skia2/infra/bots/recipe_modules/flavor/android.py

# Copyright 2016 The Chromium Authors. All rights reserved.
# Use of this source code is governed by a BSD-style license that can be
# found in the LICENSE file.


from recipe_engine import recipe_api

from . import default
import subprocess  # TODO(borenet): No! Remove this.


"""Android flavor, used for running code on Android."""


class AndroidFlavor(default.DefaultFlavor):
  def __init__(self, m):
    super(AndroidFlavor, self).__init__(m)
    self._ever_ran_adb = False
    self.ADB_BINARY = '/usr/bin/adb.1.0.35'
    self.ADB_PUB_KEY = '/home/chrome-bot/.android/adbkey'
    if 'skia' not in self.m.vars.swarming_bot_id:
      self.ADB_BINARY = '/opt/infra-android/tools/adb'
      self.ADB_PUB_KEY = ('/home/chrome-bot/.android/'
                          'chrome_infrastructure_adbkey')

    # Data should go in android_data_dir, which may be preserved across runs.
    android_data_dir = '/sdcard/revenge_of_the_skiabot/'
    self.device_dirs = default.DeviceDirs(
        bin_dir        = '/data/local/tmp/',
        dm_dir         = android_data_dir + 'dm_out',
        perf_data_dir  = android_data_dir + 'perf',
        resource_dir   = android_data_dir + 'resources',
        images_dir     = android_data_dir + 'images',
        lotties_dir    = android_data_dir + 'lotties',
        skp_dir        = android_data_dir + 'skps',
        svg_dir        = android_data_dir + 'svgs',
        mskp_dir       = android_data_dir + 'mskp',
        tmp_dir        = android_data_dir,
        texttraces_dir = android_data_dir + 'text_blob_traces')

    # A list of devices we can't root.  If rooting fails and a device is not
    # on the list, we fail the task to avoid perf inconsistencies.
    self.rootable_blacklist = ['GalaxyS6', 'GalaxyS7_G930FD', 'GalaxyS9',
                               'MotoG4', 'NVIDIA_Shield', 'P30',
                               'TecnoSpark3Pro']

    # Maps device type -> CPU ids that should be scaled for nanobench.
    # Many devices have two (or more) different CPUs (e.g. big.LITTLE
    # on Nexus5x). The CPUs listed are the biggest cpus on the device.
    # The CPUs are grouped together, so we only need to scale one of them
    # (the one listed) in order to scale them all.
    # E.g. Nexus5x has cpu0-3 as one chip and cpu4-5 as the other. Thus,
    # if one wants to run a single-threaded application (e.g. nanobench), one
    # can disable cpu0-3 and scale cpu 4 to have only cpu4 and 5 at the same
    # frequency.  See also disable_for_nanobench.
    self.cpus_to_scale = {
      'Nexus5x': [4],
      'Pixel': [2],
      'Pixel2XL': [4]
    }

    # Maps device type -> CPU ids that should be turned off when running
    # single-threaded applications like nanobench. The devices listed have
    # multiple, differnt CPUs. We notice a lot of noise that seems to be
    # caused by nanobench running on the slow CPU, then the big CPU. By
    # disabling this, we see less of that noise by forcing the same CPU
    # to be used for the performance testing every time.
    self.disable_for_nanobench = {
      'Nexus5x': range(0, 4),
      'Pixel': range(0, 2),
      'Pixel2XL': range(0, 4)
    }

    self.gpu_scaling = {
      "Nexus5":  450000000,
      "Nexus5x": 600000000,
    }

  def _run(self, title, *cmd, **kwargs):
    with self.m.context(cwd=self.m.path['start_dir'].join('skia')):
      return self.m.run(self.m.step, title, cmd=list(cmd), **kwargs)

  def _adb(self, title, *cmd, **kwargs):
    # The only non-infra adb steps (dm / nanobench) happen to not use _adb().
    if 'infra_step' not in kwargs:
      kwargs['infra_step'] = True

    self._ever_ran_adb = True
    # ADB seems to be occasionally flaky on every device, so always retry.
    attempts = 3

    def wait_for_device(attempt):
      self.m.run(self.m.step,
                 'kill adb server after failure of \'%s\' (attempt %d)' % (
                     title, attempt),
                 cmd=[self.ADB_BINARY, 'kill-server'],
                 infra_step=True, timeout=30, abort_on_failure=False,
                 fail_build_on_failure=False)
      self.m.run(self.m.step,
                 'wait for device after failure of \'%s\' (attempt %d)' % (
                     title, attempt),
                 cmd=[self.ADB_BINARY, 'wait-for-device'], infra_step=True,
                 timeout=180, abort_on_failure=False,
                 fail_build_on_failure=False)

    with self.m.context(cwd=self.m.path['start_dir'].join('skia')):
      with self.m.env({'ADB_VENDOR_KEYS': self.ADB_PUB_KEY}):
        return self.m.run.with_retry(self.m.step, title, attempts,
                                     cmd=[self.ADB_BINARY]+list(cmd),
                                     between_attempts_fn=wait_for_device,
                                     **kwargs)

  def _scale_for_dm(self):
    device = self.m.vars.builder_cfg.get('model')
    if (device in self.rootable_blacklist or
        self.m.vars.internal_hardware_label):
      return

    # This is paranoia... any CPUs we disabled while running nanobench
    # ought to be back online now that we've restarted the device.
    for i in self.disable_for_nanobench.get(device, []):
      self._set_cpu_online(i, 1) # enable

    scale_up = self.cpus_to_scale.get(device, [0])
    # For big.LITTLE devices, make sure we scale the LITTLE cores up;
    # there is a chance they are still in powersave mode from when
    # swarming slows things down for cooling down and charging.
    if 0 not in scale_up:
      scale_up.append(0)
    for i in scale_up:
      # AndroidOne doesn't support ondemand governor. hotplug is similar.
      if device == 'AndroidOne':
        self._set_governor(i, 'hotplug')
      elif device in ['Pixel3a', 'Pixel4']:
        # Pixel3a/4 have userspace powersave performance schedutil.
        # performance seems like a reasonable choice.
        self._set_governor(i, 'performance')
      else:
        self._set_governor(i, 'ondemand')

  def _scale_for_nanobench(self):
    device = self.m.vars.builder_cfg.get('model')
    if (device in self.rootable_blacklist or
      self.m.vars.internal_hardware_label):
      return

    for i in self.cpus_to_scale.get(device, [0]):
      self._set_governor(i, 'userspace')
      self._scale_cpu(i, 0.6)

    for i in self.disable_for_nanobench.get(device, []):
      self._set_cpu_online(i, 0) # disable

    if device in self.gpu_scaling:
      #https://developer.qualcomm.com/qfile/28823/lm80-p0436-11_adb_commands.pdf
      # Section 3.2.1 Commands to put the GPU in performance mode
      # Nexus 5 is  320000000 by default
      # Nexus 5x is 180000000 by default
      gpu_freq = self.gpu_scaling[device]
      self.m.run.with_retry(self.m.python.inline,
        "Lock GPU to %d (and other perf tweaks)" % gpu_freq,
        3, # attempts
        program="""
import os
import subprocess
import sys
import time
ADB = sys.argv[1]
freq = sys.argv[2]
idle_timer = "10000"

log = subprocess.check_output([ADB, 'root'])
# check for message like 'adbd cannot run as root in production builds'
print log
if 'cannot' in log:
  raise Exception('adb root failed')

subprocess.check_output([ADB, 'shell', 'stop', 'thermald'])

subprocess.check_output([ADB, 'shell', 'echo "%s" > '
    '/sys/class/kgsl/kgsl-3d0/gpuclk' % freq])

actual_freq = subprocess.check_output([ADB, 'shell', 'cat '
    '/sys/class/kgsl/kgsl-3d0/gpuclk']).strip()
if actual_freq != freq:
  raise Exception('Frequency (actual, expected) (%s, %s)'
                  % (actual_freq, freq))

subprocess.check_output([ADB, 'shell', 'echo "%s" > '
    '/sys/class/kgsl/kgsl-3d0/idle_timer' % idle_timer])

actual_timer = subprocess.check_output([ADB, 'shell', 'cat '
    '/sys/class/kgsl/kgsl-3d0/idle_timer']).strip()
if actual_timer != idle_timer:
  raise Exception('idle_timer (actual, expected) (%s, %s)'
                  % (actual_timer, idle_timer))

for s in ['force_bus_on', 'force_rail_on', 'force_clk_on']:
  subprocess.check_output([ADB, 'shell', 'echo "1" > '
      '/sys/class/kgsl/kgsl-3d0/%s' % s])
  actual_set = subprocess.check_output([ADB, 'shell', 'cat '
      '/sys/class/kgsl/kgsl-3d0/%s' % s]).strip()
  if actual_set != "1":
    raise Exception('%s (actual, expected) (%s, 1)'
                    % (s, actual_set))
""",
        args = [self.ADB_BINARY, gpu_freq],
        infra_step=True,
        timeout=30)

  def _set_governor(self, cpu, gov):
    self._ever_ran_adb = True
    self.m.run.with_retry(self.m.python.inline,
        "Set CPU %d's governor to %s" % (cpu, gov),
        3, # attempts
        program="""
import os
import subprocess
import sys
import time
ADB = sys.argv[1]
cpu = int(sys.argv[2])
gov = sys.argv[3]

log = subprocess.check_output([ADB, 'root'])
# check for message like 'adbd cannot run as root in production builds'
print log
if 'cannot' in log:
  raise Exception('adb root failed')

subprocess.check_output([ADB, 'shell', 'echo "%s" > '
    '/sys/devices/system/cpu/cpu%d/cpufreq/scaling_governor' % (gov, cpu)])
actual_gov = subprocess.check_output([ADB, 'shell', 'cat '
    '/sys/devices/system/cpu/cpu%d/cpufreq/scaling_governor' % cpu]).strip()
if actual_gov != gov:
  raise Exception('(actual, expected) (%s, %s)'
                  % (actual_gov, gov))
""",
        args = [self.ADB_BINARY, cpu, gov],
        infra_step=True,
        timeout=30)


  def _set_cpu_online(self, cpu, value):
    """Set /sys/devices/system/cpu/cpu{N}/online to value (0 or 1)."""
    self._ever_ran_adb = True
    msg = 'Disabling'
    if value:
      msg = 'Enabling'
    self.m.run.with_retry(self.m.python.inline,
        '%s CPU %d' % (msg, cpu),
        3, # attempts
        program="""
import os
import subprocess
import sys
import time
ADB = sys.argv[1]
cpu = int(sys.argv[2])
value = int(sys.argv[3])

log = subprocess.check_output([ADB, 'root'])
# check for message like 'adbd cannot run as root in production builds'
print log
if 'cannot' in log:
  raise Exception('adb root failed')

# If we try to echo 1 to an already online cpu, adb returns exit code 1.
# So, check the value before trying to write it.
prior_status = subprocess.check_output([ADB, 'shell', 'cat '
    '/sys/devices/system/cpu/cpu%d/online' % cpu]).strip()
if prior_status == str(value):
  print 'CPU %d online already %d' % (cpu, value)
  sys.exit()

subprocess.check_output([ADB, 'shell', 'echo %s > '
    '/sys/devices/system/cpu/cpu%d/online' % (value, cpu)])
actual_status = subprocess.check_output([ADB, 'shell', 'cat '
    '/sys/devices/system/cpu/cpu%d/online' % cpu]).strip()
if actual_status != str(value):
  raise Exception('(actual, expected) (%s, %d)'
                  % (actual_status, value))
""",
        args = [self.ADB_BINARY, cpu, value],
        infra_step=True,
        timeout=30)


  def _scale_cpu(self, cpu, target_percent):
    self._ever_ran_adb = True
    self.m.run.with_retry(self.m.python.inline,
        'Scale CPU %d to %f' % (cpu, target_percent),
        3, # attempts
        program="""
import os
import subprocess
import sys
import time
ADB = sys.argv[1]
target_percent = float(sys.argv[2])
cpu = int(sys.argv[3])
log = subprocess.check_output([ADB, 'root'])
# check for message like 'adbd cannot run as root in production builds'
print log
if 'cannot' in log:
  raise Exception('adb root failed')

root = '/sys/devices/system/cpu/cpu%d/cpufreq' %cpu

# All devices we test on give a list of their available frequencies.
available_freqs = subprocess.check_output([ADB, 'shell',
    'cat %s/scaling_available_frequencies' % root])

# Check for message like '/system/bin/sh: file not found'
if available_freqs and '/system/bin/sh' not in available_freqs:
  available_freqs = sorted(
      int(i) for i in available_freqs.strip().split())
else:
  raise Exception('Could not get list of available frequencies: %s' %
                  available_freqs)

maxfreq = available_freqs[-1]
target = int(round(maxfreq * target_percent))
freq = maxfreq
for f in reversed(available_freqs):
  if f <= target:
    freq = f
    break

print 'Setting frequency to %d' % freq

# If scaling_max_freq is lower than our attempted setting, it won't take.
# We must set min first, because if we try to set max to be less than min
# (which sometimes happens after certain devices reboot) it returns a
# perplexing permissions error.
subprocess.check_output([ADB, 'shell', 'echo 0 > '
    '%s/scaling_min_freq' % root])
subprocess.check_output([ADB, 'shell', 'echo %d > '
    '%s/scaling_max_freq' % (freq, root)])
subprocess.check_output([ADB, 'shell', 'echo %d > '
    '%s/scaling_setspeed' % (freq, root)])
time.sleep(5)
actual_freq = subprocess.check_output([ADB, 'shell', 'cat '
    '%s/scaling_cur_freq' % root]).strip()
if actual_freq != str(freq):
  raise Exception('(actual, expected) (%s, %d)'
                  % (actual_freq, freq))
""",
        args = [self.ADB_BINARY, str(target_percent), cpu],
        infra_step=True,
        timeout=30)


  def install(self):
    self._adb('mkdir ' + self.device_dirs.resource_dir,
              'shell', 'mkdir', '-p', self.device_dirs.resource_dir)
    if 'ASAN' in self.m.vars.extra_tokens:
      self._ever_ran_adb = True
      asan_setup = self.m.vars.slave_dir.join(
            'android_ndk_linux', 'toolchains', 'llvm', 'prebuilt',
            'linux-x86_64', 'lib64', 'clang', '8.0.7', 'bin',
            'asan_device_setup')
      self.m.run(self.m.python.inline, 'Setting up device to run ASAN',
        program="""
import os
import subprocess
import sys
import time
ADB = sys.argv[1]
ASAN_SETUP = sys.argv[2]

def wait_for_device():
  while True:
    time.sleep(5)
    print 'Waiting for device'
    subprocess.check_output([ADB, 'wait-for-device'])
    bit1 = subprocess.check_output([ADB, 'shell', 'getprop',
                                   'dev.bootcomplete'])
    bit2 = subprocess.check_output([ADB, 'shell', 'getprop',
                                   'sys.boot_completed'])
    if '1' in bit1 and '1' in bit2:
      print 'Device detected'
      break

log = subprocess.check_output([ADB, 'root'])
# check for message like 'adbd cannot run as root in production builds'
print log
if 'cannot' in log:
  raise Exception('adb root failed')

output = subprocess.check_output([ADB, 'disable-verity'])
print output

if 'already disabled' not in output:
  print 'Rebooting device'
  subprocess.check_output([ADB, 'reboot'])
  wait_for_device()

def installASAN(revert=False):
  # ASAN setup script is idempotent, either it installs it or
  # says it's installed.  Returns True on success, false otherwise.
  out = subprocess.check_output([ADB, 'wait-for-device'])
  print out
  cmd = [ASAN_SETUP]
  if revert:
    cmd = [ASAN_SETUP, '--revert']
  process = subprocess.Popen(cmd, env={'ADB': ADB},
                             stdout=subprocess.PIPE, stderr=subprocess.PIPE)

  # this also blocks until command finishes
  (stdout, stderr) = process.communicate()
  print stdout
  print 'Stderr: %s' % stderr
  return process.returncode == 0

if not installASAN():
  print 'Trying to revert the ASAN install and then re-install'
  # ASAN script sometimes has issues if it was interrupted or partially applied
  # Try reverting it, then re-enabling it
  if not installASAN(revert=True):
    raise Exception('reverting ASAN install failed')

  # Sleep because device does not reboot instantly
  time.sleep(10)

  if not installASAN():
    raise Exception('Tried twice to setup ASAN and failed.')

# Sleep because device does not reboot instantly
time.sleep(10)
wait_for_device()
# Sleep again to hopefully avoid error "secure_mkdirs failed: No such file or
# directory" when pushing resources to the device.
time.sleep(60)
""",
        args = [self.ADB_BINARY, asan_setup],
          infra_step=True,
          timeout=300,
          abort_on_failure=True)


  def cleanup_steps(self):
    if 'ASAN' in self.m.vars.extra_tokens:
      self._ever_ran_adb = True
      # Remove ASAN.
      asan_setup = self.m.vars.slave_dir.join(
            'android_ndk_linux', 'toolchains', 'llvm', 'prebuilt',
            'linux-x86_64', 'lib64', 'clang', '8.0.2', 'bin',
            'asan_device_setup')
      self.m.run(self.m.step,
                 'wait for device before uninstalling ASAN',
                 cmd=[self.ADB_BINARY, 'wait-for-device'], infra_step=True,
                 timeout=180, abort_on_failure=False,
                 fail_build_on_failure=False)
      self.m.run(self.m.step, 'uninstall ASAN',
                 cmd=[asan_setup, '--revert'], infra_step=True, timeout=300,
                 abort_on_failure=False, fail_build_on_failure=False)

    if self._ever_ran_adb:
      self.m.run(self.m.python.inline, 'dump log', program="""
          import os
          import subprocess
          import sys
          out = sys.argv[1]
          log = subprocess.check_output(['%s', 'logcat', '-d'])
          for line in log.split('\\n'):
            tokens = line.split()
            if len(tokens) == 11 and tokens[-7] == 'F' and tokens[-3] == 'pc':
              addr, path = tokens[-2:]
              local = os.path.join(out, os.path.basename(path))
              if os.path.exists(local):
                sym = subprocess.check_output(['addr2line', '-Cfpe', local, addr])
                line = line.replace(addr, addr + ' ' + sym.strip())
            print line
          """ % self.ADB_BINARY,
          args=[self.host_dirs.bin_dir],
          infra_step=True,
          timeout=300,
          abort_on_failure=False)

    # Only quarantine the bot if the first failed step
    # is an infra step. If, instead, we did this for any infra failures, we
    # would do this too much. For example, if a Nexus 10 died during dm
    # and the following pull step would also fail "device not found" - causing
    # us to run the shutdown command when the device was probably not in a
    # broken state; it was just rebooting.
    if (self.m.run.failed_steps and
        isinstance(self.m.run.failed_steps[0], recipe_api.InfraFailure)):
      bot_id = self.m.vars.swarming_bot_id
      self.m.file.write_text('Quarantining Bot',
                             '/home/chrome-bot/%s.force_quarantine' % bot_id,
                             ' ')

    if self._ever_ran_adb:
      self._adb('kill adb server', 'kill-server')

  def step(self, name, cmd, **kwargs):
    if not kwargs.get('skip_binary_push', False):
      if (cmd[0] == 'nanobench'):
        self._scale_for_nanobench()
      else:
        self._scale_for_dm()
      app = self.host_dirs.bin_dir.join(cmd[0])
      self._adb('push %s' % cmd[0],
                'push', app, self.device_dirs.bin_dir)

    sh = '%s.sh' % cmd[0]
    self.m.run.writefile(self.m.vars.tmp_dir.join(sh),
        'set -x; %s%s; echo $? >%src' % (
            self.device_dirs.bin_dir, subprocess.list2cmdline(map(str, cmd)),
            self.device_dirs.bin_dir))
    self._adb('push %s' % sh,
              'push', self.m.vars.tmp_dir.join(sh), self.device_dirs.bin_dir)

    self._adb('clear log', 'logcat', '-c')
    self.m.python.inline('%s' % cmd[0], """
    import subprocess
    import sys
    bin_dir = sys.argv[1]
    sh      = sys.argv[2]
    subprocess.check_call(['%s', 'shell', 'sh', bin_dir + sh])
    try:
      sys.exit(int(subprocess.check_output(['%s', 'shell', 'cat',
                                            bin_dir + 'rc'])))
    except ValueError:
      print "Couldn't read the return code.  Probably killed for OOM."
      sys.exit(1)
    """ % (self.ADB_BINARY, self.ADB_BINARY),
      args=[self.device_dirs.bin_dir, sh])

  def copy_file_to_device(self, host, device):
    self._adb('push %s %s' % (host, device), 'push', host, device)

  def copy_directory_contents_to_device(self, host, device):
    # Copy the tree, avoiding hidden directories and resolving symlinks.
    sep = self.m.path.sep
    host_str = str(host).rstrip(sep) + sep
    device = device.rstrip('/')
    with self.m.step.nest('push %s* %s' % (host_str, device)):
      contents = self.m.file.listdir('list %s' % host, host, recursive=True,
                                     test_data=['file1',
                                                'subdir' + sep + 'file2',
                                                '.file3',
                                                '.ignore' + sep + 'file4'])
      for path in contents:
        path_str = str(path)
        assert path_str.startswith(host_str), (
            'expected %s to have %s as a prefix' % (path_str, host_str))
        relpath = path_str[len(host_str):]
        # NOTE(dogben): Previous logic used os.walk and skipped directories
        # starting with '.', but not files starting with '.'. It's not clear
        # what the reason was (maybe skipping .git?), but I'm keeping that
        # behavior here.
        if self.m.path.dirname(relpath).startswith('.'):
          continue
        device_path = device + '/' + relpath  # Android paths use /
        self._adb('push %s' % path, 'push',
                  self.m.path.realpath(path), device_path)

  def copy_directory_contents_to_host(self, device, host):
    # TODO(borenet): When all of our devices are on Android 6.0 and up, we can
    # switch to using tar to zip up the results before pulling.
    with self.m.step.nest('adb pull'):
      tmp = self.m.path.mkdtemp('adb_pull')
      self._adb('pull %s' % device, 'pull', device, tmp)
      paths = self.m.file.glob_paths(
          'list pulled files',
          tmp,
          self.m.path.basename(device) + self.m.path.sep + '*',
          test_data=['%d.png' % i for i in (1, 2)])
      for p in paths:
        self.m.file.copy('copy %s' % self.m.path.basename(p), p, host)

  def read_file_on_device(self, path, **kwargs):
    rv = self._adb('read %s' % path,
                   'shell', 'cat', path, stdout=self.m.raw_io.output(),
                   **kwargs)
    return rv.stdout.rstrip() if rv and rv.stdout else None

  def remove_file_on_device(self, path):
    self._adb('rm %s' % path, 'shell', 'rm', '-f', path)

  def create_clean_device_dir(self, path):
    self._adb('rm %s' % path, 'shell', 'rm', '-rf', path)
    self._adb('mkdir %s' % path, 'shell', 'mkdir', '-p', path)