def write_pickle(obj, dest, tmp=None, pickle_protocol=0):
"""
Ensure that a pickle file gets written to disk. The file is first written
to a tmp file location in the destination directory path, ensured it is
synced to disk, then moved to its final destination name.
This version takes advantage of Gluster's dot-prefix-dot-suffix naming
where the a file named ".thefile.name.9a7aasv" is hashed to the same
Gluster node as "thefile.name". This ensures the renaming of a temp file
once written does not move it to another Gluster node.
:param obj: python object to be pickled
:param dest: path of final destination file
:param tmp: path to tmp to use, defaults to None (ignored)
:param pickle_protocol: protocol to pickle the obj with, defaults to 0
"""
dirname = os.path.dirname(dest)
basename = os.path.basename(dest)
tmpname = '.' + basename + '.' + \
md5(basename + str(random.random())).hexdigest()
tmppath = os.path.join(dirname, tmpname)
with open(tmppath, 'wb') as fo:
pickle.dump(obj, fo, pickle_protocol)
# TODO: This flush() method call turns into a flush() system call
# We'll need to wrap this as well, but we would do this by writing
#a context manager for our own open() method which returns an object
# in fo which makes the gluster API call.
fo.flush()
do_fsync(fo)
do_rename(tmppath, dest)
def test_do_rename_err(self):
try:
srcpath = os.path.join('/tmp', str(random.random()))
destpath = os.path.join('/tmp', str(random.random()))
fs.do_rename(srcpath, destpath)
except GlusterFileSystemOSError:
pass
else:
self.fail("GlusterFileSystemOSError expected")
def test_do_rename(self):
srcpath = mkdtemp()
try:
destpath = os.path.join('/tmp', str(random.random()))
fs.do_rename(srcpath, destpath)
assert not os.path.exists(srcpath)
assert os.path.exists(destpath)
finally:
os.rmdir(destpath)
def write_pickle(obj, dest, tmp=None, pickle_protocol=0):
"""
Ensure that a pickle file gets written to disk. The file is first written
to a tmp file location in the destination directory path, ensured it is
synced to disk, then moved to its final destination name.
This version takes advantage of Gluster's dot-prefix-dot-suffix naming
where the a file named ".thefile.name.9a7aasv" is hashed to the same
Gluster node as "thefile.name". This ensures the renaming of a temp file
once written does not move it to another Gluster node.
:param obj: python object to be pickled
:param dest: path of final destination file
:param tmp: path to tmp to use, defaults to None (ignored)
:param pickle_protocol: protocol to pickle the obj with, defaults to 0
"""
dirname = os.path.dirname(dest)
# Create destination directory
try:
os.makedirs(dirname)
except OSError as err:
if err.errno != errno.EEXIST:
raise
basename = os.path.basename(dest)
tmpname = '.' + basename + '.' + \
md5(basename + str(random.random())).hexdigest()
tmppath = os.path.join(dirname, tmpname)
with open(tmppath, 'wb') as fo:
pickle.dump(obj, fo, pickle_protocol)
# TODO: This flush() method call turns into a flush() system call
# We'll need to wrap this as well, but we would do this by writing
# a context manager for our own open() method which returns an object
# in fo which makes the gluster API call.
fo.flush()
do_fsync(fo)
do_rename(tmppath, dest)
def _finalize_put(self, metadata):
# Write out metadata before fsync() to ensure it is also forced to
# disk.
write_metadata(self._fd, metadata)
# We call fsync() before calling drop_cache() to lower the
# amount of redundant work the drop cache code will perform on
# the pages (now that after fsync the pages will be all
# clean).
do_fsync(self._fd)
# From the Department of the Redundancy Department, make sure
# we call drop_cache() after fsync() to avoid redundant work
# (pages all clean).
do_fadvise64(self._fd, self._last_sync, self._upload_size)
# At this point we know that the object's full directory path
# exists, so we can just rename it directly without using Swift's
# swift.common.utils.renamer(), which makes the directory path and
# adds extra stat() calls.
df = self._disk_file
attempts = 1
while True:
try:
do_rename(self._tmppath, df._data_file)
except OSError as err:
if err.errno in (errno.ENOENT, errno.EIO) \
and attempts < MAX_RENAME_ATTEMPTS:
# FIXME: Why either of these two error conditions is
# happening is unknown at this point. This might be a
# FUSE issue of some sort or a possible race
# condition. So let's sleep on it, and double check
# the environment after a good nap.
_random_sleep()
# Tease out why this error occurred. The man page for
# rename reads:
# "The link named by tmppath does not exist; or, a
# directory component in data_file does not exist;
# or, tmppath or data_file is an empty string."
assert len(self._tmppath) > 0 and len(df._data_file) > 0
tpstats = do_stat(self._tmppath)
tfstats = do_fstat(self._fd)
assert tfstats
if not tpstats or tfstats.st_ino != tpstats.st_ino:
# Temporary file name conflict
raise DiskFileError(
'DiskFile.put(): temporary file, %s, was'
' already renamed (targeted for %s)' % (
self._tmppath, df._data_file))
else:
# Data file target name now has a bad path!
dfstats = do_stat(df._put_datadir)
if not dfstats:
raise DiskFileError(
'DiskFile.put(): path to object, %s, no'
' longer exists (targeted for %s)' % (
df._put_datadir, df._data_file))
else:
is_dir = stat.S_ISDIR(dfstats.st_mode)
if not is_dir:
raise DiskFileError(
'DiskFile.put(): path to object, %s,'
' no longer a directory (targeted for'
' %s)' % (self._put_datadir,
df._data_file))
else:
# Let's retry since everything looks okay
logging.warn(
"DiskFile.put(): os.rename('%s','%s')"
" initially failed (%s) but a"
" stat('%s') following that succeeded:"
" %r" % (
self._tmppath, df._data_file, str(err),
df._put_datadir, dfstats))
attempts += 1
continue
else:
raise GlusterFileSystemOSError(
err.errno, "%s, os.rename('%s', '%s')" % (
err.strerror, self._tmppath, df._data_file))
else:
# Success!
break
# Close here so the calling context does not have to perform this
# in a thread.
self.close()
def _finalize_put(self, metadata):
# Write out metadata before fsync() to ensure it is also forced to
# disk.
write_metadata(self._fd, metadata)
# We call fsync() before calling drop_cache() to lower the
# amount of redundant work the drop cache code will perform on
# the pages (now that after fsync the pages will be all
# clean).
do_fsync(self._fd)
# From the Department of the Redundancy Department, make sure
# we call drop_cache() after fsync() to avoid redundant work
# (pages all clean).
do_fadvise64(self._fd, self._last_sync, self._upload_size)
# At this point we know that the object's full directory path
# exists, so we can just rename it directly without using Swift's
# swift.common.utils.renamer(), which makes the directory path and
# adds extra stat() calls.
df = self._disk_file
attempts = 1
while True:
try:
do_rename(self._tmppath, df._data_file)
except OSError as err:
if err.errno in (errno.ENOENT, errno.EIO) \
and attempts < MAX_RENAME_ATTEMPTS:
# FIXME: Why either of these two error conditions is
# happening is unknown at this point. This might be a
# FUSE issue of some sort or a possible race
# condition. So let's sleep on it, and double check
# the environment after a good nap.
_random_sleep()
# Tease out why this error occurred. The man page for
# rename reads:
# "The link named by tmppath does not exist; or, a
# directory component in data_file does not exist;
# or, tmppath or data_file is an empty string."
assert len(self._tmppath) > 0 and len(df._data_file) > 0
tpstats = do_stat(self._tmppath)
tfstats = do_fstat(self._fd)
assert tfstats
if not tpstats or tfstats.st_ino != tpstats.st_ino:
# Temporary file name conflict
raise DiskFileError(
'DiskFile.put(): temporary file, %s, was'
' already renamed (targeted for %s)' %
(self._tmppath, df._data_file))
else:
# Data file target name now has a bad path!
dfstats = do_stat(df._put_datadir)
if not dfstats:
raise DiskFileError(
'DiskFile.put(): path to object, %s, no'
' longer exists (targeted for %s)' %
(df._put_datadir, df._data_file))
else:
is_dir = stat.S_ISDIR(dfstats.st_mode)
if not is_dir:
raise DiskFileError(
'DiskFile.put(): path to object, %s,'
' no longer a directory (targeted for'
' %s)' %
(self._put_datadir, df._data_file))
else:
# Let's retry since everything looks okay
logging.warn(
"DiskFile.put(): rename('%s','%s')"
" initially failed (%s) but a"
" stat('%s') following that succeeded:"
" %r" %
(self._tmppath, df._data_file, str(err),
df._put_datadir, dfstats))
attempts += 1
continue
else:
raise GlusterFileSystemOSError(
err.errno, "%s, rename('%s', '%s')" %
(err.strerror, self._tmppath, df._data_file))
else:
# Success!
break
# Close here so the calling context does not have to perform this
# in a thread.
self.close()