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 LinuxKernel filtering, but when we speak of BPF or LSF in Linux context, wemean the very same mechanism of filtering in the Linux kernel.

BPF allows a user-space program to attach a filter onto any socket andallow or disallow certain types of data to come through the socket. LSFfollows exactly the same filter code structure as BSD’s BPF, so referringto 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 worryabout 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 filtercode passes the kernel check on it, you then immediately begin filteringdata on that socket.

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

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

The biggest user of this construct might be libpcap. Issuing a high-levelfilter command liketcpdump -i em1 port 22 passes through the libpcapinternal compiler that generates a structure that can eventually be loadedvia SO_ATTACH_FILTER to the kernel.tcpdump -i em1 port 22 -ddddisplays what is being placed into this structure.

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

[1]Documentation/userspace-api/seccomp_filter.rst

Original BPF paper:

Steven McCanne and Van Jacobson. 1993. The BSD packet filter: a newarchitecture for user-level packet capture. In Proceedings of theUSENIX Winter 1993 Conference Proceedings on USENIX Winter 1993Conference 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 thefollowing 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 containsa code, jt, jf and k value. jt and jf are jump offsets and k a genericvalue 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 infollow-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 socketin order to let all IPv4/IPv6 packets with port 22 pass. The rest willbe dropped for this socket.

The setsockopt(2) call to SO_DETACH_FILTER doesn’t need any argumentsand SO_LOCK_FILTER for preventing the filter to be detached, takes aninteger 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 becovered by libpcap in high-level syntax, so as an application developeryou 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 BPFfilters use Linux extensions that are not supported by libpcap’s compiler,iii) a filter might be more complex and not cleanly implementable withlibpcap’s compiler, or iv) particular filter codes should be optimizeddifferently than libpcap’s internal compiler does; then in such caseswriting such a filter “by hand” can be of an alternative. For example,xt_bpf and cls_bpf users might have requirements that could result inmore complex filter code, or one that cannot be expressed with libpcap(e.g. different return codes for various code paths). Moreover, BPF JITimplementors may wish to manually write test cases and thus need low-levelaccess to BPF code as well.

BPF engine and instruction set

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

The BPF architecture consists of the following basic elements:

ElementDescription
A32 bit wide accumulator
X32 bit wide X register
M[]16 x 32 bit wide misc registers aka “scratch memorystore”, addressable from 0 to 15

A program, that is translated by bpf_asm into “opcodes” is an array thatconsists 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 instructionencoded. jt and jf are two 8 bit wide jump targets, one for condition“jump if true”, the other one “jump if false”. Eventually, element kcontains a miscellaneous argument that can be interpreted in differentways depending on the given instruction in op.

The instruction set consists of load, store, branch, alu, miscellaneousand return instructions that are also represented in bpf_asm syntax. Thistable lists all bpf_asm instructions available resp. what their underlyingopcodes as defined in linux/filter.h stand for:

InstructionAddressing modeDescription
ld1, 2, 3, 4, 12Load word into A
ldi4Load word into A
ldh1, 2Load half-word into A
ldb1, 2Load byte into A
ldx3, 4, 5, 12Load word into X
ldxi4Load word into X
ldxb5Load byte into X
st3Store A into M[]
stx3Store X into M[]
jmp6Jump to label
ja6Jump to label
jeq7, 8, 9, 10Jump on A == <x>
jneq9, 10Jump on A != <x>
jne9, 10Jump on A != <x>
jlt9, 10Jump on A < <x>
jle9, 10Jump on A <= <x>
jgt7, 8, 9, 10Jump on A > <x>
jge7, 8, 9, 10Jump on A >= <x>
jset7, 8, 9, 10Jump on A & <x>
add0, 4A + <x>
sub0, 4A - <x>
mul0, 4A * <x>
div0, 4A / <x>
mod0, 4A % <x>
neg !A
and0, 4A & <x>
or0, 4A | <x>
xor0, 4A ^ <x>
lsh0, 4A << <x>
rsh0, 4A >> <x>
tax Copy A into X
txa Copy X into A
ret4, 11Return

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

Addressing modeSyntaxDescription
0x/%xRegister X
1[k]BHW at byte offset k in the packet
2[x + k]BHW at the offset X + k in the packet
3M[k]Word at offset k in M[]
4#kLiteral value stored in k
54*([k]&0xf)Lower nibble * 4 at byte offset k in the packet
6LJump label L
7#k,Lt,LfJump to Lt if true, otherwise jump to Lf
8x/%x,Lt,LfJump to Lt if true, otherwise jump to Lf
9#k,LtJump to Lt if predicate is true
10x/%x,LtJump to Lt if predicate is true
11a/%aAccumulator A
12extensionBPF extension

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

Possible BPF extensions are shown in the following table:

ExtensionDescription
lenskb->len
protoskb->protocol
typeskb->pkt_type
poffPayload start offset
ifidxskb->dev->ifindex
nlaNetlink attribute of type X with offset A
nlanNested Netlink attribute of type X with offset A
markskb->mark
queueskb->queue_mapping
hatypeskb->dev->type
rxhashskb->hash
cpuraw_smp_processor_id()
vlan_tciskb_vlan_tag_get(skb)
vlan_availskb_vlan_tag_present(skb)
vlan_tpidskb->vlan_proto
randprandom_u32()

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

ARP packets:

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

IPv4 TCP packets:

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

(Accelerated) VLAN w/ id 10:

ld vlan_tcijneq #10, dropret #-1drop: ret #0

icmp random packet sampling, 1 in 4:

ldh [12]jne #0x800, dropldb [23]jneq #1, drop# get a random uint32 numberld randmod #4jneq #1, dropret #-1drop: 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_THREAD */good: ret #0x7fff0000   /* SECCOMP_RET_ALLOW */

The above example code can be placed into a file (here called “foo”), andthen be passed to the bpf_asm tool for generating opcodes, output that xt_bpfand cls_bpf understands and can directly be loaded with. Example with aboveARP code:

$ ./bpf_asm foo4,40 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 BPFfilters that might not be obvious at first, it’s good to test filters beforeattaching to a live system. For that purpose, there’s a small tool calledbpf_dbg under tools/bpf/ in the kernel source directory. This debugger allowsfor testing BPF filters against given pcap files, single stepping through theBPF 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 analternative stdin source as a first argument, and an alternative stdoutsink as a second one, e.g../bpf_dbg test_in.txt test_out.txt.

Other than that, a particular libreadline configuration can be set viafile “~/.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-completionsupport (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 0Loads a BPF filter from standard output of bpf_asm, or transformed viae.g.tcpdump-iem1-dddport22|tr'\n'','. Note that for JITdebugging (next section), this command creates a temporary socket andloads the BPF code into the kernel. Thus, this will also be useful forJIT 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 failsthe filter will generate. A limit of packets to traverse can be given.

  • disassemble:

    l0:     ldh [12]l1:     jeq #0x800, l2, l5l2:     ldb [23]l3:     jeq #0x1, l4, l5l4:     ret #0xffffl5:     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 arun commandwill walk through the pcap file continuing from the current packet andbreak when a breakpoint is being hit (anotherrun will continue fromthe currently active breakpoint executing next instructions):

    • run:

      -- register dump --pc:       [0]                       <-- program countercode:     [40] jt[0] jf[0] k[12]    <-- plain BPF code of current instructioncurr:     l0:   ldh [12]              <-- disassembly of current instructionA:        [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 0116: 08 00 06 04 00 01 00 14 a4 43 78 69 0a 3b 01 2632: 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 pcoffset. Thus, on each step invocation, above register dump is issued.This can go forwards and backwards in time, a plainstep will breakon the next BPF instruction, thus +1. (Norun needs to be issued here.)

  • select <n>

    Selects a given packet from the pcap file to continue from. Thus, onthe nextrun orstep, the BPF program is being evaluated againstthe user pre-selected packet. Numbering starts just as in Wiresharkwith index 1.

  • quit

    Exits bpf_dbg.

JIT compiler

The Linux kernel has a built-in BPF JIT compiler for x86_64, SPARC,PowerPC, ARM, ARM64, MIPS, RISC-V and s390 and can be enabled throughCONFIG_BPF_JIT. The JIT compiler is transparently invoked for eachattached filter from user space or for internal kernel users if it hasbeen 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 generatedopcode 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

When CONFIG_BPF_JIT_ALWAYS_ON is enabled, bpf_jit_enable is permanently set to 1 andsetting any other value than that will return in failure. This is even the case forsetting bpf_jit_enable to 2, since dumping the final JIT image into the kernel logis discouraged and introspection through bpftool (under tools/bpf/bpftool/) is thegenerally recommended approach instead.

In the kernel source tree under tools/bpf/, there’s bpf_jit_disasm forgenerating disassembly out of the kernel log’s hexdump:

# ./bpf_jit_disasm70 bytes emitted from JIT compiler (pass:3, flen:6)ffffffffa0069c8f + <x>:0:      push   %rbp1:      mov    %rsp,%rbp4:      sub    $0x60,%rsp8:      mov    %rbx,-0x8(%rbp)c:      mov    0x68(%rdi),%r9d10:     sub    0x6c(%rdi),%r9d14:     mov    0xd8(%rdi),%r81b:     mov    $0xc,%esi20:     callq  0xffffffffe0ff944225:     cmp    $0x800,%eax2a:     jne    0x00000000000000422c:     mov    $0x17,%esi31:     callq  0xffffffffe0ff945e36:     cmp    $0x1,%eax39:     jne    0x00000000000000423b:     mov    $0xffff,%eax40:     jmp    0x000000000000004442:     xor    %eax,%eax44:     leaveq45:     retqIssuing option `-o` will "annotate" opcodes to resulting assemblerinstructions, which can be very useful for JIT developers:# ./bpf_jit_disasm -o70 bytes emitted from JIT compiler (pass:3, flen:6)ffffffffa0069c8f + <x>:0:      push   %rbp        551:      mov    %rsp,%rbp        48 89 e54:      sub    $0x60,%rsp        48 83 ec 608:      mov    %rbx,-0x8(%rbp)        48 89 5d f8c:      mov    0x68(%rdi),%r9d        44 8b 4f 6810:     sub    0x6c(%rdi),%r9d        44 2b 4f 6c14:     mov    0xd8(%rdi),%r8        4c 8b 87 d8 00 00 001b:     mov    $0xc,%esi        be 0c 00 00 0020:     callq  0xffffffffe0ff9442        e8 1d 94 ff e025:     cmp    $0x800,%eax        3d 00 08 00 002a:     jne    0x0000000000000042        75 162c:     mov    $0x17,%esi        be 17 00 00 0031:     callq  0xffffffffe0ff945e        e8 28 94 ff e036:     cmp    $0x1,%eax        83 f8 0139:     jne    0x0000000000000042        75 073b:     mov    $0xffff,%eax        b8 ff ff 00 0040:     jmp    0x0000000000000044        eb 0242:     xor    %eax,%eax        31 c044:     leaveq        c945:     retq        c3

For BPF JIT developers, bpf_jit_disasm, bpf_asm and bpf_dbg provides a usefultoolchain for developing and testing the kernel’s JIT compiler.

BPF kernel internals

Internally, for the kernel interpreter, a different instruction setformat with similar underlying principles from BPF described in previousparagraphs is being used. However, the instruction set format is modelledcloser to the underlying architecture to mimic native instruction sets, sothat a better performance can be achieved (more details later). This newISA is called ‘eBPF’ or ‘internal BPF’ interchangeably. (Note: eBPF whichoriginates from [e]xtended BPF is not the same as BPF extensions! WhileeBPF 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 upthe possibility for GCC/LLVM compilers to generate optimized eBPF code throughan eBPF backend that performs almost as fast as natively compiled code.

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

Currently, the new format is being used for running user BPF programs, whichincludes seccomp BPF, classic socket filters, cls_bpf traffic classifier,team driver’s classifier for its load-balancing mode, netfilter’s xt_bpfextension, PTP dissector/classifier, and much more. They are all internallyconverted by the kernel into the new instruction set representation and runin the eBPF interpreter. For in-kernel handlers, this all works transparentlyby usingbpf_prog_create() for setting up the filter, resp.bpf_prog_destroy() for destroying it. The macroBPF_PROG_RUN(filter, ctx) transparently invokes eBPF interpreter or JITedcode to run the filter. ‘filter’ is a pointer to struct bpf_prog that wegot frombpf_prog_create(), and ‘ctx’ the given context (e.g.skb pointer). All constraints and restrictions from bpf_check_classic() applybefore a conversion to the new layout is being done behind the scenes!

Currently, the classic BPF format is being used for JITing on most32-bit architectures, whereas x86-64, aarch64, s390x, powerpc64,sparc64, arm32, riscv64, riscv32 perform JIT compilation from eBPFinstruction set.

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. Thenew layout extends this to be 10 internal registers and a read-only framepointer. Since 64-bit CPUs are passing arguments to functions via registersthe number of args from eBPF program to in-kernel function is restrictedto 5 and one register is used to accept return value from an in-kernelfunction. Natively, x86_64 passes first 6 arguments in registers, aarch64/sparcv9/mips64 have 7 - 8 registers for arguments; x86_64 has 6 callee savedregisters, 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 on64-bit architectures.

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

    R0 - R5 are scratch registers and eBPF program needs spill/fill them ifnecessary across calls. Note that there is only one eBPF program (== oneeBPF main routine) and it cannot call other eBPF functions, it can onlycall 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 preservedvia 32-bit subregisters. All eBPF registers are 64-bit with 32-bit lowersubregisters that zero-extend into 64-bit if they are being written to.That behavior maps directly to x86_64 and arm64 subregister definition, butmakes 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 intonative instruction set and let the rest being interpreted.

    Operation is 64-bit, because on 64-bit architectures, pointers are also64-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-pairABI, thus, there won’t be able to use a direct eBPF register to HW registermapping and JIT would need to do combine/split/move operations for everyregister 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 asif(cond)jump_true;elsejump_false;, they are being replaced into alternative constructs likeif(cond)jump_true;/*elsefall-through*/.

  • Introduces bpf_call insn and register passing convention for zero overheadcalls from/to other kernel functions:

    Before an in-kernel function call, the internal BPF program needs toplace function arguments into R1 to R5 registers to satisfy callingconvention, then the interpreter will take them from registers and passto in-kernel function. If R1 - R5 registers are mapped to CPU registersthat are used for argument passing on given architecture, the JIT compilerdoesn’t need to emit extra moves. Function arguments will be in the correctregisters and BPF_CALL instruction will be JITed as single ‘call’ HWinstruction. This calling convention was picked to cover common callsituations without performance penalty.

    After an in-kernel function call, R1 - R5 are reset to unreadable and R0 hasa return value of the function. Since R6 - R9 are callee saved, their stateis 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  *%raxf3:    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 andreturns will be seamless. Without JIT, __bpf_prog_run() interpreter needs tobe used to call into f2.

    For practical reasons all eBPF programs have only one argument ‘ctx’ which isalready placed into R1 (e.g. on __bpf_prog_run() startup) and the programscan call kernel functions with up to 5 arguments. Calls with 6 or more argumentsare currently not supported, but these restrictions can be lifted if necessaryin the future.

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

    R0 - raxR1 - rdiR2 - rsiR3 - rdxR4 - rcxR5 - r8R6 - rbxR7 - r13R8 - r14R9 - r15R10 - rbp

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

    Then the following internal BPF pseudo-program:

    bpf_mov R6, R1 /* save ctx */bpf_mov R2, 2bpf_mov R3, 3bpf_mov R4, 4bpf_mov R5, 5bpf_call foobpf_mov R7, R0 /* save foo() return value */bpf_mov R1, R6 /* restore ctx for next call */bpf_mov R2, 6bpf_mov R3, 7bpf_mov R4, 8bpf_mov R5, 9bpf_call barbpf_add R0, R7bpf_exit

    After JIT to x86_64 may look like:

    push %rbpmov %rsp,%rbpsub $0x228,%rspmov %rbx,-0x228(%rbp)mov %r13,-0x220(%rbp)mov %rdi,%rbxmov $0x2,%esimov $0x3,%edxmov $0x4,%ecxmov $0x5,%r8dcallq foomov %rax,%r13mov %rbx,%rdimov $0x6,%esimov $0x7,%edxmov $0x8,%ecxmov $0x9,%r8dcallq baradd %r13,%raxmov -0x228(%rbp),%rbxmov -0x220(%rbp),%r13leaveqretq

    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, u64arg2, u64 arg3, u64 arg4, u64 arg5); will receive arguments in properregisters and place their return value into%rax which is R0 in eBPF.Prologue and epilogue are emitted by JIT and are implicit in theinterpreter. R0-R5 are scratch registers, so eBPF program needs to preservethem across the calls as defined by calling convention.

    For example the following program is invalid:

    bpf_mov R1, 1bpf_call foobpf_mov R0, R1bpf_exit

    After the call the registers R1-R5 contain junk values and cannot be read.An in-kernel eBPF verifier is used to validate internal BPF programs.

Also in the new design, eBPF is limited to 4096 insns, which means that anyprogram will terminate quickly and will only call a fixed number of kernelfunctions. 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 pointsto 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 fieldhas room for new instructions. Some of them may use 16/24/32 byte encoding. Newinstructions must be multiple of 8 bytes to preserve backward compatibility.

Internal BPF is a general purpose RISC instruction set. Not every register andevery instruction are used during translation from original BPF to new format.For example, socket filters are not usingexclusiveadd instruction, buttracing filters may do to maintain counters of events, for example. Register R9is not used by socket filters either, but more complex filters may be runningout of registers and would have to resort to spill/fill to stack.

Internal BPF can be used as a generic assembler for last step performanceoptimizations, socket filters and seccomp are using it as assembler. Tracingfilters may use it as assembler to generate code from kernel. In kernel usagemay not be bounded by security considerations, since generated internal BPF codemay 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 asdescribed, 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 programcan be determined in two steps: first step does depth-first-search to disallowloops and other CFG validation; second step starts from the first insn anddescends all possible paths. It simulates execution of every insn and observesthe state change of registers and stack.

eBPF opcode encoding

eBPF is reusing most of the opcode encoding from classic to simplify conversionof 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 classeseBPF classes
BPF_LD 0x00BPF_LD 0x00
BPF_LDX 0x01BPF_LDX 0x01
BPF_ST 0x02BPF_ST 0x02
BPF_STX 0x03BPF_STX 0x03
BPF_ALU 0x04BPF_ALU 0x04
BPF_JMP 0x05BPF_JMP 0x05
BPF_RET 0x06BPF_JMP32 0x06
BPF_MISC 0x07BPF_ALU64 0x07

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

BPF_K     0x00BPF_X     0x08

  • in classic BPF, this means:

    BPF_SRC(code) == BPF_X - use register X as source operandBPF_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 operandBPF_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   0x00BPF_SUB   0x10BPF_MUL   0x20BPF_DIV   0x30BPF_OR    0x40BPF_AND   0x50BPF_LSH   0x60BPF_RSH   0x70BPF_NEG   0x80BPF_MOD   0x90BPF_XOR   0xa0BPF_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 or BPF_JMP32 [ in eBPF ], BPF_OP(code) is one of:

BPF_JA    0x00  /* BPF_JMP only */BPF_JEQ   0x10BPF_JGT   0x20BPF_JGE   0x30BPF_JSET  0x40BPF_JNE   0x50  /* eBPF only: jump != */BPF_JSGT  0x60  /* eBPF only: signed '>' */BPF_JSGE  0x70  /* eBPF only: signed '>=' */BPF_CALL  0x80  /* eBPF BPF_JMP only: function call */BPF_EXIT  0x90  /* eBPF BPF_JMP only: function return */BPF_JLT   0xa0  /* eBPF only: unsigned '<' */BPF_JLE   0xb0  /* eBPF only: unsigned '<=' */BPF_JSLT  0xc0  /* eBPF only: signed '<' */BPF_JSLE  0xd0  /* eBPF only: signed '<=' */

So BPF_ADD | BPF_X | BPF_ALU means 32-bit addition in both classic BPFand 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 analogoussrc_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 noBPF_MISC operations in eBPF, the class 7 is used as BPF_ALU64 to meanexactly the same operations as BPF_ALU, but with 64-bit wide operandsinstead. 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 singleretoperation. Classic BPF_RET | BPF_K means copy imm32 into return registerand perform function exit. eBPF is modeled to match CPU, so BPF_JMP | BPF_EXITin eBPF means function exit only. The eBPF program needs to store returnvalue into register R0 before doing a BPF_EXIT. Class 6 in eBPF is used asBPF_JMP32 to mean exactly the same operations as BPF_JMP, but with 32-bit wideoperands for the comparisons instead.

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 byteH  - 2 byteW  - 4 byteDW - 8 byte (eBPF only)

Mode modifier is one of:

BPF_IMM  0x00  /* used for 32-bit mov in classic BPF and 64-bit in eBPF */BPF_ABS  0x20BPF_IND  0x40BPF_MEM  0x60BPF_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 ofsocket filters running in eBPF interpreter. These instructions can onlybe used when interpreter context is a pointer tostructsk_buff andhave seven implicit operands. Register R6 is an implicit input that mustcontain pointer to sk_buff. Register R0 is an implicit output which containsthe data fetched from the packet. Registers R1-R5 are scratch registersand must not be used to store the data across BPF_ABS | BPF_LD orBPF_IND | BPF_LD instructions.

These instructions have implicit program exit condition as well. WheneBPF program is trying to access the data beyond the packet boundary,the interpreter will abort the execution of the program. JIT compilerstherefore must preserve this property. src_reg and imm32 fields areexplicit 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_regBPF_MEM | <size> | BPF_ST:   *(size *) (dst_reg + off) = imm32BPF_MEM | <size> | BPF_LDX:  dst_reg = *(size *) (src_reg + off)BPF_XADD | BPF_W  | BPF_STX: lock xadd *(u32 *)(dst_reg + off16) += src_regBPF_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 and2 byte atomic increments are not supported.

eBPF has one 16-byte instruction: BPF_LD | BPF_DW | BPF_IMM which consistsof two consecutivestructbpf_insn 8-byte blocks and interpreted as singleinstruction that loads 64-bit immediate value into a dst_reg.Classic BPF has similar instruction: BPF_LD | BPF_W | BPF_IMM which loads32-bit immediate value into a register.

eBPF verifier

The safety of the eBPF program is determined in two steps.

First step does DAG check to disallow loops and other CFG validation.In particular it will detect programs that have unreachable instructions.(though classic BPF checker allows them)

Second step starts from the first insn and descends all possible paths.It simulates execution of every insn and observes the state change ofregisters and stack.

At the start of the program the register R1 contains a pointer to contextand has type PTR_TO_CTX.If verifier sees an insn that does R2=R1, then R2 has now typePTR_TO_CTX as well and can be used on the right hand side of expression.If R1=PTR_TO_CTX and insn is R2=R1+R1, then R2=SCALAR_VALUE,since addition of two valid pointers makes invalid pointer.(In ‘secure’ mode verifier will reject any type of pointer arithmetic to makesure that kernel addresses don’t leak to unprivileged users)

If register was never written to, it’s not readable:

bpf_mov R0 = R2bpf_exit

will be rejected, since R2 is unreadable at the start of the program.

After kernel function call, R1-R5 are reset to unreadable andR0 has a return type of the function.

Since R6-R9 are callee saved, their state is preserved across the call.

bpf_mov R6 = 1bpf_call foobpf_mov R0 = R6bpf_exit

is a correct program. If there was R1 instead of R6, it would havebeen rejected.

load/store instructions are allowed only with registers of valid types, whichare PTR_TO_CTX, PTR_TO_MAP, PTR_TO_STACK. They are bounds and alignment checked.For example:

bpf_mov R1 = 1bpf_mov R2 = 2bpf_xadd *(u32 *)(R1 + 3) += R2bpf_exit

will be rejected, since R1 doesn’t have a valid pointer type at the time ofexecution of instruction bpf_xadd.

At the start R1 type is PTR_TO_CTX (a pointer to genericstructbpf_context)A callback is used to customize verifier to restrict eBPF program access to onlycertain fields within ctx structure with specified size and alignment.

For example, the following insn:

bpf_ld R0 = *(u32 *)(R6 + 8)

intends to load a word from address R6 + 8 and store it into R0If R6=PTR_TO_CTX, via is_valid_access() callback the verifier will knowthat offset 8 of size 4 bytes can be accessed for reading, otherwisethe verifier will reject the program.If R6=PTR_TO_STACK, then access should be aligned and be withinstack bounds, which are [-MAX_BPF_STACK, 0). In this example offset is 8,so it will fail verification, since it’s out of bounds.

The verifier will allow eBPF program to read data from stack only afterit wrote into it.

Classic BPF verifier does similar check with M[0-15] memory slots.For example:

bpf_ld R0 = *(u32 *)(R10 - 4)bpf_exit

is invalid program.Though R10 is correct read-only register and has type PTR_TO_STACKand R10 - 4 is within stack bounds, there were no stores into that location.

Pointer register spill/fill is tracked as well, since four (R6-R9)callee saved registers may not be enough for some programs.

Allowed function calls are customized with bpf_verifier_ops->get_func_proto()The eBPF verifier will check that registers match argument constraints.After the call register R0 will be set to return type of the function.

Function calls is a main mechanism to extend functionality of eBPF programs.Socket filters may let programs to call one set of functions, whereas tracingfilters may allow completely different set.

If a function made accessible to eBPF program, it needs to be thought throughfrom safety point of view. The verifier will guarantee that the function iscalled with valid arguments.

seccomp vs socket filters have different security restrictions for classic BPF.Seccomp solves this by two stage verifier: classic BPF verifier is followedby seccomp verifier. In case of eBPF one configurable verifier is shared forall use cases.

See details of eBPF verifier in kernel/bpf/verifier.c

Register value tracking

In order to determine the safety of an eBPF program, the verifier must trackthe range of possible values in each register and also in each stack slot.This is done withstructbpf_reg_state, defined in include/linux/bpf_verifier.h, which unifies tracking of scalar and pointer values. Eachregister state has a type, which is either NOT_INIT (the register has not beenwritten to), SCALAR_VALUE (some value which is not usable as a pointer), or apointer type. The types of pointers describe their base, as follows:

PTR_TO_CTX
Pointer to bpf_context.
CONST_PTR_TO_MAP
Pointer to struct bpf_map. “Const” because arithmeticon these pointers is forbidden.
PTR_TO_MAP_VALUE
Pointer to the value stored in a map element.
PTR_TO_MAP_VALUE_OR_NULL
Either a pointer to a map value, or NULL; map accesses(see section ‘eBPF maps’, below) return this type,which becomes a PTR_TO_MAP_VALUE when checked != NULL.Arithmetic on these pointers is forbidden.
PTR_TO_STACK
Frame pointer.
PTR_TO_PACKET
skb->data.
PTR_TO_PACKET_END
skb->data + headlen; arithmetic forbidden.
PTR_TO_SOCKET
Pointer to struct bpf_sock_ops, implicitly refcounted.
PTR_TO_SOCKET_OR_NULL
Either a pointer to a socket, or NULL; socket lookupreturns this type, which becomes a PTR_TO_SOCKET whenchecked != NULL. PTR_TO_SOCKET is reference-counted,so programs must release the reference through thesocket release function before the end of the program.Arithmetic on these pointers is forbidden.

However, a pointer may be offset from this base (as a result of pointerarithmetic), and this is tracked in two parts: the ‘fixed offset’ and ‘variableoffset’. The former is used when an exactly-known value (e.g. an immediateoperand) is added to a pointer, while the latter is used for values which arenot exactly known. The variable offset is also used in SCALAR_VALUEs, to trackthe range of possible values in the register.

The verifier’s knowledge about the variable offset consists of:

  • minimum and maximum values as unsigned
  • minimum and maximum values as signed
  • knowledge of the values of individual bits, in the form of a ‘tnum’: a u64‘mask’ and a u64 ‘value’. 1s in the mask represent bits whose value is unknown;1s in the value represent bits known to be 1. Bits known to be 0 have 0 in bothmask and value; no bit should ever be 1 in both. For example, if a byte is readinto a register from memory, the register’s top 56 bits are known zero, whilethe low 8 are unknown - which is represented as the tnum (0x0; 0xff). If wethen OR this with 0x40, we get (0x40; 0xbf), then if we add 1 we get (0x0;0x1ff), because of potential carries.

Besides arithmetic, the register state can also be updated by conditionalbranches. For instance, if a SCALAR_VALUE is compared > 8, in the ‘true’ branchit will have a umin_value (unsigned minimum value) of 9, whereas in the ‘false’branch it will have a umax_value of 8. A signed compare (with BPF_JSGT orBPF_JSGE) would instead update the signed minimum/maximum values. Informationfrom the signed and unsigned bounds can be combined; for instance if a value isfirst tested < 8 and then tested s> 4, the verifier will conclude that the valueis also > 4 and s< 8, since the bounds prevent crossing the sign boundary.

PTR_TO_PACKETs with a variable offset part have an ‘id’, which is common to allpointers sharing that same variable offset. This is important for packet rangechecks: after adding a variable to a packet pointer register A, if you then copyit to another register B and then add a constant 4 to A, both registers willshare the same ‘id’ but the A will have a fixed offset of +4. Then if A isbounds-checked and found to be less than a PTR_TO_PACKET_END, the register B isnow known to have a safe range of at least 4 bytes. See ‘Direct packet access’,below, for more on PTR_TO_PACKET ranges.

The ‘id’ field is also used on PTR_TO_MAP_VALUE_OR_NULL, common to all copies ofthe pointer returned from a map lookup. This means that when one copy ischecked and found to be non-NULL, all copies can become PTR_TO_MAP_VALUEs.As well as range-checking, the tracked information is also used for enforcingalignment of pointer accesses. For instance, on most systems the packet pointeris 2 bytes after a 4-byte alignment. If a program adds 14 bytes to that to jumpover the Ethernet header, then reads IHL and addes (IHL * 4), the resultingpointer will have a variable offset known to be 4n+2 for some n, so adding the 2bytes (NET_IP_ALIGN) gives a 4-byte alignment and so word-sized accesses throughthat pointer are safe.The ‘id’ field is also used on PTR_TO_SOCKET and PTR_TO_SOCKET_OR_NULL, commonto all copies of the pointer returned from a socket lookup. This has similarbehaviour to the handling for PTR_TO_MAP_VALUE_OR_NULL->PTR_TO_MAP_VALUE, butit also handles reference tracking for the pointer. PTR_TO_SOCKET implicitlyrepresents a reference to the correspondingstructsock. To ensure that thereference is not leaked, it is imperative to NULL-check the reference and inthe non-NULL case, and pass the valid reference to the socket release function.

Direct packet access

In cls_bpf and act_bpf programs the verifier allows direct access to the packetdata via skb->data and skb->data_end pointers.Ex:

1:  r4 = *(u32 *)(r1 +80)  /* load skb->data_end */2:  r3 = *(u32 *)(r1 +76)  /* load skb->data */3:  r5 = r34:  r5 += 145:  if r5 > r4 goto pc+16R1=ctx R3=pkt(id=0,off=0,r=14) R4=pkt_end R5=pkt(id=0,off=14,r=14) R10=fp6:  r0 = *(u16 *)(r3 +12) /* access 12 and 13 bytes of the packet */

this 2byte load from the packet is safe to do, since the program authordid checkif(skb->data+14>skb->data_end)gotoerr at insn #5 whichmeans that in the fall-through case the register R3 (which points to skb->data)has at least 14 directly accessible bytes. The verifier marks itas R3=pkt(id=0,off=0,r=14).id=0 means that no additional variables were added to the register.off=0 means that no additional constants were added.r=14 is the range of safe access which means that bytes [R3, R3 + 14) are ok.Note that R5 is marked as R5=pkt(id=0,off=14,r=14). It also pointsto the packet data, but constant 14 was added to the register, soit now points toskb->data+14 and accessible range is [R5, R5 + 14 - 14)which is zero bytes.

More complex packet access may look like:

R0=inv1 R1=ctx R3=pkt(id=0,off=0,r=14) R4=pkt_end R5=pkt(id=0,off=14,r=14) R10=fp6:  r0 = *(u8 *)(r3 +7) /* load 7th byte from the packet */7:  r4 = *(u8 *)(r3 +12)8:  r4 *= 149:  r3 = *(u32 *)(r1 +76) /* load skb->data */10:  r3 += r411:  r2 = r112:  r2 <<= 4813:  r2 >>= 4814:  r3 += r215:  r2 = r316:  r2 += 817:  r1 = *(u32 *)(r1 +80) /* load skb->data_end */18:  if r2 > r1 goto pc+2R0=inv(id=0,umax_value=255,var_off=(0x0; 0xff)) R1=pkt_end R2=pkt(id=2,off=8,r=8) R3=pkt(id=2,off=0,r=8) R4=inv(id=0,umax_value=3570,var_off=(0x0; 0xfffe)) R5=pkt(id=0,off=14,r=14) R10=fp19:  r1 = *(u8 *)(r3 +4)

The state of the register R3 is R3=pkt(id=2,off=0,r=8)id=2 means that twor3+=rX instructions were seen, so r3 points to someoffset within a packet and since the program author didif(r3+8>r1)gotoerr at insn #18, the safe range is [R3, R3 + 8).The verifier only allows ‘add’/’sub’ operations on packet registers. Any otheroperation will set the register state to ‘SCALAR_VALUE’ and it won’t beavailable for direct packet access.

Operationr3+=rX may overflow and become less than original skb->data,therefore the verifier has to prevent that. So when it seesr3+=rXinstruction and rX is more than 16-bit value, any subsequent bounds-check of r3against skb->data_end will not give us ‘range’ information, so attempts to readthrough the pointer will give “invalid access to packet” error.

Ex. after insnr4=*(u8*)(r3+12) (insn #7 above) the state of r4 isR4=inv(id=0,umax_value=255,var_off=(0x0; 0xff)) which means that upper 56 bitsof the register are guaranteed to be zero, and nothing is known about the lower8 bits. After insnr4*=14 the state becomesR4=inv(id=0,umax_value=3570,var_off=(0x0; 0xfffe)), since multiplying an 8-bitvalue by constant 14 will keep upper 52 bits as zero, also the least significantbit will be zero as 14 is even. Similarlyr2>>=48 will makeR2=inv(id=0,umax_value=65535,var_off=(0x0; 0xffff)), since the shift is not signextending. This logic is implemented in adjust_reg_min_max_vals() function,which calls adjust_ptr_min_max_vals() for adding pointer to scalar (or viceversa) and adjust_scalar_min_max_vals() for operations on two scalars.

The end result is that bpf program author can access packet directlyusing normal C code as:

void *data = (void *)(long)skb->data;void *data_end = (void *)(long)skb->data_end;struct eth_hdr *eth = data;struct iphdr *iph = data + sizeof(*eth);struct udphdr *udp = data + sizeof(*eth) + sizeof(*iph);if (data + sizeof(*eth) + sizeof(*iph) + sizeof(*udp) > data_end)        return 0;if (eth->h_proto != htons(ETH_P_IP))        return 0;if (iph->protocol != IPPROTO_UDP || iph->ihl != 5)        return 0;if (udp->dest == 53 || udp->source == 9)        ...;

which makes such programs easier to write comparing to LD_ABS insnand significantly faster.

eBPF maps

‘maps’ is a generic storage of different types for sharing data between kerneland userspace.

The maps are accessed from user space via BPF syscall, which has commands:

  • create a map with given type and attributesmap_fd=bpf(BPF_MAP_CREATE,unionbpf_attr*attr,u32size)using attr->map_type, attr->key_size, attr->value_size, attr->max_entriesreturns process-local file descriptor or negative error
  • lookup key in a given maperr=bpf(BPF_MAP_LOOKUP_ELEM,unionbpf_attr*attr,u32size)using attr->map_fd, attr->key, attr->valuereturns zero and stores found elem into value or negative error
  • create or update key/value pair in a given maperr=bpf(BPF_MAP_UPDATE_ELEM,unionbpf_attr*attr,u32size)using attr->map_fd, attr->key, attr->valuereturns zero or negative error
  • find and delete element by key in a given maperr=bpf(BPF_MAP_DELETE_ELEM,unionbpf_attr*attr,u32size)using attr->map_fd, attr->key
  • to delete map: close(fd)Exiting process will delete maps automatically

userspace programs use this syscall to create/access maps that eBPF programsare concurrently updating.

maps can have different types: hash, array, bloom filter, radix-tree, etc.

The map is defined by:

  • type
  • max number of elements
  • key size in bytes
  • value size in bytes

Pruning

The verifier does not actually walk all possible paths through the program. Foreach new branch to analyse, the verifier looks at all the states it’s previouslybeen in when at this instruction. If any of them contain the current state as asubset, the branch is ‘pruned’ - that is, the fact that the previous state wasaccepted implies the current state would be as well. For instance, if in theprevious state, r1 held a packet-pointer, and in the current state, r1 holds apacket-pointer with a range as long or longer and at least as strict analignment, then r1 is safe. Similarly, if r2 was NOT_INIT before then it can’thave been used by any path from that point, so any value in r2 (includinganother NOT_INIT) is safe. The implementation is in the function regsafe().Pruning considers not only the registers but also the stack (and any spilledregisters it may hold). They must all be safe for the branch to be pruned.This is implemented in states_equal().

Understanding eBPF verifier messages

The following are few examples of invalid eBPF programs and verifier errormessages as seen in the log:

Program with unreachable instructions:

static struct bpf_insn prog[] = {BPF_EXIT_INSN(),BPF_EXIT_INSN(),};

Error:

unreachable insn 1

Program that reads uninitialized register:

BPF_MOV64_REG(BPF_REG_0, BPF_REG_2),BPF_EXIT_INSN(),

Error:

0: (bf) r0 = r2R2 !read_ok

Program that doesn’t initialize R0 before exiting:

BPF_MOV64_REG(BPF_REG_2, BPF_REG_1),BPF_EXIT_INSN(),

Error:

0: (bf) r2 = r11: (95) exitR0 !read_ok

Program that accesses stack out of bounds:

BPF_ST_MEM(BPF_DW, BPF_REG_10, 8, 0),BPF_EXIT_INSN(),

Error:

0: (7a) *(u64 *)(r10 +8) = 0invalid stack off=8 size=8

Program that doesn’t initialize stack before passing its address into function:

BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -8),BPF_LD_MAP_FD(BPF_REG_1, 0),BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),BPF_EXIT_INSN(),

Error:

0: (bf) r2 = r101: (07) r2 += -82: (b7) r1 = 0x03: (85) call 1invalid indirect read from stack off -8+0 size 8

Program that uses invalid map_fd=0 while calling to map_lookup_elem() function:

BPF_ST_MEM(BPF_DW, BPF_REG_10, -8, 0),BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -8),BPF_LD_MAP_FD(BPF_REG_1, 0),BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),BPF_EXIT_INSN(),

Error:

0: (7a) *(u64 *)(r10 -8) = 01: (bf) r2 = r102: (07) r2 += -83: (b7) r1 = 0x04: (85) call 1fd 0 is not pointing to valid bpf_map

Program that doesn’t check return value of map_lookup_elem() before accessingmap element:

BPF_ST_MEM(BPF_DW, BPF_REG_10, -8, 0),BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -8),BPF_LD_MAP_FD(BPF_REG_1, 0),BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),BPF_ST_MEM(BPF_DW, BPF_REG_0, 0, 0),BPF_EXIT_INSN(),

Error:

0: (7a) *(u64 *)(r10 -8) = 01: (bf) r2 = r102: (07) r2 += -83: (b7) r1 = 0x04: (85) call 15: (7a) *(u64 *)(r0 +0) = 0R0 invalid mem access 'map_value_or_null'

Program that correctly checks map_lookup_elem() returned value for NULL, butaccesses the memory with incorrect alignment:

BPF_ST_MEM(BPF_DW, BPF_REG_10, -8, 0),BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -8),BPF_LD_MAP_FD(BPF_REG_1, 0),BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 1),BPF_ST_MEM(BPF_DW, BPF_REG_0, 4, 0),BPF_EXIT_INSN(),

Error:

0: (7a) *(u64 *)(r10 -8) = 01: (bf) r2 = r102: (07) r2 += -83: (b7) r1 = 14: (85) call 15: (15) if r0 == 0x0 goto pc+1 R0=map_ptr R10=fp6: (7a) *(u64 *)(r0 +4) = 0misaligned access off 4 size 8

Program that correctly checks map_lookup_elem() returned value for NULL andaccesses memory with correct alignment in one side of ‘if’ branch, but failsto do so in the other side of ‘if’ branch:

BPF_ST_MEM(BPF_DW, BPF_REG_10, -8, 0),BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -8),BPF_LD_MAP_FD(BPF_REG_1, 0),BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 2),BPF_ST_MEM(BPF_DW, BPF_REG_0, 0, 0),BPF_EXIT_INSN(),BPF_ST_MEM(BPF_DW, BPF_REG_0, 0, 1),BPF_EXIT_INSN(),

Error:

0: (7a) *(u64 *)(r10 -8) = 01: (bf) r2 = r102: (07) r2 += -83: (b7) r1 = 14: (85) call 15: (15) if r0 == 0x0 goto pc+2 R0=map_ptr R10=fp6: (7a) *(u64 *)(r0 +0) = 07: (95) exitfrom 5 to 8: R0=imm0 R10=fp8: (7a) *(u64 *)(r0 +0) = 1R0 invalid mem access 'imm'

Program that performs a socket lookup then sets the pointer to NULL withoutchecking it:

BPF_MOV64_IMM(BPF_REG_2, 0),BPF_STX_MEM(BPF_W, BPF_REG_10, BPF_REG_2, -8),BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -8),BPF_MOV64_IMM(BPF_REG_3, 4),BPF_MOV64_IMM(BPF_REG_4, 0),BPF_MOV64_IMM(BPF_REG_5, 0),BPF_EMIT_CALL(BPF_FUNC_sk_lookup_tcp),BPF_MOV64_IMM(BPF_REG_0, 0),BPF_EXIT_INSN(),

Error:

0: (b7) r2 = 01: (63) *(u32 *)(r10 -8) = r22: (bf) r2 = r103: (07) r2 += -84: (b7) r3 = 45: (b7) r4 = 06: (b7) r5 = 07: (85) call bpf_sk_lookup_tcp#658: (b7) r0 = 09: (95) exitUnreleased reference id=1, alloc_insn=7

Program that performs a socket lookup but does not NULL-check the returnedvalue:

BPF_MOV64_IMM(BPF_REG_2, 0),BPF_STX_MEM(BPF_W, BPF_REG_10, BPF_REG_2, -8),BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -8),BPF_MOV64_IMM(BPF_REG_3, 4),BPF_MOV64_IMM(BPF_REG_4, 0),BPF_MOV64_IMM(BPF_REG_5, 0),BPF_EMIT_CALL(BPF_FUNC_sk_lookup_tcp),BPF_EXIT_INSN(),

Error:

0: (b7) r2 = 01: (63) *(u32 *)(r10 -8) = r22: (bf) r2 = r103: (07) r2 += -84: (b7) r3 = 45: (b7) r4 = 06: (b7) r5 = 07: (85) call bpf_sk_lookup_tcp#658: (95) exitUnreleased reference id=1, alloc_insn=7

Testing

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

CONFIG_TEST_BPF=m

After the module has been built and installed, the test suite can be executedvia insmod or modprobe against ‘test_bpf’ module. Results of the test casesincluding timings in nsec can be found in the kernel log (dmesg).

Misc

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

Written by

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