xref: /aosp_15_r20/external/perfetto/infra/ci/worker/run_job.py (revision 6dbdd20afdafa5e3ca9b8809fa73465d530080dc)
1#!/usr/bin/env python3
2# Copyright (C) 2019 The Android Open Source Project
3#
4# Licensed under the Apache License, Version 2.0 (the "License");
5# you may not use this file except in compliance with the License.
6# You may obtain a copy of the License at
7#
8#      http://www.apache.org/licenses/LICENSE-2.0
9#
10# Unless required by applicable law or agreed to in writing, software
11# distributed under the License is distributed on an "AS IS" BASIS,
12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13# See the License for the specific language governing permissions and
14# limitations under the License.
15''' Runs the given job in an isolated docker container.
16
17Also streams stdout/err onto the firebase realtime DB.
18'''
19
20import fcntl
21import logging
22import os
23import queue
24import signal
25import socket
26import subprocess
27import sys
28import threading
29import time
30
31from config import DB, SANDBOX_IMG
32from common_utils import init_logging, req, SCOPES
33
34CUR_DIR = os.path.dirname(__file__)
35SCOPES.append('https://www.googleapis.com/auth/firebase.database')
36SCOPES.append('https://www.googleapis.com/auth/userinfo.email')
37
38
39def read_nonblock(fd):
40  fcntl.fcntl(fd, fcntl.F_SETFL, fcntl.fcntl(fd, fcntl.F_GETFL) | os.O_NONBLOCK)
41  res = ''
42  while True:
43    try:
44      buf = os.read(fd.fileno(), 8192)
45      if not buf:
46        break
47      # There are two reasons for the errors='ignore' here:
48      # 1: By reading the pipe in chunks of N bytes, we can end up truncating
49      #    a valid multi-byte character and cause an "unexpected end of data".
50      #    This means that we will skip valid unicode chars if they happen to
51      #    span across two read() chunks.
52      # 2: The job output might just emit some invalid unicode in stdout. We
53      #    don't want to crash when that happens.
54      # See b/194053229 for more context.
55      res += buf.decode('utf-8', errors='ignore')
56    except OSError:
57      break
58  return res
59
60
61def log_thread(job_id, queue):
62  ''' Uploads stdout/stderr from the queue to the firebase DB.
63
64  Each line is logged as an invidivual entry in the DB, as follows:
65  MMMMMM-NNNN log line, where M: hex-encodeed timestamp, N:  monotonic counter.
66  '''
67  uri = '%s/logs/%s.json' % (DB, job_id)
68  req('DELETE', uri)
69  while True:
70    batch = queue.get()
71    if batch is None:
72      break  # EOF
73    req('PATCH', uri, body=batch)
74  logging.debug('Uploader thread terminated')
75
76
77def main(argv):
78  init_logging()
79  if len(argv) != 2:
80    print('Usage: %s job_id' % argv[0])
81    return 1
82
83  job_id = argv[1]
84  res = 42
85
86  # The container name will be worker-N-sandbox.
87  container = socket.gethostname() + '-sandbox'
88
89  # Remove stale jobs, if any.
90  subprocess.call(['sudo', 'docker', 'rm', '-f', container])
91
92  q = queue.Queue()
93
94  # Conversely to real programs, signal handlers in python aren't really async
95  # but are queued on the main thread. Hence We need to keep the main thread
96  # responsive to react to signals. This is to handle timeouts and graceful
97  # termination of the worker container, which dispatches a SIGTERM on stop.
98  def sig_handler(sig, _):
99    logging.warning('Job runner got signal %s, terminating job %s', sig, job_id)
100    subprocess.call(['sudo', 'docker', 'kill', container])
101    os._exit(1)  # sys.exit throws a SystemExit exception, _exit really exits.
102
103  signal.signal(signal.SIGTERM, sig_handler)
104
105  log_thd = threading.Thread(target=log_thread, args=(job_id, q))
106  log_thd.start()
107
108  # SYS_PTRACE is required for gtest death tests and LSan.
109  cmd = [
110      'sudo', 'docker', 'run', '--name', container, '--hostname', container,
111      '--cap-add', 'SYS_PTRACE', '--rm', '--env',
112      'PERFETTO_TEST_JOB=%s' % job_id, '--tmpfs', '/tmp:exec'
113  ]
114
115  # Propagate environment variables coming from the job config.
116  for kv in [kv for kv in os.environ.items() if kv[0].startswith('PERFETTO_')]:
117    cmd += ['--env', '%s=%s' % kv]
118
119  # We use the tmpfs mount created by gce-startup-script.sh, if present. The
120  # problem is that Docker doesn't allow to both override the tmpfs-size and
121  # prevent the "-o noexec". In turn the default tmpfs-size depends on the host
122  # phisical memory size.
123  if os.getenv('SANDBOX_TMP'):
124    cmd += ['-v', '%s:/ci/ramdisk' % os.getenv('SANDBOX_TMP')]
125  else:
126    cmd += ['--tmpfs', '/ci/ramdisk:exec']
127
128  # Rationale for the conditional branches below: when running in the real GCE
129  # environment, the gce-startup-script.sh mounts these directories in the right
130  # locations, so that they are shared between all workers.
131  # When running the worker container outside of GCE (i.e.for local testing) we
132  # leave these empty. The VOLUME directive in the dockerfile will cause docker
133  # to automatically mount a scratch volume for those.
134  # This is so that the CI containers can be tested without having to do the
135  # work that gce-startup-script.sh does.
136  if os.getenv('SHARED_WORKER_CACHE'):
137    cmd += ['--volume=%s:/ci/cache' % os.getenv('SHARED_WORKER_CACHE')]
138
139  artifacts_dir = None
140  if os.getenv('ARTIFACTS_DIR'):
141    artifacts_dir = os.path.join(os.getenv('ARTIFACTS_DIR'), job_id)
142    subprocess.call(['sudo', 'rm', '-rf', artifacts_dir])
143    os.mkdir(artifacts_dir)
144    cmd += ['--volume=%s:/ci/artifacts' % artifacts_dir]
145
146  cmd += os.getenv('SANDBOX_NETWORK_ARGS', '').split()
147  cmd += [SANDBOX_IMG]
148
149  logging.info('Starting %s', ' '.join(cmd))
150  proc = subprocess.Popen(
151      cmd,
152      stdin=open(os.devnull),
153      stdout=subprocess.PIPE,
154      stderr=subprocess.STDOUT,
155      bufsize=65536)
156  stdout = ''
157  tstart = time.time()
158  while True:
159    ms_elapsed = int((time.time() - tstart) * 1000)
160    stdout += read_nonblock(proc.stdout)
161
162    # stdout/err pipes are not atomic w.r.t. '\n'. Extract whole lines out into
163    # |olines| and keep the last partial line (-1) in the |stdout| buffer.
164    lines = stdout.split('\n')
165    stdout = lines[-1]
166    lines = lines[:-1]
167
168    # Each line has a key of the form <time-from-start><out|err><counter>
169    # |counter| is relative to the batch and is only used to disambiguate lines
170    # fetched at the same time, preserving the ordering.
171    batch = {}
172    for counter, line in enumerate(lines):
173      batch['%06x-%04x' % (ms_elapsed, counter)] = line
174    if batch:
175      q.put(batch)
176    if proc.poll() is not None:
177      res = proc.returncode
178      logging.info('Job subprocess terminated with code %s', res)
179      break
180
181    # Large sleeps favour batching in the log uploader.
182    # Small sleeps favour responsiveness of the signal handler.
183    time.sleep(1)
184
185  q.put(None)  # EOF maker
186  log_thd.join()
187
188  if artifacts_dir:
189    artifacts_uploader = os.path.join(CUR_DIR, 'artifacts_uploader.py')
190    cmd = ['setsid', artifacts_uploader, '--job-id=%s' % job_id, '--rm']
191    subprocess.call(cmd)
192
193  return res
194
195
196if __name__ == '__main__':
197  sys.exit(main(sys.argv))
198