Skip to content

Latest commit

 

History

History
285 lines (223 loc) · 11.3 KB

README.md

File metadata and controls

285 lines (223 loc) · 11.3 KB
Raw

BPFtrace

BPFtrace is a high-level tracing language for Linux enhanced Berkeley Packet Filter (eBPF) available in recent Linux kernels (4.x). BPFtrace uses LLVM as a backend to compile scripts to BPF-bytecode and makes use of BCC for interacting with the Linux BPF system, as well as existing Linux tracing capabilities: kernel dynamic tracing (kprobes), user-level dynamic tracing (uprobes), and tracepoints. The BPFtrace language is inspired by awk and C, and predecessor tracers such as DTrace and SystemTap. BPFtrace was created by Alastair Robertson.

To learn more about BPFtrace, see the Reference Guide and One-Liner Tutorial.

Install

For build and install instructions, see INSTALL.md.

Examples

Count system calls using tracepoints:

# bpftrace -e 'tracepoint:syscalls:sys_enter_* { @[name] = count(); }'
Attaching 320 probes...
^C

...
@[tracepoint:syscalls:sys_enter_access]: 3291
@[tracepoint:syscalls:sys_enter_close]: 3897
@[tracepoint:syscalls:sys_enter_newstat]: 4268
@[tracepoint:syscalls:sys_enter_open]: 4609
@[tracepoint:syscalls:sys_enter_mmap]: 4781

Produce a histogram of time (in nanoseconds) spent in the read() system call:

// read.bt file
tracepoint:syscalls:sys_enter_read
{
  @start[tid] = nsecs;
}

tracepoint:syscalls:sys_exit_read / @start[tid] /
{
  @times = hist(nsecs - @start[tid]);
  delete(@start[tid]);
}

# bpftrace read.bt
Attaching 2 probes...
^C

@times:
[256, 512)           326 |@                                                   |
[512, 1k)           7715 |@@@@@@@@@@@@@@@@@@@@@@@@@@                          |
[1k, 2k)           15306 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[2k, 4k)             609 |@@                                                  |
[4k, 8k)             611 |@@                                                  |
[8k, 16k)            438 |@                                                   |
[16k, 32k)            59 |                                                    |
[32k, 64k)            36 |                                                    |
[64k, 128k)            5 |                                                    |

Print process name and paths for file opens, using kprobes (kernel dynamic tracing) of do_sys_open():

# bpftrace -e 'kprobe:do_sys_open { printf("%s: %s\n", comm, str(arg1)) }'
Attaching 1 probe...
git: .git/objects/da
git: .git/objects/pack
git: /etc/localtime
systemd-journal: /var/log/journal/72d0774c88dc4943ae3d34ac356125dd
DNS Res~ver #15: /etc/hosts
^C

CPU profiling, sampling kernel stacks at 99 Hertz:

# bpftrace -e 'profile:hz:99 { @[stack] = count() }'
Attaching 1 probe...
^C

...
@[
queue_work_on+41
tty_flip_buffer_push+43
pty_write+83
n_tty_write+434
tty_write+444
__vfs_write+55
vfs_write+177
sys_write+85
entry_SYSCALL_64_fastpath+26
]: 97
@[
cpuidle_enter_state+299
cpuidle_enter+23
call_cpuidle+35
do_idle+394
cpu_startup_entry+113
rest_init+132
start_kernel+1083
x86_64_start_reservations+41
x86_64_start_kernel+323
verify_cpu+0
]: 150

One-Liners

The following one-liners demonstrate different capabilities:

# Files opened by process
bpftrace -e 'tracepoint:syscalls:sys_enter_open { printf("%s %s\n", comm, str(args->filename)); }'

# Syscall count by program
bpftrace -e 'tracepoint:raw_syscalls:sys_enter { @[comm] = count(); }'

# Read bytes by process:
bpftrace -e 'tracepoint:syscalls:sys_exit_read /args->ret/ { @[comm] = sum(args->ret); }'

# Read size distribution by process:
bpftrace -e 'tracepoint:syscalls:sys_exit_read { @[comm] = hist(args->ret); }'

# Show per-second syscall rates:
bpftrace -e 'tracepoint:raw_syscalls:sys_enter { @ = count(); } interval:s:1 { print(@); clear(@); }'

# Trace disk size by process
bpftrace -e 'tracepoint:block:block_rq_issue { printf("%d %s %d\n", pid, comm, args->bytes); }'

# Count page faults by process
bpftrace -e 'software:faults:1 { @[comm] = count(); }'

# Count LLC cache misses by process name and PID (uses PMCs):
bpftrace -e 'hardware:cache-misses:1000000 { @[comm, pid] = count(); }'

# Profile user-level stacks at 99 Hertz, for PID 189:
bpftrace -e 'profile:hz:99 /pid == 189/ { @[ustack] = count(); }'

# Files opened, for processes in the root cgroup-v2
bpftrace -e 'tracepoint:syscalls:sys_enter_openat /cgroup == cgroupid("/sys/fs/cgroup/unified/mycg")/ { printf("%s\n", str(args->filename)); }'

Tools

bpftrace contains various tools, which also serve as examples of programming in the bpftrace language.

For more eBPF observability tools, see bcc tools.

Probe types

kprobes

Attach a BPFtrace script to a kernel function, to be executed when that function is called:

kprobe:vfs_read { ... }

uprobes

Attach script to a userland function:

uprobe:/bin/bash:readline { ... }

tracepoints

Attach script to a statically defined tracepoint in the kernel:

tracepoint:sched:sched_switch { ... }

Tracepoints are guaranteed to be stable between kernel versions, unlike kprobes.

software

Attach script to kernel software events, executing once every provided count or use a default:

software:faults:100 software:faults:

hardware

Attach script to hardware events (PMCs), executing once every provided count or use a default:

hardware:cache-references:1000000 hardware:cache-references:

profile

Run the script on all CPUs at specified time intervals:

profile:hz:99 { ... }

profile:s:1 { ... }

profile:ms:20 { ... }

profile:us:1500 { ... }

interval

Run the script once per interval, for printing interval output:

interval:s:1 { ... }

interval:ms:20 { ... }

Multiple attachment points

A single probe can be attached to multiple events:

kprobe:vfs_read,kprobe:vfs_write { ... }

Wildcards

Some probe types allow wildcards to be used when attaching a probe:

kprobe:vfs_* { ... }

Predicates

Define conditions for which a probe should be executed:

kprobe:sys_open / uid == 0 / { ... }

Builtins

The following variables and functions are available for use in bpftrace scripts:

Variables:

  • pid - Process ID (kernel tgid)
  • tid - Thread ID (kernel pid)
  • cgroup - Cgroup ID of the current process
  • uid - User ID
  • gid - Group ID
  • nsecs - Nanosecond timestamp
  • cpu - Processor ID
  • comm - Process name
  • stack - Kernel stack trace
  • ustack - User stack trace
  • arg0, arg1, ... etc. - Arguments to the function being traced
  • retval - Return value from function being traced
  • func - Name of the function currently being traced
  • name - Full name of the probe
  • curtask - Current task_struct as a u64.
  • rand - Random number of type u32.

Functions:

  • hist(int n) - Produce a log2 histogram of values of n
  • lhist(int n, int min, int max, int step) - Produce a linear histogram of values of n
  • count() - Count the number of times this function is called
  • sum(int n) - Sum this value
  • min(int n) - Record the minimum value seen
  • max(int n) - Record the maximum value seen
  • avg(int n) - Average this value
  • stats(int n) - Return the count, average, and total for this value
  • delete(@x) - Delete the map element passed in as an argument
  • str(char *s) - Returns the string pointed to by s
  • printf(char *fmt, ...) - Print formatted to stdout
  • print(@x[, int top [, int div]]) - Print a map, with optional top entry count and divisor
  • clear(@x) - Delete all key/values from a map
  • sym(void *p) - Resolve kernel address
  • usym(void *p) - Resolve user space address
  • kaddr(char *name) - Resolve kernel symbol name
  • uaddr(char *name) - Resolve user space symbol name
  • reg(char *name) - Returns the value stored in the named register
  • join(char *arr[]) - Prints the string array
  • time(char *fmt) - Print the current time
  • system(char *fmt) - Execute shell command
  • exit() - Quit bpftrace

See the Reference Guide for more detail.

Internals

bpftrace employs various techniques for efficiency, minimizing the instrumentation overhead. Summary statistics are stored in kernel BPF maps, which are asynchronously copied from kernel to user-space, only when needed. Other data, and asynchronous actions, are passed from kernel to user-space via the perf output buffer.