[Mailman-Developers] New LockFile.py
Barry Warsaw
bwarsaw@python.org
Mon, 1 May 2000 13:43:04 -0400 (EDT)
Folks,
Based on much input from Thomas and Harald, I've got a new version of
LockFile.py which should work much better for high volume sites. This
has the fixed linking direction, the implementation has been cleaned
up considerably, and I think I've eliminated race conditions as much
as possible.
One casualty is steal(). The current Mailman CVS source doesn't use
it, and I didn't like the semantics anyway, so I simply got rid of the
method. I don't know if any of the "unofficial" patches out there
still use it though.
The one questionable aspect of the new LockFile.py is that it uses the
mtime of the file as the lock's lifetime. This works great for
avoiding race conditions but means that Mailman is setting mtime's of
the files to some time in the future. Is anybody aware of OSes where
this will be a problem? Linux and Solaris both seem to stay happy.
I've added some unit testing which gives me confidence in this module,
but I haven't tried it yet on python.org. And I'm still waiting for
Thomas to give it his usual bulletproofing test. Still, I think it's
time for other eyes to take a look. I'd appreciate any feedback you
folks can provide.
-Barry
-------------------- snip snip --------------------
# Copyright (C) 1998,1999,2000 by the Free Software Foundation, Inc.
#
# This program is free software; you can redistribute it and/or
# modify it under the terms of the GNU General Public License
# as published by the Free Software Foundation; either version 2
# of the License, or (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program; if not, write to the Free Software
# Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
"""Portable, NFS-safe file locking with timeouts.
This code implements an NFS-safe file-based locking algorithm influenced by
the GNU/Linux open(2) manpage, under the description of the O_EXCL option.
From RH6.1:
[...] O_EXCL is broken on NFS file systems, programs which rely on it
for performing locking tasks will contain a race condition. The
solution for performing atomic file locking using a lockfile is to
create a unique file on the same fs (e.g., incorporating hostname and
pid), use link(2) to make a link to the lockfile. If link() returns
0, the lock is successful. Otherwise, use stat(2) on the unique file
to check if its link count has increased to 2, in which case the lock
is also successful.
The assumption made here is that there will be no `outside interference',
e.g. no agent external to this code will have access to link() to the affected
lock files.
LockFile objects support lock-breaking so that you can't wedge a process
forever. This is especially helpful in a web environment, but may not be
appropriate for all applications.
Locks have a `lifetime', which is the maximum length of time the process
expects to retain the lock. It is important to pick a good number here
because other processes will not break an existing lock until the expected
lifetime has expired. Too long and other processes will hang; too short and
you'll end up trampling on existing process locks -- and possibly corrupting
data. In a distributed (NFS) environment, you also need to make sure that
your clocks are properly synchronized.
Locks can also log their state to a log file. When running under Mailman, the
log file is placed in a Mailman-specific location, otherwise, the log file is
called `LockFile.log' and placed in the temp directory (calculated from
tempfile.mktemp()).
"""
# This code has undergone several revisions, with contributions from Barry
# Warsaw, Thomas Wouters, Harald Meland, and John Viega. It should also work
# well outside of Mailman so it could be used for other Python projects
# requiring file locking. See the __main__ section at the bottom of the file
# for unit testing.
import os
import socket
import time
import errno
import random
from stat import ST_NLINK, ST_MTIME
# Units are floating-point seconds.
DEFAULT_LOCK_LIFETIME = 15
# Allowable a bit of clock skew
CLOCK_SLOP = 10
# Figure out what logfile to use. This is different depending on whether
# we're running in a Mailman context or not.
_logfile = None
def _create_logfile():
try:
from Mailman.Logging.StampedLogger import StampedLogger
return StampedLogger('locks')
except ImportError:
# not running inside Mailman
import tempfile
dir = os.path.split(tempfile.mktemp())[0]
path = os.path.join(dir, 'LockFile.log')
# open in line-buffered mode
class SimpleUserFile:
def __init__(self, path):
self.__fp = open(path, 'a', 1)
self.__prefix = '(%d) ' % os.getpid()
def write(self, msg):
now = '%.3f' % time.time()
self.__fp.write(self.__prefix + now + ' ' + msg)
return SimpleUserFile(path)
# Exceptions that can be raised by this module
class LockError(Exception):
"""Base class for all exceptions in this module."""
class AlreadyLockedError(LockError):
"""An attempt is made to lock an already locked object."""
class NotLockedError(LockError):
"""An attempt is made to unlock an object that isn't locked."""
class TimeOutError(LockError):
"""The timeout interval elapsed before the lock succeeded."""
class LockFile:
"""A portable way to lock resources by way of the file system.
This class supports the following methods:
__init__(lockfile[, lifetime[, withlogging]]):
Create the resource lock using lockfile as the global lock file. Each
process laying claim to this resource lock will create their own
temporary lock files based on the path specified by lockfile.
Optional lifetime is the number of seconds the process expects to hold
the lock. Optional withlogging, when true, turns on lockfile logging
(see the module docstring for details).
set_lifetime(lifetime):
Set a new lock lifetime. This takes affect the next time the file is
locked, but does not refresh a locked file.
get_lifetime():
Return the lock's lifetime.
refresh([newlifetime[, unconditionally]]):
Refreshes the lifetime of a locked file. Use this if you realize that
you need to keep a resource locked longer than you thought. With
optional newlifetime, set the lock's lifetime. Raises NotLockedError
if the lock is not set, unless optional unconditionally flag is set to
true.
lock([timeout]):
Acquire the lock. This blocks until the lock is acquired unless
optional timeout is greater than 0, in which case, a TimeOutError is
raised when timeout number of seconds (or possibly more) expires
without lock acquisition. Raises AlreadyLockedError if the lock is
already set.
unlock([unconditionally]):
Relinquishes the lock. Raises a NotLockedError if the lock is not
set, unless optional unconditionally is true.
locked():
Return 1 if the lock is set, otherwise 0. To avoid race conditions,
this refreshes the lock (on set locks).
"""
def __init__(self, lockfile,
lifetime=DEFAULT_LOCK_LIFETIME,
withlogging=0):
"""Create the resource lock using lockfile as the global lock file.
Each process laying claim to this resource lock will create their own
temporary lock files based on the path specified by lockfile.
Optional lifetime is the number of seconds the process expects to hold
the lock. Optional withlogging, when true, turns on lockfile logging
(see the module docstring for details).
"""
self.__lockfile = lockfile
self.__lifetime = lifetime
self.__tmpfname = '%s.%s.%d' % (
lockfile, socket.gethostname(), os.getpid())
self.__withlogging = withlogging
self.__logprefix = os.path.split(self.__lockfile)[1]
def set_lifetime(self, lifetime):
"""Set a new lock lifetime.
This takes affect the next time the file is locked, but does not
refresh a locked file.
"""
self.__lifetime = lifetime
def get_lifetime(self):
"""Return the lock's lifetime."""
return self.__lifetime
def refresh(self, newlifetime=None, unconditionally=0):
"""Refreshes the lifetime of a locked file.
Use this if you realize that you need to keep a resource locked longer
than you thought. With optional newlifetime, set the lock's lifetime.
Raises NotLockedError if the lock is not set, unless optional
unconditionally flag is set to true.
"""
if newlifetime is not None:
self.set_lifetime(newlifetime)
# Do we have the lock? As a side effect, this refreshes the lock!
if not self.locked() and not unconditionally:
raise NotLockedError
def lock(self, timeout=0):
"""Acquire the lock.
This blocks until the lock is acquired unless optional timeout is
greater than 0, in which case, a TimeOutError is raised when timeout
number of seconds (or possibly more) expires without lock acquisition.
Raises AlreadyLockedError if the lock is already set.
"""
if timeout:
timeout_time = time.time() + timeout
# Make sure my temp lockfile exists, and that its contents are
# up-to-date (e.g. the temp file name, and the lock lifetime).
self.__write()
self.__touch()
self.__writelog('laying claim')
# for quieting the logging output
loopcount = -1
while 1:
loopcount = loopcount + 1
# Create the hard link and test for exactly 2 links to the file
try:
os.link(self.__tmpfname, self.__lockfile)
# If we got here, we know we know we got the lock, and never
# had it before, so we're done. Just touch it again for the
# fun of it.
self.__writelog('got the lock')
self.__touch()
break
except OSError, e:
# The link failed for some reason, possibly because someone
# else already has the lock (i.e. we got an EEXIST), or for
# some other bizarre reason.
if e.errno == errno.ENOENT:
# TBD: in some Linux environments, it is possible to get
# an ENOENT, which is truly strange, because this means
# that self.__tmpfname doesn't exist at the time of the
# os.link(), but self.__write() is supposed to guarantee
# that this happens! I don't honestly know why this
# happens, but for now we just say we didn't acquire the
# lock, and try again next time.
pass
elif e.errno <> errno.EEXIST:
# Something very bizarre happened. Clean up our state and
# pass the error on up.
self.__writelog('unexpected link error: %s' % e)
os.unlink(self.__tmpfname)
raise
elif self.__linkcount() <> 2:
# Somebody's messin' with us! Log this, and try again
# later. TBD: should we raise an exception?
self.__writelog('unexpected linkcount <> 2: %d' %
self.__linkcount())
elif self.__read() == self.__tmpfname:
# It was us that already had the link.
self.__writelog('already locked')
os.unlink(self.__tmpfname)
raise AlreadyLockedError
# otherwise, someone else has the lock
pass
# We did not acquire the lock, because someone else already has
# it. Have we timed out in our quest for the lock?
if timeout and timeout_time < time.time():
os.unlink(self.__tmpfname)
self.__writelog('timed out')
raise TimeOutError
# Okay, we haven't timed out, but we didn't get the lock. Let's
# find if the lock lifetime has expired.
if time.time() > self.__releasetime() + CLOCK_SLOP:
# Yes, so break the lock.
self.__break()
self.__writelog('lifetime has expired, breaking')
# Okay, someone else has the lock, our claim hasn't timed out yet,
# and the expected lock lifetime hasn't expired yet. So let's
# wait a while for the owner of the lock to give it up.
elif not loopcount % 100:
self.__writelog('waiting for claim')
self.__sleep()
def unlock(self, unconditionally=0):
"""Unlock the lock.
If we don't already own the lock (either because of unbalanced unlock
calls, or because the lock was stolen out from under us), raise a
NotLockedError, unless optional `unconditional' is true.
"""
islocked = 0
try:
islocked = self.locked()
except OSError, e:
if e.errno <> errno.ENOENT: raise
if not islocked and not unconditionally:
raise NotLockedError
# Remove our tempfile
try:
os.unlink(self.__tmpfname)
except OSError, e:
if e.errno <> errno.ENOENT: raise
# If we owned the lock, remove the global file, relinquishing it.
if islocked:
try:
os.unlink(self.__lockfile)
except OSError, e:
if e.errno <> errno.ENOENT: raise
self.__writelog('unlocked')
def locked(self):
"""Returns 1 if we own the lock, 0 if we do not.
Checking the status of the lockfile resets the lock's lifetime, which
helps avoid race conditions during the lock status test.
"""
# Discourage breaking the lock for a while.
self.__touch()
# TBD: can the link count ever be > 2?
if self.__linkcount() <> 2:
return 0
return self.__read() == self.__tmpfname
def finalize(self):
self.unlock(unconditionally=1)
def __del__(self):
self.finalize()
#
# Private interface
#
def __writelog(self, msg):
global _logfile
if self.__withlogging:
if not _logfile:
_logfile = _create_logfile()
_logfile.write('%s %s\n' % (self.__logprefix, msg))
def __write(self):
# Make sure it's group writable
oldmask = os.umask(002)
try:
fp = open(self.__tmpfname, 'w')
fp.write(self.__tmpfname)
fp.close()
finally:
os.umask(oldmask)
def __read(self):
try:
fp = open(self.__lockfile)
filename = fp.read()
fp.close()
return filename
except (OSError, IOError), e:
if e.errno <> errno.ENOENT: raise
return None
def __touch(self, filename=None):
t = time.time() + self.__lifetime
try:
# TBD: We probably don't need to modify atime, but this is easier.
os.utime(filename or self.__tmpfname, (t, t))
except OSError, e:
if e.errno <> errno.ENOENT: raise
def __releasetime(self):
try:
return os.stat(self.__lockfile)[ST_MTIME]
except OSError, e:
if e.errno <> errno.ENOENT: raise
return -1
def __linkcount(self):
try:
return os.stat(self.__lockfile)[ST_NLINK]
except OSError, e:
if e.errno <> errno.ENOENT: raise
return -1
def __break(self):
# First, touch the global lock file so no other process will try to
# break the lock while we're doing it. Specifically, this avoids the
# race condition where we've decided to break the lock at the same
# time someone else has, but between the time we made this decision
# and the time we read the winner out of the global lock file, they've
# gone ahead and claimed the lock.
self.__touch(self.__lockfile)
# Try to remove the old winner's temp file, since we're assuming the
# winner process has hung or died. Don't worry too much if we can't
# unlink their temp file -- this doesn't break the locking algorithm,
# but will leave temp file turds laying around, a minor inconvenience.
try:
winner = self.__read()
if winner:
os.unlink(winner)
except OSError, e:
if e.errno <> errno.ENOENT: raise
# Now remove the global lockfile, which actually breaks the lock.
try:
os.unlink(self.__lockfile)
except OSError, e:
if e.errno <> errno.ENOENT: raise
def __sleep(self):
interval = random.random() * 2.0 + 0.01
time.sleep(interval)
# Unit test framework
def _dochild():
prefix = '[%d]' % os.getpid()
# Create somewhere between 1 and 1000 locks
lockfile = LockFile('/tmp/LockTest', withlogging=1, lifetime=120)
# Use a lock lifetime of between 1 and 15 seconds. Under normal
# situations, Mailman's usage patterns (untested) shouldn't be much longer
# than this.
workinterval = 5 * random.random()
hitwait = 20 * random.random()
print prefix, 'workinterval:', workinterval
islocked = 0
t0 = 0
t1 = 0
t2 = 0
try:
try:
t0 = time.time()
print prefix, 'acquiring...'
lockfile.lock()
print prefix, 'acquired...'
islocked = 1
except TimeOutError:
print prefix, 'timed out'
else:
t1 = time.time()
print prefix, 'acquisition time:', t1-t0, 'seconds'
time.sleep(workinterval)
finally:
if islocked:
try:
lockfile.unlock()
t2 = time.time()
print prefix, 'lock hold time:', t2-t1, 'seconds'
except NotLockedError:
print prefix, 'lock was broken'
# wait for next web hit
print prefix, 'webhit sleep:', hitwait
time.sleep(hitwait)
def _onetest():
loopcount = random.randint(1, 100)
for i in range(loopcount):
pid = os.fork()
if pid:
# parent, wait for child to exit
pid, status = os.waitpid(pid, 0)
else:
# child
try:
_dochild()
except KeyboardInterrupt:
pass
os._exit(0)
def _reap(kids):
if not kids:
return
pid, status = os.waitpid(-1, os.WNOHANG)
if pid <> 0:
del kids[pid]
def _test(numtests):
kids = {}
for i in range(numtests):
pid = os.fork()
if pid:
# parent
kids[pid] = pid
else:
# child
try:
import sha
random.seed(sha.new(`os.getpid()`+`time.time()`).hexdigest())
_onetest()
except KeyboardInterrupt:
pass
os._exit(0)
# slightly randomize each kid's seed
while kids:
_reap(kids)
if __name__ == '__main__':
import sys
import random
_test(int(sys.argv[1]))