#output #protobuf #produces #formatted #perfetto #track #bpftrace

app btetto

A tool that produces Perfetto protobuf from formatted bpftrace output

5 releases

new 0.1.4 Nov 18, 2024
0.1.3 Nov 15, 2024
0.1.2 Nov 11, 2024
0.1.1 Nov 8, 2024
0.1.0 Nov 7, 2024

#174 in Encoding

Download history 52/week @ 2024-11-01 286/week @ 2024-11-08 248/week @ 2024-11-15

586 downloads per month

BSD-2-Clause OR LGPL-2.1

390KB
690 lines

btetto

A tool that produces Perfetto protobuf from formatted bpftrace output.

Rust Crate available here

Usage

$ sudo bpftrace my_script.bt -f json | btetto
Attached probes: 4
^C
Writing 149 events to trace file: bpftrace_trace.binpb

You can also pass a bpftrace output file to btetto e.g.

btetto my_bpftrace_output

btetto.py produces a bpftrace_trace.binpb protobuf file, which can then be loaded into the Perfetto UI.

bpftrace Output Format

The print output from bpftrace should be tuples (in JSON format e.g. -f json) where the first item in the tuple is the event type and the rest of the items are key/value tuples.

Examples

Event Types

  • track_event
  • call_stack
  • stdout

Track Events (Spans)

Required Fields:

  • name (string)
  • ts (timestamp)
  • type (string - see below)

Optional Fields:

  • pid (number)
  • thread_name (string)
  • tid (number)
  • track (string or number)
  • track_parent (string or number)
  • unit (string - see below)
  • flow_id (string or number)
  • log (tuple - see below)

Track Event Types

  • BEGIN
  • END
  • INSTANT
  • COUNTER

If the field is not listed above it will get logged as an annotation on the event like "bananas" and "greeting" below. pid, tid, and thread_name also get logged as annotations by default.

print(("track_event",
    ("name", "page_fault_user"),
    ("type", "BEGIN"),
    ("ts", $start),
    ("pid", pid),
    ("tid", tid),
    ("thread_name", comm),
    ("bananas", 10),
    ("greeting", "hello"),
    ("log", ("WARN", "this is my log message"))
));

print(("track_event",
    ("name", "page_fault_user"),
    ("type", "END"),
    ("ts", nsecs),
    ("pid", pid),
    ("tid", tid),
    ("thread_name", comm)
));

track and track_parent

These are used to name the "tracks" where these events exist. At the moment they can be nested one level, where you would provide both a track and a track_parent tuple (both strings).

If track is not provided, you must then provide pid, tid, and thread_name tuples and then these track events will go into global pid/tid "tracks".

unit

These are for COUNTER type track events and can be:

  • unspecified
  • count (default if no "unit" is provided)
  • size_bytes
  • time_ns

Example:

print(("track_event",
    ("name", "Max Duration"),
    ("type", "COUNTER"),
    ("ts", nsecs),
    ("track", "Max Duration"),
    ("unit", "count"),
    ("counter_value", @mx)
));

log

The log tuple is a little different in that the value is another tuple where the first field is the log level and the second field is the log message e.g. ("log", ("FATAL", "This is an error message")). These show up as "Android Logs" in Perfetto.

Valid Log Levels

  • UNSPECIFIED
  • UNUSED
  • VERBOSE
  • DEBUG
  • INFO
  • WARN
  • ERROR
  • FATAL

Call Stack Sample

These are for logging call stacks (kernel, user, or both) at specific points in time. They do not have durations.

Required Fields:

  • pid (number)
  • tid (number)
  • ts (timestamp)
  • kstack and/or ustack (array of strings)

Optional Fields:

  • thread_name (string)
print(("call_stack",
    ("ts", nsecs),
    ("pid", pid),
    ("tid", tid),
    ("thread_name", comm),
    ("kstack", kstack),
    ("ustack", ustack)
));

stdout

This just prints the value to the command line e.g.

BEGIN {
    print(("stdout", "Tracks the duration of page faults"));
}

Dependencies

~4–12MB
~152K SLoC