1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
|
# frozen_string_literal: true
# trigger stackprof by sending a SIGUSR2 signal
#
# default settings:
# * collect raw samples
# * sample at 100hz (every 10k microseconds)
# * timeout profile after 30 seconds
# * write to $TMPDIR/stackprof.$PID.$RAND.profile
module Gitlab
class StackProf
# this is a workaround for sidekiq, which defines its own SIGUSR2 handler.
# by defering to the sidekiq startup event, we get to set up our own
# handler late enough.
# see also: https://github.com/mperham/sidekiq/pull/4653
def self.install
require 'stackprof'
require 'tmpdir'
if Gitlab::Runtime.sidekiq?
Sidekiq.configure_server do |config|
config.on :startup do
on_worker_start
end
end
else
Gitlab::Cluster::LifecycleEvents.on_worker_start do
on_worker_start
end
end
end
def self.on_worker_start
Gitlab::AppJsonLogger.info(
event: "stackprof",
message: "listening on SIGUSR2 signal",
pid: Process.pid
)
# create a pipe in order to propagate signal out of the signal handler
# see also: https://cr.yp.to/docs/selfpipe.html
read, write = IO.pipe
# create a separate thread that polls for signals on the pipe.
#
# this way we do not execute in signal handler context, which
# lifts restrictions and also serializes the calls in a thread-safe
# manner.
#
# it's very similar to a goroutine and channel design.
#
# another nice benefit of this method is that we can timeout the
# IO.select call, allowing the profile to automatically stop after
# a given interval (by default 30 seconds), avoiding unbounded memory
# growth from a profile that was started and never stopped.
t = Thread.new do
timeout_s = ENV['STACKPROF_TIMEOUT_S']&.to_i || 30
current_timeout_s = nil
loop do
got_value = IO.select([read], nil, nil, current_timeout_s)
read.getbyte if got_value
if ::StackProf.running?
stackprof_file_prefix = ENV['STACKPROF_FILE_PREFIX'] || Dir.tmpdir
stackprof_out_file = "#{stackprof_file_prefix}/stackprof.#{Process.pid}.#{SecureRandom.hex(6)}.profile"
Gitlab::AppJsonLogger.info(
event: "stackprof",
message: "stopping profile",
output_filename: stackprof_out_file,
pid: Process.pid,
timeout_s: timeout_s,
timed_out: got_value.nil?
)
::StackProf.stop
::StackProf.results(stackprof_out_file)
current_timeout_s = nil
else
Gitlab::AppJsonLogger.info(
event: "stackprof",
message: "starting profile",
pid: Process.pid
)
::StackProf.start(
mode: :cpu,
raw: Gitlab::Utils.to_boolean(ENV['STACKPROF_RAW'] || 'true'),
interval: ENV['STACKPROF_INTERVAL_US']&.to_i || 10_000
)
current_timeout_s = timeout_s
end
end
end
t.abort_on_exception = true
# in the case of puma, this will override the existing SIGUSR2 signal handler
# that can be used to trigger a restart.
#
# puma cluster has two types of restarts:
# * SIGUSR1: phased restart
# * SIGUSR2: restart
#
# phased restart is not supported in our configuration, because we use
# preload_app. this means we will always perform a normal restart.
# additionally, phased restart is not supported when sending a SIGUSR2
# directly to a puma worker (as opposed to the master process).
#
# the result is that the behaviour of SIGUSR1 and SIGUSR2 is identical in
# our configuration, and we can always use a SIGUSR1 to perform a restart.
#
# thus, it is acceptable for us to re-appropriate the SIGUSR2 signal, and
# override the puma behaviour.
#
# see also:
# * https://github.com/puma/puma/blob/master/docs/signals.md#puma-signals
# * https://github.com/phusion/unicorn/blob/master/SIGNALS
# * https://github.com/mperham/sidekiq/wiki/Signals
Signal.trap('SIGUSR2') do
write.write('.')
end
end
end
end
if Gitlab::Utils.to_boolean(ENV['STACKPROF_ENABLED'].to_s)
Gitlab::StackProf.install
end
|
