xref: /aosp_15_r20/external/autotest/utils/frozen_chromite/lib/locking.py (revision 9c5db1993ded3edbeafc8092d69fe5de2ee02df7)

# -*- coding: utf-8 -*-
# Copyright (c) 2012 The Chromium OS Authors. All rights reserved.
# Use of this source code is governed by a BSD-style license that can be
# found in the LICENSE file.

"""Basic locking functionality."""

from __future__ import print_function

import contextlib
import os
import errno
import fcntl
import stat
import tempfile

from autotest_lib.utils.frozen_chromite.lib import cros_build_lib
from autotest_lib.utils.frozen_chromite.lib import cros_logging as logging
from autotest_lib.utils.frozen_chromite.lib import osutils
from autotest_lib.utils.frozen_chromite.lib import retry_util
from autotest_lib.utils.frozen_chromite.lib import timeout_util


LOCKF = 'lockf'
FLOCK = 'flock'


class LockNotAcquiredError(Exception):
  """Signals that the lock was not acquired."""


class LockingError(Exception):
  """Signals miscellaneous problems in the locking process."""


@contextlib.contextmanager
def _optional_timer_context(timeout):
  """Use the timeout_util.Timeout contextmanager if timeout is set."""
  if timeout:
    with timeout_util.Timeout(timeout):
      yield
  else:
    yield


class _Lock(cros_build_lib.MasterPidContextManager):
  """Base lockf based locking.  Derivatives need to override _GetFd"""

  def __init__(self, description=None, verbose=True, locktype=LOCKF,
               blocking=True, blocking_timeout=None):
    """Initialize this instance.

    Two types of locks are available: LOCKF and FLOCK.

    Use LOCKF (POSIX locks) if:
      - you need to lock a file between processes created by the
        parallel/multiprocess libraries

    Use FLOCK (BSD locks) if these scenarios apply:
      - you need to lock a file between shell scripts running the flock program
      - you need the lock to be bound to the fd and thus inheritable across
        execs

    Note: These two locks are completely independent; using one on a path will
          not block using the other on the same path.

    Args:
      path: On disk pathway to lock.  Can be a directory or a file.
      description: A description for this lock- what is it protecting?
      verbose: Verbose logging?
      locktype: Type of lock to use (lockf or flock).
      blocking: If True, use a blocking lock.
      blocking_timeout: If not None, time is seconds to wait on blocking calls.
    """
    cros_build_lib.MasterPidContextManager.__init__(self)
    self._verbose = verbose
    self.description = description
    self._fd = None
    self.locking_mechanism = fcntl.flock if locktype == FLOCK else fcntl.lockf
    # Store (to log) the locktype string.
    self.locktype = locktype
    self.blocking = blocking
    self.blocking_timeout = blocking_timeout

  @property
  def fd(self):
    if self._fd is None:
      self._fd = self._GetFd()
      # Ensure that all derivatives of this lock can't bleed the fd
      # across execs.
      fcntl.fcntl(self._fd, fcntl.F_SETFD,
                  fcntl.fcntl(self._fd, fcntl.F_GETFD) | fcntl.FD_CLOEXEC)
    return self._fd

  def _GetFd(self):
    raise NotImplementedError(self, '_GetFd')

  def _enforce_lock(self, flags, message):
    # Try nonblocking first, if it fails, display the context/message,
    # and then wait on the lock.
    try:
      self.locking_mechanism(self.fd, flags|fcntl.LOCK_NB)
      return
    except EnvironmentError as e:
      if e.errno == errno.EDEADLK:
        self.unlock()
      elif e.errno != errno.EAGAIN:
        raise
    if self.description:
      message = '%s: blocking (LOCK_NB) (%s) while %s' % (self.description,
                                                          self.locktype,
                                                          message)
    if not self.blocking:
      self.close()
      raise LockNotAcquiredError(message)
    if self._verbose:
      logging.info(message)

    try:
      with _optional_timer_context(self.blocking_timeout):
        self.locking_mechanism(self.fd, flags)
    except timeout_util.TimeoutError:
      description = self.description or 'locking._enforce_lock'
      logging.error(
          'Timed out after waiting %d seconds for blocking lock (%s): %s',
          self.blocking_timeout, self.locktype, description)
      raise
    except EnvironmentError as e:
      if e.errno != errno.EDEADLK:
        message = ('%s: blocking wait failed errno %s'
                   % (self.description, e))
        raise
      self.unlock()
      self.locking_mechanism(self.fd, flags)
    logging.debug('%s: lock has been acquired (%s), continuing.',
                  self.description, self.locktype)

  def lock(self, shared=False):
    """Take a lock of type |shared|.

    Any existing lock will be updated if need be.

    Args:
      shared: If True make the lock shared.

    Returns:
      self, allowing it to be used as a `with` target.

    Raises:
      IOError if the operation fails in some way.
      LockNotAcquiredError if the lock couldn't be acquired (non-blocking
        mode only).
    """
    self._enforce_lock(
        fcntl.LOCK_SH if shared else fcntl.LOCK_EX,
        'taking a %s lock' % ('shared' if shared else 'exclusive'))
    return self

  def read_lock(self, message='taking read lock'):
    """Take a read lock (shared), downgrading from write if required.

    Args:
      message: A description of what/why this lock is being taken.

    Returns:
      self, allowing it to be used as a `with` target.

    Raises:
      IOError if the operation fails in some way.
    """
    self._enforce_lock(fcntl.LOCK_SH, message)
    return self

  def write_lock(self, message='taking write lock'):
    """Take a write lock (exclusive), upgrading from read if required.

    Note that if the lock state is being upgraded from read to write,
    a deadlock potential exists- as such we *will* release the lock
    to work around it.  Any consuming code should not assume that
    transitioning from shared to exclusive means no one else has
    gotten at the critical resource in between for this reason.

    Args:
      message: A description of what/why this lock is being taken.

    Returns:
      self, allowing it to be used as a `with` target.

    Raises:
      IOError if the operation fails in some way.
    """
    self._enforce_lock(fcntl.LOCK_EX, message)
    return self

  def unlock(self):
    """Release any locks held.  Noop if no locks are held.

    Raises:
      IOError if the operation fails in some way.
    """
    if self._fd is not None:
      logging.debug('%s: lock is being released (%s).',
                    self.description, self.locktype)
      self.locking_mechanism(self._fd, fcntl.LOCK_UN)

  def __del__(self):
    # TODO(ferringb): Convert this to snakeoil.weakref.WeakRefFinalizer
    # if/when that rebasing occurs.
    self.close()

  def close(self):
    """Release the underlying lock and close the fd."""
    if self._fd is not None:
      self.unlock()
      os.close(self._fd)
      self._fd = None

  def _enter(self):
    # Force the fd to be opened via touching the property.
    # We do this to ensure that even if entering a context w/out a lock
    # held, we can do locking in that critical section if the code requests it.
    # pylint: disable=pointless-statement
    self.fd
    return self

  def _exit(self, exc_type, exc, exc_tb):
    try:
      self.unlock()
    finally:
      self.close()

  def IsLocked(self):
    """Return True if the lock is grabbed."""
    return bool(self._fd)


class FileLock(_Lock):
  """Use a specified file as a locking mechanism."""

  def __init__(self, path, description=None, verbose=True,
               locktype=LOCKF, world_writable=False, blocking=True,
               blocking_timeout=None):
    """Initializer for FileLock.

    Args:
      path: On disk pathway to lock.  Can be a directory or a file.
      description: A description for this lock- what is it protecting?
      verbose: Verbose logging?
      locktype: Type of lock to use (lockf or flock).
      world_writable: If true, the lock file will be created as root and be made
        writable to all users.
      blocking: If True, use a blocking lock.
      blocking_timeout: If not None, time is seconds to wait on blocking calls.
    """
    if description is None:
      description = 'lock %s' % (path,)
    _Lock.__init__(self, description=description, verbose=verbose,
                   locktype=locktype, blocking=blocking,
                   blocking_timeout=blocking_timeout)
    self.path = os.path.abspath(path)
    self.world_writable = world_writable

  def _GetFd(self):
    if self.world_writable:
      create = True
      try:
        create = stat.S_IMODE(os.stat(self.path).st_mode) != 0o666
      except OSError as e:
        if e.errno != errno.ENOENT:
          raise
      if create:
        osutils.SafeMakedirs(os.path.dirname(self.path), sudo=True)
        cros_build_lib.sudo_run(['touch', self.path], print_cmd=False)
        cros_build_lib.sudo_run(['chmod', '666', self.path], print_cmd=False)

    # If we're on py3.4 and this attribute is exposed, use it to close
    # the threading race between open and fcntl setting; this is
    # extremely paranoid code, but might as well.
    cloexec = getattr(os, 'O_CLOEXEC', 0)
    # There exist race conditions where the lock may be created by
    # root, thus denying subsequent accesses from others. To prevent
    # this, we create the lock with mode 0o666.
    try:
      value = os.umask(000)
      fd = os.open(self.path, os.W_OK|os.O_CREAT|cloexec, 0o666)
    finally:
      os.umask(value)
    return fd


class ProcessLock(_Lock):
  """Process level locking visible to parent/child only.

  This lock is basically a more robust version of what
  multiprocessing.Lock does.  That implementation uses semaphores
  internally which require cleanup/deallocation code to run to release
  the lock; a SIGKILL hitting the process holding the lock violates those
  assumptions leading to a stuck lock.

  Thus this implementation is based around locking of a deleted tempfile;
  lockf locks are guranteed to be released once the process/fd is closed.
  """

  def _GetFd(self):
    with tempfile.TemporaryFile() as f:
      # We don't want to hold onto the object indefinitely; we just want
      # the fd to a temporary inode, preferably one that isn't vfs accessible.
      # Since TemporaryFile closes the fd once the object is GC'd, we just
      # dupe the fd so we retain a copy, while the original TemporaryFile
      # goes away.
      return os.dup(f.fileno())


class PortableLinkLock(object):
  """A more primitive lock that relies on the atomicity of creating hardlinks.

  Use this lock if you need to be compatible with shadow utils like groupadd
  or useradd.
  """

  def __init__(self, path, max_retry=0, sleep=1):
    """Construct an instance.

    Args:
      path: path to file to lock on.  Multiple processes attempting to lock the
        same path will compete for a system wide lock.
      max_retry: maximum number of times to attempt to acquire the lock.
      sleep: See retry_util.GenericRetry's sleep parameter.
    """
    self._path = path
    self._target_path = None
    # These two poorly named variables are just passed straight through to
    # retry_util.RetryException.
    self._max_retry = max_retry
    self._sleep = sleep

  def __enter__(self):
    fd, self._target_path = tempfile.mkstemp(
        prefix=self._path + '.chromite.portablelock.')
    os.close(fd)
    try:
      retry_util.RetryException(OSError, self._max_retry,
                                os.link, self._target_path, self._path,
                                sleep=self._sleep)
    except OSError:
      raise LockNotAcquiredError('Timeout while trying to lock %s' % self._path)
    finally:
      osutils.SafeUnlink(self._target_path)

    return self

  def __exit__(self, exc_type, exc_val, exc_tb):
    try:
      if self._target_path:
        osutils.SafeUnlink(self._target_path)
    finally:
      osutils.SafeUnlink(self._path)


class PipeLock(object):
  """A simple one-way lock based on pipe().

  This is used when code is calling os.fork() directly and needs to synchronize
  behavior between the two.  The same process should not try to use Wait/Post
  as it will just see its own results.  If you need bidirection locks, you'll
  need to create two yourself.

  Be sure to delete the lock when you're done to prevent fd leakage.
  """

  def __init__(self):
    # TODO(vapier): Simplify this when we're Python 3 only.
    # pylint: disable=using-constant-test
    pipe2 = getattr(os, 'pipe2', None)
    if pipe2:
      cloexec = getattr(os, 'O_CLOEXEC', 0)
      # Pylint-1.7 is unable to handle this conditional logic.
      # pylint: disable=not-callable
      pipes = pipe2(cloexec)
    else:
      pipes = os.pipe()
    self.read_fd, self.write_fd = pipes

  def Wait(self, size=1):
    """Read |size| bytes from the pipe.

    Args:
      size: How many bytes to read.  It must match the length of |data| passed
        by the other end during its call to Post.

    Returns:
      The data read back.
    """
    return os.read(self.read_fd, size)

  def Post(self, data=b'!'):
    """Write |data| to the pipe.

    Args:
      data: The data to send to the other side calling Wait.  It must be of the
        exact length that is passed to Wait.
    """
    os.write(self.write_fd, data)

  def __del__(self):
    os.close(self.read_fd)
    os.close(self.write_fd)