/libbpf-bootstrap

Scaffolding for BPF application development with libbpf and BPF CO-RE

Primary LanguageCBSD 3-Clause "New" or "Revised" LicenseBSD-3-Clause

Demo BPF applications

Minimal

minimal is just that – a minimal practical BPF application example. It doesn't use or require BPF CO-RE, so should run on quite old kernels. It installs a tracepoint handler which is triggered once every second. It uses bpf_printk() BPF helper to communicate with the world. To see it's output, read /sys/kernel/debug/tracing/trace_pipe file as a root:

$ cd examples/c
$ make minimal
$ sudo ./minimal
$ sudo cat /sys/kernel/debug/tracing/trace_pipe
           <...>-3840345 [010] d... 3220701.101143: bpf_trace_printk: BPF triggered from PID 3840345.
           <...>-3840345 [010] d... 3220702.101265: bpf_trace_printk: BPF triggered from PID 3840345.

minimal is great as a bare-bones experimental playground to quickly try out new ideas or BPF features.

Minimal_Legacy

This version of minimal is modified to allow running on even older kernels that do not allow global variables. bpf_printk uses global variables unless BPF_NO_GLOBAL_DATA is defined before including bpf_helpers.h. Additionally, the global variable my_pid has been replaced with an array of one element to hold the process pid.

$ cd examples/c
$ make minimal_legacy
$ sudo ./minimal_legacy
$ sudo cat /sys/kernel/debug/tracing/trace_pipe
  minimal_legacy-52030 [001] .... 491227.784078: 0x00000001: BPF triggered from PID 52030.
  minimal_legacy-52030 [001] .... 491228.840571: 0x00000001: BPF triggered from PID 52030.
  minimal_legacy-52030 [001] .... 491229.841643: 0x00000001: BPF triggered from PID 52030.
  minimal_legacy-52030 [001] .... 491230.842432: 0x00000001: BPF triggered from PID 52030.

Bootstrap

bootstrap is an example of a simple (but realistic) BPF application. It tracks process starts (exec() family of syscalls, to be precise) and exits and emits data about filename, PID and parent PID, as well as exit status and duration of the process life. With -d <min-duration-ms> you can specify minimum duration of the process to log. In such mode process start (technically, exec()) events are not output (see example output below).

bootstrap was created in the similar spirit as libbpf-tools from BCC package, but is designed to be more stand-alone and with simpler Makefile to simplify adoption to user's particular needs. It demonstrates the use of typical BPF features:

  • cooperating BPF programs (tracepoint handlers for process exec and exit events, in this particular case);
  • BPF map for maintaining the state;
  • BPF ring buffer for sending data to user-space;
  • global variables for application behavior parameterization.
  • it utilizes BPF CO-RE and vmlinux.h to read extra process information from kernel's struct task_struct.

bootstrap is intended to be the starting point for your own BPF application, with things like BPF CO-RE and vmlinux.h, consuming BPF ring buffer data, command line arguments parsing, graceful Ctrl-C handling, etc. all taken care of for you, which are crucial but mundane tasks that are no fun, but necessary to be able to do anything useful. Just copy/paste and do simple renaming to get yourself started.

Here's an example output in minimum process duration mode:

$ sudo ./bootstrap -d 50
TIME     EVENT COMM             PID     PPID    FILENAME/EXIT CODE
19:18:32 EXIT  timeout          3817109 402466  [0] (126ms)
19:18:32 EXIT  sudo             3817117 3817111 [0] (259ms)
19:18:32 EXIT  timeout          3817110 402466  [0] (264ms)
19:18:33 EXIT  python3.7        3817083 1       [0] (1026ms)
19:18:38 EXIT  python3          3817429 3817424 [1] (60ms)
19:18:38 EXIT  sh               3817424 3817420 [0] (79ms)
19:18:38 EXIT  timeout          3817420 402466  [0] (80ms)
19:18:43 EXIT  timeout          3817610 402466  [0] (70ms)
19:18:43 EXIT  grep             3817619 3817617 [1] (271ms)
19:18:43 EXIT  timeout          3817609 402466  [0] (321ms)
19:18:44 EXIT  iostat           3817585 3817531 [0] (3006ms)
19:18:44 EXIT  tee              3817587 3817531 [0] (3005ms)
...

Uprobe

uprobe is an example of dealing with user-space entry and exit (return) probes, uprobe and uretprobe in libbpf lingo. It attached uprobe and uretprobe BPF programs to its own functions (uprobed_add() and uprobed_sub()) and logs input arguments and return result, respectively, using bpf_printk() macro. The user-space function is triggered once every second:

$ sudo ./uprobe
libbpf: loading object 'uprobe_bpf' from buffer
...
Successfully started!
...........

You can see uprobe demo output in /sys/kernel/debug/tracing/trace_pipe:

$ sudo cat /sys/kernel/debug/tracing/trace_pipe
          uprobe-1809291 [007] .... 4017233.106596: 0: uprobed_add ENTRY: a = 0, b = 1
          uprobe-1809291 [007] .... 4017233.106605: 0: uprobed_add EXIT: return = 1
          uprobe-1809291 [007] .... 4017233.106606: 0: uprobed_sub ENTRY: a = 0, b = 0
          uprobe-1809291 [007] .... 4017233.106607: 0: uprobed_sub EXIT: return = 0
          uprobe-1809291 [007] .... 4017234.106694: 0: uprobed_add ENTRY: a = 1, b = 2
          uprobe-1809291 [007] .... 4017234.106697: 0: uprobed_add EXIT: return = 3
          uprobe-1809291 [007] .... 4017234.106700: 0: uprobed_sub ENTRY: a = 1, b = 1
          uprobe-1809291 [007] .... 4017234.106701: 0: uprobed_sub EXIT: return = 0

USDT

usdt is an example of dealing with USDT probe. It attaches USDT BPF programs to the libc:setjmp probe, which is triggered by calling setjmp in user-space program once per second and logs USDT arguments using bpf_printk() macro:

$ sudo ./usdt
libbpf: loading object 'usdt_bpf' from buffer
...
Successfully started!
...........

You can see usdt demo output in /sys/kernel/debug/tracing/trace_pipe:

$ sudo cat /sys/kernel/debug/tracing/trace_pipe
            usdt-1919077 [005] d..21 537310.886092: bpf_trace_printk: USDT auto attach to libc:setjmp: arg1 = 55d03d6a42a0, arg2 = 0, arg3 = 55d03d65e54e
            usdt-1919077 [005] d..21 537310.886105: bpf_trace_printk: USDT manual attach to libc:setjmp: arg1 = 55d03d6a42a0, arg2 = 0, arg3 = 55d03d65e54e
            usdt-1919077 [005] d..21 537311.886214: bpf_trace_printk: USDT auto attach to libc:setjmp: arg1 = 55d03d6a42a0, arg2 = 0, arg3 = 55d03d65e54e
            usdt-1919077 [005] d..21 537311.886227: bpf_trace_printk: USDT manual attach to libc:setjmp: arg1 = 55d03d6a42a0, arg2 = 0, arg3 = 55d03d65e54e

Fentry

fentry is an example that uses fentry and fexit BPF programs for tracing. It attaches fentry and fexit traces to do_unlinkat() which is called when a file is deleted and logs the return value, PID, and filename to the trace pipe.

Important differences, compared to kprobes, are improved performance and usability. In this example, better usability is shown with the ability to directly dereference pointer arguments, like in normal C, instead of using various read helpers. The big distinction between fexit and kretprobe programs is that fexit one has access to both input arguments and returned result, while kretprobe can only access the result.

fentry and fexit programs are available starting from 5.5 kernels.

$ sudo ./fentry
libbpf: loading object 'fentry_bpf' from buffer
...
Successfully started!
..........

The fentry output in /sys/kernel/debug/tracing/trace_pipe should look something like this:

$ sudo cat /sys/kernel/debug/tracing/trace_pipe
              rm-9290    [004] d..2  4637.798698: bpf_trace_printk: fentry: pid = 9290, filename = test_file
              rm-9290    [004] d..2  4637.798843: bpf_trace_printk: fexit: pid = 9290, filename = test_file, ret = 0
              rm-9290    [004] d..2  4637.798698: bpf_trace_printk: fentry: pid = 9290, filename = test_file2
              rm-9290    [004] d..2  4637.798843: bpf_trace_printk: fexit: pid = 9290, filename = test_file2, ret = 0

Kprobe

kprobe is an example of dealing with kernel-space entry and exit (return) probes, kprobe and kretprobe in libbpf lingo. It attaches kprobe and kretprobe BPF programs to the do_unlinkat() function and logs the PID, filename, and return result, respectively, using bpf_printk() macro.

$ sudo ./kprobe
libbpf: loading object 'kprobe_bpf' from buffer
...
Successfully started!
...........

The kprobe demo output in /sys/kernel/debug/tracing/trace_pipe should look something like this:

$ sudo cat /sys/kernel/debug/tracing/trace_pipe
              rm-9346    [005] d..3  4710.951696: bpf_trace_printk: KPROBE ENTRY pid = 9346, filename = test1
              rm-9346    [005] d..4  4710.951819: bpf_trace_printk: KPROBE EXIT: ret = 0
              rm-9346    [005] d..3  4710.951852: bpf_trace_printk: KPROBE ENTRY pid = 9346, filename = test2
              rm-9346    [005] d..4  4710.951895: bpf_trace_printk: KPROBE EXIT: ret = 0

XDP

xdp is an example written in Rust (using libbpf-rs). It attaches to the ingress path of networking device and logs the size of each packet, returning XDP_PASS to allow the packet to be passed up to the kernel’s networking stack.

$ sudo ./target/release/xdp 1
..........

The xdp output in /sys/kernel/debug/tracing/trace_pipe should look something like this:

$ sudo cat /sys/kernel/debug/tracing/trace_pipe
           <...>-823887  [000] d.s1 602386.079100: bpf_trace_printk: packet size: 75
           <...>-823887  [000] d.s1 602386.079141: bpf_trace_printk: packet size: 66
           <...>-2813507 [000] d.s1 602386.696702: bpf_trace_printk: packet size: 77
           <...>-2813507 [000] d.s1 602386.696735: bpf_trace_printk: packet size: 66

TC

tc (short for Traffic Control) is an example of handling ingress network traffics. It creates a qdisc on the lo interface and attaches the tc_ingress BPF program to it. It reports the metadata of the IP packets that coming into the lo interface.

$ sudo ./tc
...
Successfully started! Please run `sudo cat /sys/kernel/debug/tracing/trace_pipe` to see output of the BPF program.
......

The tc output in /sys/kernel/debug/tracing/trace_pipe should look something like this:

$ sudo cat /sys/kernel/debug/tracing/trace_pipe
            node-1254811 [007] ..s1 8737831.671074: 0: Got IP packet: tot_len: 79, ttl: 64
            sshd-1254728 [006] ..s1 8737831.674334: 0: Got IP packet: tot_len: 79, ttl: 64
            sshd-1254728 [006] ..s1 8737831.674349: 0: Got IP packet: tot_len: 72, ttl: 64
            node-1254811 [007] ..s1 8737831.674550: 0: Got IP packet: tot_len: 71, ttl: 64

Profile

profile is an example written in Rust and C with BlazeSym. It attaches to perf events, sampling on every processor periodically. It shows addresses, symbols, file names, and line numbers of stacktraces.

$ sudo ./target/release/profile
COMM: swapper/6 (pid=0) @ CPU 6
Kernel:
  0 [<ffffffff81bdf010>] intel_idle+0x96
  1 [<ffffffff819959b0>] cpuidle_enter_state+0x80 /ro/source/drivers/cpuidle/cpuidle.c:238
  2 [<ffffffff81995cc9>] cpuidle_enter+0x29 /ro/source/drivers/cpuidle/cpuidle.c:353
  3 [<ffffffff810f8c0b>] do_idle+0x1bb /ro/source/kernel/sched/idle.c:243
  4 [<ffffffff810f8de9>] cpu_startup_entry+0x19 /ro/source/kernel/sched/idle.c:396
  5 [<ffffffff81044f46>] start_secondary+0x116 /ro/source/arch/x86/kernel/smpboot.c:272
  6 [<ffffffff810000f5>] secondary_startup_64_no_verify+0xb0 /ro/source/arch/x86/kernel/head_64.S:283
No Userspace Stack

C version and Rust version show the same content. Both of them use BlazeSym to symbolize stacktraces.

Socket filter

sockfilter is an example of monitoring packet and dealing with __sk_buff structure. It attaches socket BPF program to sock_queue_rcv_skb() function and retrieve information from BPF_MAP_TYPE_RINGBUF, then print protocol, src IP, src port, dst IP, dst port in standard output. Currently, most of the IPv4 protocols defined in uapi/linux/in.h are included, please check ipproto_mapping of examples/c/sockfilter.c for the supported protocols.

$ sudo ./sockfilter
interface:lo    protocol: UDP   127.0.0.1:51845(src) -> 127.0.0.1:53(dst)
interface:lo    protocol: UDP   127.0.0.1:41552(src) -> 127.0.0.1:53(dst)

Building

libbpf-bootstrap supports multiple build systems that do the same thing. This serves as a cross reference for folks coming from different backgrounds.

Install Dependencies

You will need clang, libelf and zlib to build the examples, package names may vary across distros.

On Ubuntu/Debian, you need:

$ apt install clang libelf1 libelf-dev zlib1g-dev

On CentOS/Fedora, you need:

$ dnf install clang elfutils-libelf elfutils-libelf-devel zlib-devel

C Examples

Makefile build:

$ git submodule update --init --recursive       # check out libbpf
$ cd examples/c
$ make
$ sudo ./bootstrap
TIME     EVENT COMM             PID     PPID    FILENAME/EXIT CODE
00:21:22 EXIT  python3.8        4032353 4032352 [0] (123ms)
00:21:22 EXEC  mkdir            4032379 4032337 /usr/bin/mkdir
00:21:22 EXIT  mkdir            4032379 4032337 [0] (1ms)
00:21:22 EXEC  basename         4032382 4032381 /usr/bin/basename
00:21:22 EXIT  basename         4032382 4032381 [0] (0ms)
00:21:22 EXEC  sh               4032381 4032380 /bin/sh
00:21:22 EXEC  dirname          4032384 4032381 /usr/bin/dirname
00:21:22 EXIT  dirname          4032384 4032381 [0] (1ms)
00:21:22 EXEC  readlink         4032387 4032386 /usr/bin/readlink
^C

CMake build:

$ git submodule update --init --recursive       # check out libbpf
$ mkdir build && cd build
$ cmake ../examples/c
$ make
$ sudo ./bootstrap
<...>

XMake build (Linux):

$ git submodule update --init --recursive       # check out libbpf
$ cd examples/c
$ xmake
$ xmake run bootstrap

XMake build (Android):

$ git submodule update --init --recursive       # check out libbpf
$ cd examples/c
$ xmake f -p android
$ xmake

Install Xmake

$ bash <(wget https://xmake.io/shget.text -O -)
$ source ~/.xmake/profile

Rust Examples

Install libbpf-cargo:

$ cargo install libbpf-cargo

Build using cargo:

$ cd examples/rust
$ cargo build --release
$ sudo ./target/release/xdp 1
<...>

Troubleshooting

Libbpf debug logs are quire helpful to pinpoint the exact source of problems, so it's usually a good idea to look at them before starting to debug or posting question online.

./minimal is always running with libbpf debug logs turned on.

For ./bootstrap, run it in verbose mode (-v) to see libbpf debug logs:

$ sudo ./bootstrap -v
libbpf: loading object 'bootstrap_bpf' from buffer
libbpf: elf: section(2) tp/sched/sched_process_exec, size 384, link 0, flags 6, type=1
libbpf: sec 'tp/sched/sched_process_exec': found program 'handle_exec' at insn offset 0 (0 bytes), code size 48 insns (384 bytes)
libbpf: elf: section(3) tp/sched/sched_process_exit, size 432, link 0, flags 6, type=1
libbpf: sec 'tp/sched/sched_process_exit': found program 'handle_exit' at insn offset 0 (0 bytes), code size 54 insns (432 bytes)
libbpf: elf: section(4) license, size 13, link 0, flags 3, type=1
libbpf: license of bootstrap_bpf is Dual BSD/GPL
...