is not the same as BPF extensions! While
eBPF is an ISA, BPF extensions date back to classic BPF's 'overloading'
of BPF_LD | BPF_{B,H,W} | BPF_ABS instruction.)

It is designed to be JITed with one to one mapping, which can also open up
the possibility for GCC/LLVM compilers to generate optimized eBPF code through
an eBPF backend that performs almost as fast as natively compiled code.

The new instruction set was originally designed with the possible goal in
mind to write programs in "restricted C" and compile into eBPF with a optional
GCC/LLVM backend, so that it can just-in-time map to modern 64-bit CPUs with
minimal performance overhead over two steps, that is, C -> eBPF -> native code.

Currently, the new format is being used for running user BPF programs, which
includes seccomp BPF, classic socket filters, cls_bpf traffic classifier,
team driver's classifier for its load-balancing mode, netfilter's xt_bpf
extension, PTP dissector/classifier, and much more. They are all internally
converted by the kernel into the new instruction set representation and run
in the eBPF interpreter. For in-kernel handlers, this all works transparently
by using sk_unattached_filter_create() for setting up the filter, resp.
sk_unattached_filter_destroy() for destroying it. The macro
SK_RUN_FILTER(filter, ctx) transparently invokes eBPF interpreter or JITed
code to run the filter. 'filter' is a pointer to struct sk_filter that we
got from sk_unattached_filter_create(), and 'ctx' the given context (e.g.
skb pointer). All constraints and restrictions from sk_chk_filter() apply
before a conversion to the new layout is being done behind the scenes!

Currently, the classic BPF format is being used for JITing on most of the
architectures. Only x86-64 performs JIT compilation from eBPF instruction set,
however, future work will migrate other JIT compilers as well, so that they
will profit from the very same benefits.

Some core changes of the new internal format:

- Number of registers increase from 2 to 10:

  The old format had two registers A and X, and a hidden frame pointer. The
  new layout extends this to be 10 internal registers and a read-only frame
  pointer. Since 64-bit CPUs are passing arguments to functions via registers
  the number of args from eBPF program to in-kernel function is restricted
  to 5 and one register is used to accept return value from an in-kernel
  function. Natively, x86_64 passes first 6 arguments in registers, aarch64/
  sparcv9/mips64 have 7 - 8 registers for arguments; x86_64 has 6 callee saved
  registers, and aarch64/sparcv9/mips64 have 11 or more callee saved registers.

  Therefore, eBPF calling convention is defined as:

    * R0	- return value from in-kernel function, and exit value for eBPF program
    * R1 - R5	- arguments from eBPF program to in-kernel function
    * R6 - R9	- callee saved registers that in-kernel function will preserve
    * R10	- read-only frame pointer to access stack

  Thus, all eBPF registers map one to one to HW registers on x86_64, aarch64,
  etc, and eBPF calling convention maps directly to ABIs used by the kernel on
  64-bit architectures.

  On 32-bit architectures JIT may map programs that use only 32-bit arithmetic
  and may let more complex programs to be interpreted.

  R0 - R5 are scratch registers and eBPF program needs spill/fill them if
  necessary across calls. Note that there is only one eBPF program (== one
  eBPF main routine) and it cannot call other eBPF functions, it can only
  call predefined in-kernel functions, though.

- Register width increases from 32-bit to 64-bit:

  Still, the semantics of the original 32-bit ALU operations are preserved
  via 32-bit subregisters. All eBPF registers are 64-bit with 32-bit lower
  subregisters that zero-extend into 64-bit if they are being written to.
  That behavior maps directly to x86_64 and arm64 subregister definition, but
  makes other JITs more difficult.

  32-bit architectures run 64-bit internal BPF programs via interpreter.
  Their JITs may convert BPF programs that only use 32-bit subregisters into
  native instruction set and let the rest being interpreted.

  Operation is 64-bit, because on 64-bit architectures, pointers are also
  64-bit wide, and we want to pass 64-bit values in/out of kernel functions,
  so 32-bit eBPF registers would otherwise require to define register-pair
  ABI, thus, there won't be able to use a direct eBPF register to HW register
  mapping and JIT would need to do combine/split/move operations for every
  register in and out of the function, which is complex, bug prone and slow.
  Another reason is the use of atomic 64-bit counters.

- Conditional jt/jf targets replaced with jt/fall-through:

  While the original design has constructs such as "if (cond) jump_true;
  else jump_false;", they are being replaced into alternative constructs like
  "if (cond) jump_true; /* else fall-through */".

- Introduces bpf_call insn and register passing convention for zero overhead
  calls from/to other kernel functions:

  Before an in-kernel function call, the internal BPF program needs to
  place function arguments into R1 to R5 registers to satisfy calling
  convention, then the interpreter will take them from registers and pass
  to in-kernel function. If R1 - R5 registers are mapped to CPU registers
  that are used for argument passing on given architecture, the JIT compiler
  doesn't need to emit extra moves. Function arguments will be in the correct
  registers and BPF_CALL instruction will be JITed as single 'call' HW
  instruction. This calling convention was picked to cover common call
  situations without performance penalty.

  After an in-kernel function call, R1 - R5 are reset to unreadable and R0 has
  a return value of the function. Since R6 - R9 are callee saved, their state
  is preserved across the call.

  For example, consider three C functions:

  u64 f1() { return (*_f2)(1); }
  u64 f2(u64 a) { return f3(a + 1, a); }
  u64 f3(u64 a, u64 b) { return a - b; }

  GCC can compile f1, f3 into x86_64:

  f1:
    movl $1, %edi
    movq _f2(%rip), %rax
    jmp  *%rax
  f3:
    movq %rdi, %rax
    subq %rsi, %rax
    ret

  Function f2 in eBPF may look like:

  f2:
    bpf_mov R2, R1
    bpf_add R1, 1
    bpf_call f3
    bpf_exit

  If f2 is JITed and the pointer stored to '_f2'. The calls f1 -> f2 -> f3 and
  returns will be seamless. Without JIT, __sk_run_filter() interpreter needs to
  be used to call into f2.

  For practical reasons all eBPF programs have only one argument 'ctx' which is
  already placed into R1 (e.g. on __sk_run_filter() startup) and the programs
  can call kernel functions with up to 5 arguments. Calls with 6 or more arguments
  are currently not supported, but these restrictions can be lifted if necessary
  in the future.

  On 64-bit architectures all register map to HW registers one to one. For
  example, x86_64 JIT compiler can map them as ...

    R0 - rax
    R1 - rdi
    R2 - rsi
    R3 - rdx
    R4 - rcx
    R5 - r8
    R6 - rbx
    R7 - r13
    R8 - r14
    R9 - r15
    R10 - rbp

  ... since x86_64 ABI mandates rdi, rsi, rdx, rcx, r8, r9 for argument passing
  and rbx, r12 - r15 are callee saved.

  Then the following internal BPF pseudo-program:

    bpf_mov R6, R1 /* save ctx */
    bpf_mov R2, 2
    bpf_mov R3, 3
    bpf_mov R4, 4
    bpf_mov R5, 5
    bpf_call foo
    bpf_mov R7, R0 /* save foo() return value */
    bpf_mov R1, R6 /* restore ctx for next call */
    bpf_mov R2, 6
    bpf_mov R3, 7
    bpf_mov R4, 8
    bpf_mov R5, 9
    bpf_call bar
    bpf_add R0, R7
    bpf_exit

  After JIT to x86_64 may look like:

    push %rbp
    mov %rsp,%rbp
    sub $0x228,%rsp
    mov %rbx,-0x228(%rbp)
    mov %r13,-0x220(%rbp)
    mov %rdi,%rbx
    mov $0x2,%esi
    mov $0x3,%edx
    mov $0x4,%ecx
    mov $0x5,%r8d
    callq foo
    mov %rax,%r13
    mov %rbx,%rdi
    mov $0x2,%esi
    mov $0x3,%edx
    mov $0x4,%ecx
    mov $0x5,%r8d
    callq bar
    add %r13,%rax
    mov -0x228(%rbp),%rbx
    mov -0x220(%rbp),%r13
    leaveq
    retq

  Which is in this example equivalent in C to:

    u64 bpf_filter(u64 ctx)
    {
        return foo(ctx, 2, 3, 4, 5) + bar(ctx, 6, 7, 8, 9);
    }

  In-kernel functions foo() and bar() with prototype: u64 (*)(u64 arg1, u64
  arg2, u64 arg3, u64 arg4, u64 arg5); will receive arguments in proper
  registers and place their return value into '%rax' which is R0 in eBPF.
  Prologue and epilogue are emitted by JIT and are implicit in the
  interpreter. R0-R5 are scratch registers, so eBPF program needs to preserve
  them across the calls as defined by calling convention.

  For example the following program is invalid:

    bpf_mov R1, 1
    bpf_call foo
    bpf_mov R0, R1
    bpf_exit

  After the call the registers R1-R5 contain junk values and cannot be read.
  In the future an eBPF verifier can be used to validate internal BPF programs.

Also in the new design, eBPF is limited to 4096 insns, which means that any
program will terminate quickly and will only call a fixed number of kernel
functions. Original BPF and the new format are two operand instructions,
which helps to do one-to-one mapping between eBPF insn and x86 insn during JIT.

The input context pointer for invoking the interpreter function is generic,
its content is defined by a specific use case. For seccomp register R1 points
to seccomp_data, for converted BPF filters R1 points to a skb.

A program, that is translated internally consists of the following elements:

  op:16, jt:8, jf:8, k:32    ==>    op:8, dst_reg:4, src_reg:4, off:16, imm:32

So far 87 internal BPF instructions were implemented. 8-bit 'op' opcode field
has room for new instructions. Some of them may use 16/24/32 byte encoding. New
instructions must be multiple of 8 bytes to preserve backward compatibility.

Internal BPF is a general purpose RISC instruction set. Not every register and
every instruction are used during translation from original BPF to new format.
For example, socket filters are not using 'exclusive add' instruction, but
tracing filters may do to maintain counters of events, for example. Register R9
is not used by socket filters either, but more complex filters may be running
out of registers and would have to resort to spill/fill to stack.

Internal BPF can used as generic assembler for last step performance
optimizations, socket filters and seccomp are using it as assembler. Tracing
filters may use it as assembler to generate code from kernel. In kernel usage
may not be bounded by security considerations, since generated internal BPF code
may be optimizing internal code path and not being exposed to the user space.
Safety of internal BPF can come from a verifier (TBD). In such use cases as
described, it may be used as safe instruction set.

Just like the original BPF, the new format runs within a controlled environment,
is deterministic and the kernel can easily prove that. The safety of the program
can be determined in two steps: first step does depth-first-search to disallow
loops and other CFG validation; second step starts from the first insn and
descends all possible paths. It simulates execution of every insn and observes
the state change of registers and stack.

eBPF opcode encoding
--------------------

eBPF is reusing most of the opcode encoding from classic to simplify conversion
of classic BPF to eBPF. For arithmetic and jump instructions the 8-bit 'code'
field is divided into three parts:

  +----------------+--------+--------------------+
  |   4 bits       |  1 bit |   3 bits           |
  | operation code | source | instruction class  |
  +----------------+--------+--------------------+
  (MSB)                                      (LSB)

Three LSB bits store instruction class which is one of:

  Classic BPF classes:    eBPF classes:

  BPF_LD    0x00          BPF_LD    0x00
  BPF_LDX   0x01          BPF_LDX   0x01
  BPF_ST    0x02          BPF_ST    0x02
  BPF_STX   0x03          BPF_STX   0x03
  BPF_ALU   0x04          BPF_ALU   0x04
  BPF_JMP   0x05          BPF_JMP   0x05
  BPF_RET   0x06          [ class 6 unused, for future if needed ]
  BPF_MISC  0x07          BPF_ALU64 0x07

When BPF_CLASS(code) == BPF_ALU or BPF_JMP, 4th bit encodes source operand ...

  BPF_K     0x00
  BPF_X     0x08

 * in classic BPF, this means:

  BPF_SRC(code) == BPF_X - use register X as source operand
  BPF_SRC(code) == BPF_K - use 32-bit immediate as source operand

 * in eBPF, this means:

  BPF_SRC(code) == BPF_X - use 'src_reg' register as source operand
  BPF_SRC(code) == BPF_K - use 32-bit immediate as source operand

... and four MSB bits store operation code.

If BPF_CLASS(code) == BPF_ALU or BPF_ALU64 [ in eBPF ], BPF_OP(code) is one of:

  BPF_ADD   0x00
  BPF_SUB   0x10
  BPF_MUL   0x20
  BPF_DIV   0x30
  BPF_OR    0x40
  BPF_AND   0x50
  BPF_LSH   0x60
  BPF_RSH   0x70
  BPF_NEG   0x80
  BPF_MOD   0x90
  BPF_XOR   0xa0
  BPF_MOV   0xb0  /* eBPF only: mov reg to reg */
  BPF_ARSH  0xc0  /* eBPF only: sign extending shift right */
  BPF_END   0xd0  /* eBPF only: endianness conversion */

If BPF_CLASS(code) == BPF_JMP, BPF_OP(code) is one of:

  BPF_JA    0x00
  BPF_JEQ   0x10
  BPF_JGT   0x20
  BPF_JGE   0x30
  BPF_JSET  0x40
  BPF_JNE   0x50  /* eBPF only: jump != */
  BPF_JSGT  0x60  /* eBPF only: signed '>' */
  BPF_JSGE  0x70  /* eBPF only: signed '>=' */
  BPF_CALL  0x80  /* eBPF only: function call */
  BPF_EXIT  0x90  /* eBPF only: function return */

So BPF_ADD | BPF_X | BPF_ALU means 32-bit addition in both classic BPF
and eBPF. There are only two registers in classic BPF, so it means A += X.
In eBPF it means dst_reg = (u32) dst_reg + (u32) src_reg; similarly,
BPF_XOR | BPF_K | BPF_ALU means A ^= imm32 in classic BPF and analogous
src_reg = (u32) src_reg ^ (u32) imm32 in eBPF.

Classic BPF is using BPF_MISC class to represent A = X and X = A moves.
eBPF is using BPF_MOV | BPF_X | BPF_ALU code instead. Since there are no
BPF_MISC operations in eBPF, the class 7 is used as BPF_ALU64 to mean
exactly the same operations as BPF_ALU, but with 64-bit wide operands
instead. So BPF_ADD | BPF_X | BPF_ALU64 means 64-bit addition, i.e.:
dst_reg = dst_reg + src_reg

Classic BPF wastes the whole BPF_RET class to represent a single 'ret'
operation. Classic BPF_RET | BPF_K means copy imm32 into return register
and perform function exit. eBPF is modeled to match CPU, so BPF_JMP | BPF_EXIT
in eBPF means function exit only. The eBPF program needs to store return
value into register R0 before doing a BPF_EXIT. Class 6 in eBPF is currently
unused and reserved for future use.

For load and store instructions the 8-bit 'code' field is divided as:

  +--------+--------+-------------------+
  | 3 bits | 2 bits |   3 bits          |
  |  mode  |  size  | instruction class |
  +--------+--------+-------------------+
  (MSB)                             (LSB)

Size modifier is one of ...

  BPF_W   0x00    /* word */
  BPF_H   0x08    /* half word */
  BPF_B   0x10    /* byte */
  BPF_DW  0x18    /* eBPF only, double word */

... which encodes size of load/store operation:

 B  - 1 byte
 H  - 2 byte
 W  - 4 byte
 DW - 8 byte (eBPF only)

Mode modifier is one of:

  BPF_IMM  0x00  /* classic BPF only, reserved in eBPF */
  BPF_ABS  0x20
  BPF_IND  0x40
  BPF_MEM  0x60
  BPF_LEN  0x80  /* classic BPF only, reserved in eBPF */
  BPF_MSH  0xa0  /* classic BPF only, reserved in eBPF */
  BPF_XADD 0xc0  /* eBPF only, exclusive add */

eBPF has two non-generic instructions: (BPF_ABS | <size> | BPF_LD) and
(BPF_IND | <size> | BPF_LD) which are used to access packet data.

They had to be carried over from classic to have strong performance of
socket filters running in eBPF interpreter. These instructions can only
be used when interpreter context is a pointer to 'struct sk_buff' and
have seven implicit operands. Register R6 is an implicit input that must
contain pointer to sk_buff. Register R0 is an implicit output which contains
the data fetched from the packet. Registers R1-R5 are scratch registers
and must not be used to store the data across BPF_ABS | BPF_LD or
BPF_IND | BPF_LD instructions.

These instructions have implicit program exit condition as well. When
eBPF program is trying to access the data beyond the packet boundary,
the interpreter will abort the execution of the program. JIT compilers
therefore must preserve this property. src_reg and imm32 fields are
explicit inputs to these instructions.

For example:

  BPF_IND | BPF_W | BPF_LD means:

    R0 = ntohl(*(u32 *) (((struct sk_buff *) R6)->data + src_reg + imm32))
    and R1 - R5 were scratched.

Unlike classic BPF instruction set, eBPF has generic load/store operations:

BPF_MEM | <size> | BPF_STX:  *(size *) (dst_reg + off) = src_reg
BPF_MEM | <size> | BPF_ST:   *(size *) (dst_reg + off) = imm32
BPF_MEM | <size> | BPF_LDX:  dst_reg = *(size *) (src_reg + off)
BPF_XADD | BPF_W  | BPF_STX: lock xadd *(u32 *)(dst_reg + off16) += src_reg
BPF_XADD | BPF_DW | BPF_STX: lock xadd *(u64 *)(dst_reg + off16) += src_reg

Where size is one of: BPF_B or BPF_H or BPF_W or BPF_DW. Note that 1 and
2 byte atomic increments are not supported.

Testing
-------

Next to the BPF toolchain, the kernel also ships a test module that contains
various test cases for classic and internal BPF that can be executed against
the BPF interpreter and JIT compiler. It can be found in lib/test_bpf.c and
enabled via Kconfig:

  CONFIG_TEST_BPF=m

After the module has been built and installed, the test suite can be executed
via insmod or modprobe against 'test_bpf' module. Results of the test cases
including timings in nsec can be found in the kernel log (dmesg).

Misc
----

Also trinity, the Linux syscall fuzzer, has built-in support for BPF and
SECCOMP-BPF kernel fuzzing.

Written by
----------

The document was written in the hope that it is found useful and in order
to give potential BPF hackers or security auditors a better overview of
the underlying architecture.

Jay Schulist <[e-mail]>
Daniel Borkmann <[e-mail]>
Alexei Starovoitov <[e-mail]>