README.md 15.9 KB
Newer Older
1
![BCC Logo](images/logo2.png)
Brenden's avatar
Brenden committed
2 3
# BPF Compiler Collection (BCC)

Brendan Gregg's avatar
Brendan Gregg committed
4 5 6 7 8 9 10 11 12
BCC is a toolkit for creating efficient kernel tracing and manipulation
programs, and includes several useful tools and examples. It makes use of eBPF
(Extended Berkeley Packet Filters), a new feature that was first added to
Linux 3.15. Much of what BCC uses requires Linux 4.1 and above.

eBPF was [described by](https://lkml.org/lkml/2015/4/14/232) Ingo Molnár as:

> One of the more interesting features in this cycle is the ability to attach eBPF programs (user-defined, sandboxed bytecode executed by the kernel) to kprobes. This allows user-defined instrumentation on a live kernel image that can never crash, hang or interfere with the kernel negatively.

Brendan Gregg's avatar
Brendan Gregg committed
13 14 15
BCC makes eBPF programs easier to write, with kernel instrumentation in C
and a front-end in Python. It is suited for many tasks, including performance
analysis and network traffic control.
Brendan Gregg's avatar
Brendan Gregg committed
16 17 18 19 20 21 22 23 24 25 26

## Screenshot

This example traces a disk I/O kernel function, and populates an in-kernel
power-of-2 histogram of the I/O size. For efficiency, only the histogram
summary is returned to user-level.

```Shell
# ./bitehist.py 
Tracing... Hit Ctrl-C to end.
^C
Brendan Gregg's avatar
Brendan Gregg committed
27
     kbytes          : count     distribution
Brendan Gregg's avatar
Brendan Gregg committed
28 29 30 31 32 33 34 35 36 37 38 39
       0 -> 1        : 3        |                                      |
       2 -> 3        : 0        |                                      |
       4 -> 7        : 211      |**********                            |
       8 -> 15       : 0        |                                      |
      16 -> 31       : 0        |                                      |
      32 -> 63       : 0        |                                      |
      64 -> 127      : 1        |                                      |
     128 -> 255      : 800      |**************************************|
```

The above output shows a bimodal distribution, where the largest mode of
800 I/O was between 128 and 255 Kbytes in size.
Brenden's avatar
Brenden committed
40

41 42
See the source: [bitehist.c](examples/tracing/bitehist.c) and
[bitehist.py](examples/tracing/bitehist.py). What this traces, what this stores, and how
Brendan Gregg's avatar
Brendan Gregg committed
43 44
the data is presented, can be entirely customized. This shows only some of
many possible capabilities.
Brenden's avatar
Brenden committed
45

46 47 48 49
## Installing

See [INSTALL.md](INSTALL.md) for installation steps on your platform.

Brendan Gregg's avatar
Brendan Gregg committed
50 51 52 53 54 55 56 57 58
## Contents

Some of these are single files that contain both C and Python, others have a
pair of .c and .py files, and some are directories of files.

### Tracing

Examples:

59 60
- examples/tracing/[bitehist.py](examples/tracing/bitehist.py) examples/tracing/[bitehist.c](examples/tracing/bitehist.c): Block I/O size histogram. [Examples](examples/tracing/bitehist_example.txt).
- examples/tracing/[disksnoop.py](examples/tracing/disksnoop.py) examples/tracing/[disksnoop.c](examples/tracing/disksnoop.c): Trace block device I/O latency. [Examples](examples/tracing/disksnoop_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
61
- examples/[hello_world.py](examples/hello_world.py): Prints "Hello, World!" for new processes.
62 63 64
- examples/tracing/[tcpv4connect](examples/tracing/tcpv4connect): Trace TCP IPv4 active connections. [Examples](examples/tracing/tcpv4connect_example.txt).
- examples/tracing/[trace_fields.py](examples/tracing/trace_fields.py): Simple example of printing fields from traced events.
- examples/tracing/[vfsreadlat.py](examples/tracing/vfsreadlat.py) examples/tracing/[vfsreadlat.c](examples/tracing/vfsreadlat.c): VFS read latency distribution. [Examples](examples/tracing/vfsreadlat_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
65 66 67

Tools:

68
- tools/[argdist](tools/argdist.py): Display function parameter values as a histogram or frequency count. [Examples](tools/argdist_examples.txt).
Brendan Gregg's avatar
Brendan Gregg committed
69
- tools/[bashreadline](tools/bashreadline.py): Print entered bash commands system wide. [Examples](tools/bashreadline_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
70
- tools/[biolatency](tools/biolatency.py): Summarize block device I/O latency as a histogram. [Examples](tools/biolatency_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
71
- tools/[biotop](tools/biotop.py): Top for disks: Summarize block device I/O by process. [Examples](tools/biotop_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
72
- tools/[biosnoop](tools/biosnoop.py): Trace block device I/O with PID and latency. [Examples](tools/biosnoop_example.txt).
Allan McAleavy's avatar
Allan McAleavy committed
73
- tools/[bitesize](tools/bitesize.py): Show per process I/O size histogram. [Examples](tools/bitesize_example.txt).
74
- tools/[cachestat](tools/cachestat.py): Trace page cache hit/miss ratio. [Examples](tools/cachestat_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
75
- tools/[execsnoop](tools/execsnoop.py): Trace new processes via exec() syscalls. [Examples](tools/execsnoop_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
76
- tools/[dcsnoop](tools/dcsnoop.py): Trace directory entry cache (dcache) lookups. [Examples](tools/dcsnoop_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
77
- tools/[dcstat](tools/dcstat.py): Directory entry cache (dcache) stats. [Examples](tools/dcstat_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
78
- tools/[ext4dist](tools/ext4dist.py): Summarize ext4 operation latency. [Examples](tools/ext4dist_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
79
- tools/[ext4slower](tools/ext4slower.py): Trace slow ext4 operations. [Examples](tools/ext4slower_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
80
- tools/[filelife](tools/filelife.py): Trace the lifespan of short-lived files. [Examples](tools/filelife_example.txt).
81
- tools/[fileslower](tools/fileslower.py): Trace slow synchronous file reads and writes. [Examples](tools/fileslower_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
82
- tools/[filetop](tools/filetop.py): File reads and writes by filename and process. Top for files. [Examples](tools/filetop_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
83 84
- tools/[funccount](tools/funccount.py): Count kernel function calls. [Examples](tools/funccount_example.txt).
- tools/[funclatency](tools/funclatency.py): Time kernel functions and show their latency distribution. [Examples](tools/funclatency_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
85
- tools/[gethostlatency](tools/gethostlatency.py): Show latency for getaddrinfo/gethostbyname[2] calls. [Examples](tools/gethostlatency_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
86 87
- tools/[hardirqs](tools/hardirqs.py):  Measure hard IRQ (hard interrupt) event time. [Examples](tools/hardirqs_example.txt).
- tools/[killsnoop](tools/killsnoop.py): Trace signals issued by the kill() syscall. [Examples](tools/killsnoop_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
88
- tools/[mdflush](tools/mdflush.py): Trace md flush events. [Examples](tools/mdflush.txt).
89
- tools/[memleak](tools/memleak.py): Display outstanding memory allocations to find memory leaks. [Examples](tools/memleak_examples.txt).
Brendan Gregg's avatar
Brendan Gregg committed
90
- tools/[offcputime](tools/offcputime.py): Summarize off-CPU time by kernel stack trace. [Examples](tools/offcputime_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
91
- tools/[offwaketime](tools/offwaketime.py): Summarize blocked time by kernel off-CPU stack and waker stack. [Examples](tools/offwaketime_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
92
- tools/[oomkill](tools/oomkill.py): Trace the out-of-memory (OOM) killer. [Examples](tools/oomkill_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
93 94
- tools/[opensnoop](tools/opensnoop.py): Trace open() syscalls. [Examples](tools/opensnoop_example.txt).
- tools/[pidpersec](tools/pidpersec.py): Count new processes (via fork). [Examples](tools/pidpersec_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
95
- tools/[runqlat](tools/runqlat.py): Run queue (scheduler) latency as a histogram. [Examples](tools/runqlat_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
96 97 98
- tools/[softirqs](tools/softirqs.py):  Measure soft IRQ (soft interrupt) event time. [Examples](tools/softirqs_example.txt).
- tools/[stackcount](tools/stackcount.py): Count kernel function calls and their stack traces. [Examples](tools/stackcount_example.txt).
- tools/[stacksnoop](tools/stacksnoop.py): Trace a kernel function and print all kernel stack traces. [Examples](tools/stacksnoop_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
99
- tools/[statsnoop](tools/statsnoop.py): Trace stat() syscalls. [Examples](tools/statsnoop_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
100 101 102 103 104 105
- tools/[syncsnoop](tools/syncsnoop.py): Trace sync() syscall. [Examples](tools/syncsnoop_example.txt).
- tools/[tcpaccept](tools/tcpaccept.py): Trace TCP passive connections (accept()). [Examples](tools/tcpaccept_example.txt).
- tools/[tcpconnect](tools/tcpconnect.py): Trace TCP active connections (connect()). [Examples](tools/tcpconnect_example.txt).
- tools/[vfscount](tools/vfscount.py) tools/[vfscount.c](tools/vfscount.c): Count VFS calls. [Examples](tools/vfscount_example.txt).
- tools/[vfsstat](tools/vfsstat.py) tools/[vfsstat.c](tools/vfsstat.c): Count some VFS calls, with column output. [Examples](tools/vfsstat_example.txt).
- tools/[wakeuptime](tools/wakeuptime.py): Summarize sleep to wakeup time by waker kernel stack. [Examples](tools/wakeuptime_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
106 107
- tools/[xfsdist](tools/xfsdist.py): Summarize XFS operation latency. [Examples](tools/xfsdist_example.txt).
- tools/[xfsslower](tools/xfsslower.py): Trace slow XFS operations. [Examples](tools/xfsslower_example.txt).
Brendan Gregg's avatar
Brendan Gregg committed
108 109 110 111 112

### Networking

Examples:

113 114 115 116 117 118
- examples/networking/[distributed_bridge/](examples/networking/distributed_bridge): Distributed bridge example.
- examples/networking/[simple_tc.py](examples/networking/simple_tc.py): Simple traffic control example.
- examples/networking/[simulation.py](examples/networking/simulation.py): Simulation helper.
- examples/networking/neighbor_sharing/[tc_neighbor_sharing.py](examples/networking/neighbor_sharing/tc_neighbor_sharing.py) examples/networking/neighbor_sharing/[tc_neighbor_sharing.c](examples/networking/neighbor_sharing/tc_neighbor_sharing.c): Per-IP classification and rate limiting.
- examples/networking/[tunnel_monitor/](examples/networking/tunnel_monitor): Efficiently monitor traffic flows. [Example video](https://www.youtube.com/watch?v=yYy3Cwce02k).
- examples/networking/vlan_learning/[vlan_learning.py](examples/networking/vlan_learning/vlan_learning.py) examples/[vlan_learning.c](examples/networking/vlan_learning/vlan_learning.c): Demux Ethernet traffic into worker veth+namespaces.
Brendan Gregg's avatar
Brendan Gregg committed
119

Brenden's avatar
Brenden committed
120 121 122
## Motivation

BPF guarantees that the programs loaded into the kernel cannot crash, and
123 124
cannot run forever, but yet BPF is general purpose enough to perform many
arbitrary types of computation. Currently, it is possible to write a program in
Brenden's avatar
Brenden committed
125 126
C that will compile into a valid BPF program, yet it is vastly easier to
write a C program that will compile into invalid BPF (C is like that). The user
127
won't know until trying to run the program whether it was valid or not.
Brenden's avatar
Brenden committed
128 129 130 131 132 133

With a BPF-specific frontend, one should be able to write in a language and
receive feedback from the compiler on the validity as it pertains to a BPF
backend. This toolkit aims to provide a frontend that can only create valid BPF
programs while still harnessing its full flexibility.

134 135 136 137 138 139
Furthermore, current integrations with BPF have a kludgy workflow, sometimes
involving compiling directly in a linux kernel source tree. This toolchain aims
to minimize the time that a developer spends getting BPF compiled, and instead
focus on the applications that can be written and the problems that can be
solved with BPF.

Brenden's avatar
Brenden committed
140 141
The features of this toolkit include:
* End-to-end BPF workflow in a shared library
142
  * A modified C language for BPF backends
143
  * Integration with llvm-bpf backend for JIT
Brenden's avatar
Brenden committed
144 145 146 147 148
  * Dynamic (un)loading of JITed programs
  * Support for BPF kernel hooks: socket filters, tc classifiers,
      tc actions, and kprobes
* Bindings for Python
* Examples for socket filters, tc classifiers, and kprobes
149
* Self-contained tools for tracing a running system
Brenden's avatar
Brenden committed
150

151 152
In the future, more bindings besides python will likely be supported. Feel free
to add support for the language of your choice and send a pull request!
Brenden's avatar
Brenden committed
153

Brendan Gregg's avatar
Brendan Gregg committed
154
## Tutorial
Brenden's avatar
Brenden committed
155

Brendan Gregg's avatar
Brendan Gregg committed
156 157 158
The BCC toolchain is currently composed of two parts: a C wrapper around LLVM,
and a Python API to interact with the running program. Later, we will go into
more detail of how this all works.
159

160
### Hello, World
161

162 163
First, we should include the BPF class from the bpf module:
```python
164
from bcc import BPF
165 166 167 168 169 170 171 172
```

Since the C code is so short, we will embed it inside the python script.

The BPF program always takes at least one argument, which is a pointer to the
context for this type of program. Different program types have different calling
conventions, but for this one we don't care so `void *` is fine.
```python
173
BPF(text='void kprobe__sys_clone(void *ctx) { bpf_trace_printk("Hello, World!\\n"); }').trace_print()
174
```
175

176
For this example, we will call the program every time `fork()` is called by a
177 178
userspace process. Underneath the hood, fork translates to the `clone` syscall.
BCC recognizes prefix `kprobe__`, and will auto attach our program to the kernel symbol `sys_clone`.
179

180 181 182
The python process will then print the trace printk circular buffer until ctrl-c
is pressed. The BPF program is removed from the kernel when the userspace
process that loaded it closes the fd (or exits).
183

184
Output:
185
```
186
bcc/examples$ sudo python hello_world.py
187
          python-7282  [002] d...  3757.488508: : Hello, World!
188 189
```

190 191 192
For an explanation of the meaning of the printed fields, see the trace_pipe
section of the [kernel ftrace doc](https://www.kernel.org/doc/Documentation/trace/ftrace.txt).

193
[Source code listing](examples/hello_world.py)
194

195
### Networking
196

Alex Bagehot's avatar
Alex Bagehot committed
197
At Red Hat Summit 2015, BCC was presented as part of a [session on BPF](http://www.devnation.org/#7784f1f7513e8542e4db519e79ff5eec).
198 199 200 201
A multi-host vxlan environment is simulated and a BPF program used to monitor
one of the physical interfaces. The BPF program keeps statistics on the inner
and outer IP addresses traversing the interface, and the userspace component
turns those statistics into a graph showing the traffic distribution at
202
multiple granularities. See the code [here](examples/networking/tunnel_monitor).
203 204

[![Screenshot](http://img.youtube.com/vi/yYy3Cwce02k/0.jpg)](https://youtu.be/yYy3Cwce02k)
205

206
### Tracing
207

208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232
Here is a slightly more complex tracing example than Hello World. This program
will be invoked for every task change in the kernel, and record in a BPF map
the new and old pids.

The C program below introduces two new concepts.
The first is the macro `BPF_TABLE`. This defines a table (type="hash"), with key
type `key_t` and leaf type `u64` (a single counter). The table name is `stats`,
containing 1024 entries maximum. One can `lookup`, `lookup_or_init`, `update`,
and `delete` entries from the table.
The second concept is the prev argument. This argument is treated specially by
the BCC frontend, such that accesses to this variable are read from the saved
context that is passed by the kprobe infrastructure. The prototype of the args
starting from position 1 should match the prototype of the kernel function being
kprobed. If done so, the program will have seamless access to the function
parameters.
```c
#include <uapi/linux/ptrace.h>
#include <linux/sched.h>

struct key_t {
  u32 prev_pid;
  u32 curr_pid;
};
// map_type, key_type, leaf_type, table_name, num_entry
BPF_TABLE("hash", struct key_t, u64, stats, 1024);
233 234 235
// attach to finish_task_switch in kernel/sched/core.c, which has the following
// prototype:
//   struct rq *finish_task_switch(struct task_struct *prev)
236 237 238 239 240 241 242 243 244 245 246 247
int count_sched(struct pt_regs *ctx, struct task_struct *prev) {
  struct key_t key = {};
  u64 zero = 0, *val;

  key.curr_pid = bpf_get_current_pid_tgid();
  key.prev_pid = prev->pid;

  val = stats.lookup_or_init(&key, &zero);
  (*val)++;
  return 0;
}
```
248
[Source code listing](examples/tracing/task_switch.c)
249 250

The userspace component loads the file shown above, and attaches it to the
251 252 253 254 255
`finish_task_switch` kernel function.
The [] operator of the BPF object gives access to each BPF_TABLE in the
program, allowing pass-through access to the values residing in the kernel. Use
the object as you would any other python dict object: read, update, and deletes
are all allowed.
256
```python
257
from bcc import BPF
258 259 260
from time import sleep

b = BPF(src_file="task_switch.c")
261
b.attach_kprobe(event="finish_task_switch", fn_name="count_sched")
262 263 264 265

# generate many schedule events
for i in range(0, 100): sleep(0.01)

266
for k, v in b["stats"].items():
267 268
    print("task_switch[%5d->%5d]=%u" % (k.prev_pid, k.curr_pid, v.value))
```
269
[Source code listing](examples/tracing/task_switch.py)
270

271
## Getting started
272

273
See [INSTALL.md](INSTALL.md) for installation steps on your platform.
Suchakra Sharma's avatar
Suchakra Sharma committed
274 275

## Contributing
Brendan Gregg's avatar
Brendan Gregg committed
276

Suchakra Sharma's avatar
Suchakra Sharma committed
277 278 279
Already pumped up to commit some code? Here are some resources to join the
discussions in the [IOVisor](https://www.iovisor.org/) community and see
what you want to work on.
Suchakra Sharma's avatar
Suchakra Sharma committed
280 281 282 283 284

* _Mailing List:_ http://lists.iovisor.org/mailman/listinfo/iovisor-dev
* _IRC:_ #iovisor at irc.oftc.net
* _IRC Logs:_ https://scrollback.io/iovisor/all
* _BCC Issue Tracker:_ [Github Issues](https://github.com/iovisor/bcc/issues)
Brendan Gregg's avatar
Brendan Gregg committed
285
* _A guide for contributing scripts:_ [CONTRIBUTING-SCRIPTS.md](CONTRIBUTING-SCRIPTS.md)