openstack/swift
Code Issues Proposed changes
Files
2fe18b24cd7e709c8712fcda7369b0715a21fb89
swift/swift/obj/ssync_receiver.py
Alistair Coles 2fe18b24cd ssync: fix decoding of ts_meta when ts_data has offset 
The SsyncSender encodes object file timestamps in a compact form and
the SsyncReceiver decodes the timestamps and compares them to its
object file set.

The encoding represents the meta file timestamp as a delta from the
data file timestamp, NOT INCLUDING the data file timestamp offset.

Previously, the decoding was erroneously calculating the meta file
timestamp as the sum of the delta plus the data file timestamp
INCLUDING the offset.

For example, if the SssyncSender has object file timestamps:

  ts_data = t0_1.data
  ts_meta = t1.data

then the receiver would erroneously perceive that the sender has:

  ts_data = t0_1.data
  ts_meta = t1_1.data

As described in the referenced bug report, this erroneous decoding
could cause the SsyncReceiver to request that the SsyncSender sync an
object that is already in sync, which results in a 409 Conflict at the
receiver. The 409 causes the ssync session to terminate, and the same
process repeats on the next attempt.

Closes-Bug: #2007643
Co-Authored-By: Clay Gerrard <clay.gerrard@gmail.com>
Change-Id: I74a0aac0ac29577026743f87f4b654d85e8fcc80
2023-02-27 07:27:32 -06:00

556 lines
24 KiB
Python
Raw Blame History

# Copyright (c) 2013 OpenStack Foundation
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
# implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import eventlet.greenio
import eventlet.wsgi
from eventlet import sleep
from six.moves import urllib
from swift.common import exceptions
from swift.common import http
from swift.common import swob
from swift.common import utils
from swift.common import request_helpers
from swift.common.utils import Timestamp
class SsyncClientDisconnected(Exception):
pass
def decode_missing(line):
"""
Parse a string of the form generated by
:py:func:`~swift.obj.ssync_sender.encode_missing` and return a dict
with keys ``object_hash``, ``ts_data``, ``ts_meta``, ``ts_ctype``,
``durable``.
The encoder for this line is
:py:func:`~swift.obj.ssync_sender.encode_missing`
"""
result = {}
parts = line.decode('ascii').split()
result['object_hash'] = urllib.parse.unquote(parts[0])
t_data = urllib.parse.unquote(parts[1])
result['ts_data'] = ts_data = Timestamp(t_data)
result['ts_meta'] = result['ts_ctype'] = ts_data
result['durable'] = True # default to True in case this key isn't sent
if len(parts) > 2:
# allow for a comma separated list of k:v pairs to future-proof
subparts = urllib.parse.unquote(parts[2]).split(',')
for item in [subpart for subpart in subparts if ':' in subpart]:
k, v = item.split(':')
if k == 'm':
# ignore ts_data offset when calculating ts_meta
result['ts_meta'] = Timestamp(ts_data.normal,
delta=int(v, 16))
elif k == 't':
# ignore ts_data offset when calculating ts_ctype
result['ts_ctype'] = Timestamp(ts_data.normal,
delta=int(v, 16))
elif k == 'durable':
result['durable'] = utils.config_true_value(v)
return result
def encode_wanted(remote, local):
"""
Compare a remote and local results and generate a wanted line.
:param remote: a dict, with ts_data and ts_meta keys in the form
returned by :py:func:`decode_missing`
:param local: a dict, possibly empty, with ts_data and ts_meta keys
in the form returned :py:meth:`Receiver._check_local`
The decoder for this line is
:py:func:`~swift.obj.ssync_sender.decode_wanted`
"""
want = {}
if 'ts_data' in local:
# we have something, let's get just the right stuff
if remote['ts_data'] > local['ts_data']:
want['data'] = True
if 'ts_meta' in local and remote['ts_meta'] > local['ts_meta']:
want['meta'] = True
if ('ts_ctype' in local and remote['ts_ctype'] > local['ts_ctype']
and remote['ts_ctype'] > remote['ts_data']):
want['meta'] = True
else:
# we got nothing, so we'll take whatever the remote has
want['data'] = True
want['meta'] = True
if want:
# this is the inverse of _decode_wanted's key_map
key_map = dict(data='d', meta='m')
parts = ''.join(v for k, v in sorted(key_map.items()) if want.get(k))
return '%s %s' % (urllib.parse.quote(remote['object_hash']), parts)
return None
class Receiver(object):
"""
Handles incoming SSYNC requests to the object server.
These requests come from the object-replicator daemon that uses
:py:mod:`.ssync_sender`.
The number of concurrent SSYNC requests is restricted by
use of a replication_semaphore and can be configured with the
object-server.conf [object-server] replication_concurrency
setting.
An SSYNC request is really just an HTTP conduit for
sender/receiver replication communication. The overall
SSYNC request should always succeed, but it will contain
multiple requests within its request and response bodies. This
"hack" is done so that replication concurrency can be managed.
The general process inside an SSYNC request is:
1. Initialize the request: Basic request validation, mount check,
acquire semaphore lock, etc..
2. Missing check: Sender sends the hashes and timestamps of
the object information it can send, receiver sends back
the hashes it wants (doesn't have or has an older
timestamp).
3. Updates: Sender sends the object information requested.
4. Close down: Release semaphore lock, etc.
"""
def __init__(self, app, request):
self.app = app
self.request = request
self.device = None
self.partition = None
self.fp = None
# We default to dropping the connection in case there is any exception
# raised during processing because otherwise the sender could send for
# quite some time before realizing it was all in vain.
self.disconnect = True
self.initialize_request()
def __call__(self):
"""
Processes an SSYNC request.
Acquires a semaphore lock and then proceeds through the steps
of the SSYNC process.
"""
# The general theme for functions __call__ calls is that they should
# raise exceptions.MessageTimeout for client timeouts (logged locally),
# exceptions.ChunkReadError for client disconnects (logged locally),
# swob.HTTPException classes for exceptions to return to the caller but
# not log locally (unmounted, for example), and any other Exceptions
# will be logged with a full stack trace.
# This is because the client is never just some random user but
# is instead also our code and we definitely want to know if our code
# is broken or doing something unexpected.
try:
# Double try blocks in case our main error handlers fail.
try:
# Need to send something to trigger wsgi to return response
# headers and kick off the ssync exchange.
yield b'\r\n'
# If semaphore is in use, try to acquire it, non-blocking, and
# return a 503 if it fails.
if self.app.replication_semaphore:
if not self.app.replication_semaphore.acquire(False):
raise swob.HTTPServiceUnavailable()
try:
with self.diskfile_mgr.replication_lock(self.device,
self.policy,
self.partition):
for data in self.missing_check():
yield data
for data in self.updates():
yield data
# We didn't raise an exception, so end the request
# normally.
self.disconnect = False
finally:
if self.app.replication_semaphore:
self.app.replication_semaphore.release()
except SsyncClientDisconnected:
self.app.logger.error('ssync client disconnected')
self.disconnect = True
except exceptions.LockTimeout as err:
self.app.logger.debug(
'%s/%s/%s SSYNC LOCK TIMEOUT: %s' % (
self.request.remote_addr, self.device, self.partition,
err))
yield (':ERROR: %d %r\n' % (0, str(err))).encode('utf8')
except exceptions.MessageTimeout as err:
self.app.logger.error(
'%s/%s/%s TIMEOUT in ssync.Receiver: %s' % (
self.request.remote_addr, self.device, self.partition,
err))
yield (':ERROR: %d %r\n' % (408, str(err))).encode('utf8')
except exceptions.ChunkReadError as err:
self.app.logger.error(
'%s/%s/%s read failed in ssync.Receiver: %s' % (
self.request.remote_addr, self.device, self.partition,
err))
except swob.HTTPException as err:
body = b''.join(err({}, lambda *args: None))
yield (':ERROR: %d %r\n' % (
err.status_int, body)).encode('utf8')
except Exception as err:
self.app.logger.exception(
'%s/%s/%s EXCEPTION in ssync.Receiver' %
(self.request.remote_addr, self.device, self.partition))
yield (':ERROR: %d %r\n' % (0, str(err))).encode('utf8')
except Exception:
self.app.logger.exception('EXCEPTION in ssync.Receiver')
if self.disconnect:
# This makes the socket close early so the remote side doesn't have
# to send its whole request while the lower Eventlet-level just
# reads it and throws it away. Instead, the connection is dropped
# and the remote side will get a broken-pipe exception.
try:
socket = self.request.environ['wsgi.input'].get_socket()
eventlet.greenio.shutdown_safe(socket)
socket.close()
except Exception:
pass # We're okay with the above failing.
def initialize_request(self):
"""
Basic validation of request and mount check.
This function will be called before attempting to acquire a
replication semaphore lock, so contains only quick checks.
"""
# This environ override has been supported since eventlet 0.14:
# https://bitbucket.org/eventlet/eventlet/commits/ \
# 4bd654205a4217970a57a7c4802fed7ff2c8b770
self.request.environ['eventlet.minimum_write_chunk_size'] = 0
self.device, self.partition, self.policy = \
request_helpers.get_name_and_placement(self.request, 2, 2, False)
self.frag_index = None
if self.request.headers.get('X-Backend-Ssync-Frag-Index'):
try:
self.frag_index = int(
self.request.headers['X-Backend-Ssync-Frag-Index'])
except ValueError:
raise swob.HTTPBadRequest(
'Invalid X-Backend-Ssync-Frag-Index %r' %
self.request.headers['X-Backend-Ssync-Frag-Index'])
utils.validate_device_partition(self.device, self.partition)
self.diskfile_mgr = self.app._diskfile_router[self.policy]
if not self.diskfile_mgr.get_dev_path(self.device):
raise swob.HTTPInsufficientStorage(drive=self.device)
self.fp = self.request.environ['wsgi.input']
def _readline(self, context):
# try to read a line from the wsgi input; annotate any timeout or read
# errors with a description of the calling context
with exceptions.MessageTimeout(
self.app.client_timeout, context):
try:
line = self.fp.readline(self.app.network_chunk_size)
except (eventlet.wsgi.ChunkReadError, IOError) as err:
raise exceptions.ChunkReadError('%s: %s' % (context, err))
return line
def _check_local(self, remote, make_durable=True):
"""
Parse local diskfile and return results of current
representative for comparison to remote.
:param object_hash: the hash of the remote object being offered
"""
try:
df = self.diskfile_mgr.get_diskfile_from_hash(
self.device, self.partition, remote['object_hash'],
self.policy, frag_index=self.frag_index, open_expired=True)
except exceptions.DiskFileNotExist:
return {}
try:
df.open()
except exceptions.DiskFileDeleted as err:
result = {'ts_data': err.timestamp}
except exceptions.DiskFileError:
# e.g. a non-durable EC frag
result = {}
else:
result = {
'ts_data': df.data_timestamp,
'ts_meta': df.timestamp,
'ts_ctype': df.content_type_timestamp,
}
if ((df.durable_timestamp is None or
df.durable_timestamp < remote['ts_data']) and
df.fragments and
remote['ts_data'] in df.fragments and
self.frag_index in df.fragments[remote['ts_data']]):
# The remote is offering a fragment that we already have but is
# *newer* than anything *durable* that we have
if remote['durable']:
# We have the frag, just missing durable state, so make the
# frag durable now. Try this just once to avoid looping if
# it fails.
if make_durable:
try:
with df.create() as writer:
writer.commit(remote['ts_data'])
return self._check_local(remote, make_durable=False)
except Exception:
# if commit fails then log exception and fall back to
# wanting a full update
self.app.logger.exception(
'%s/%s/%s EXCEPTION in ssync.Receiver while '
'attempting commit of %s'
% (self.request.remote_addr, self.device,
self.partition, df._datadir))
else:
# We have the non-durable frag that is on offer, but our
# ts_data may currently be set to an older durable frag, so
# bump our ts_data to prevent the remote frag being wanted.
result['ts_data'] = remote['ts_data']
return result
def _check_missing(self, line):
"""
Parse offered object from sender, and compare to local diskfile,
responding with proper protocol line to represented needed data
or None if in sync.
Anchor point for tests to mock legacy protocol changes.
"""
remote = decode_missing(line)
local = self._check_local(remote)
return encode_wanted(remote, local)
def missing_check(self):
"""
Handles the receiver-side of the MISSING_CHECK step of a
SSYNC request.
Receives a list of hashes and timestamps of object
information the sender can provide and responds with a list
of hashes desired, either because they're missing or have an
older timestamp locally.
The process is generally:
1. Sender sends ``:MISSING_CHECK: START`` and begins
sending `hash timestamp` lines.
2. Receiver gets ``:MISSING_CHECK: START`` and begins
reading the `hash timestamp` lines, collecting the
hashes of those it desires.
3. Sender sends ``:MISSING_CHECK: END``.
4. Receiver gets ``:MISSING_CHECK: END``, responds with
``:MISSING_CHECK: START``, followed by the list of
<wanted_hash> specifiers it collected as being wanted
(one per line), ``:MISSING_CHECK: END``, and flushes any
buffers.
Each <wanted_hash> specifier has the form <hash>[ <parts>] where
<parts> is a string containing characters 'd' and/or 'm'
indicating that only data or meta part of object respectively is
required to be sync'd.
5. Sender gets ``:MISSING_CHECK: START`` and reads the list
of hashes desired by the receiver until reading
``:MISSING_CHECK: END``.
The collection and then response is so the sender doesn't
have to read while it writes to ensure network buffers don't
fill up and block everything.
"""
line = self._readline('missing_check start')
if not line:
# Guess they hung up
raise SsyncClientDisconnected
if line.strip() != b':MISSING_CHECK: START':
raise Exception(
'Looking for :MISSING_CHECK: START got %r' % line[:1024])
object_hashes = []
nlines = 0
while True:
line = self._readline('missing_check line')
if not line or line.strip() == b':MISSING_CHECK: END':
break
want = self._check_missing(line)
if want:
object_hashes.append(want)
if nlines % 5 == 0:
sleep() # Gives a chance for other greenthreads to run
nlines += 1
yield b':MISSING_CHECK: START\r\n'
if object_hashes:
yield b'\r\n'.join(hsh.encode('ascii') for hsh in object_hashes)
yield b'\r\n'
yield b':MISSING_CHECK: END\r\n'
def updates(self):
"""
Handles the UPDATES step of an SSYNC request.
Receives a set of PUT and DELETE subrequests that will be
routed to the object server itself for processing. These
contain the information requested by the MISSING_CHECK step.
The PUT and DELETE subrequests are formatted pretty much
exactly like regular HTTP requests, excepting the HTTP
version on the first request line.
The process is generally:
1. Sender sends ``:UPDATES: START`` and begins sending the
PUT and DELETE subrequests.
2. Receiver gets ``:UPDATES: START`` and begins routing the
subrequests to the object server.
3. Sender sends ``:UPDATES: END``.
4. Receiver gets ``:UPDATES: END`` and sends ``:UPDATES:
START`` and ``:UPDATES: END`` (assuming no errors).
5. Sender gets ``:UPDATES: START`` and ``:UPDATES: END``.
If too many subrequests fail, as configured by
replication_failure_threshold and replication_failure_ratio,
the receiver will hang up the request early so as to not
waste any more time.
At step 4, the receiver will send back an error if there were
any failures (that didn't cause a hangup due to the above
thresholds) so the sender knows the whole was not entirely a
success. This is so the sender knows if it can remove an out
of place partition, for example.
"""
line = self._readline('updates start')
if not line:
# Guess they hung up waiting for us to process the missing check
raise SsyncClientDisconnected
if line.strip() != b':UPDATES: START':
raise Exception('Looking for :UPDATES: START got %r' % line[:1024])
successes = 0
failures = 0
updates = 0
while True:
line = self._readline('updates line')
if not line or line.strip() == b':UPDATES: END':
break
# Read first line METHOD PATH of subrequest.
method, path = swob.bytes_to_wsgi(line.strip()).split(' ', 1)
subreq = swob.Request.blank(
'/%s/%s%s' % (self.device, self.partition, path),
environ={'REQUEST_METHOD': method})
# Read header lines.
content_length = None
replication_headers = []
while True:
line = self._readline('updates line')
if not line:
raise Exception(
'Got no headers for %s %s' % (method, path))
line = line.strip()
if not line:
break
header, value = swob.bytes_to_wsgi(line).split(':', 1)
header = header.strip().lower()
value = value.strip()
subreq.headers[header] = value
if header not in ('etag', 'x-backend-no-commit'):
# we'll use X-Backend-Replication-Headers to force the
# object server to write all sync'd metadata, but with some
# exceptions:
# - make sure ssync doesn't cause 'Etag' to be added to
# obj metadata in addition to 'ETag' which object server
# sets (note capitalization)
# - filter out x-backend-no-commit which ssync sender may
# have added to the subrequest
replication_headers.append(header)
if header == 'content-length':
content_length = int(value)
# Establish subrequest body, if needed.
if method in ('DELETE', 'POST'):
if content_length not in (None, 0):
raise Exception(
'%s subrequest with content-length %s'
% (method, path))
elif method == 'PUT':
if content_length is None:
raise Exception(
'No content-length sent for %s %s' % (method, path))
def subreq_iter():
left = content_length
while left > 0:
with exceptions.MessageTimeout(
self.app.client_timeout,
'updates content'):
chunk = self.fp.read(
min(left, self.app.network_chunk_size))
if not chunk:
raise exceptions.ChunkReadError(
'Early termination for %s %s' % (method, path))
left -= len(chunk)
yield chunk
subreq.environ['wsgi.input'] = utils.FileLikeIter(
subreq_iter())
else:
raise Exception('Invalid subrequest method %s' % method)
subreq.headers['X-Backend-Storage-Policy-Index'] = int(self.policy)
subreq.headers['X-Backend-Replication'] = 'True'
if self.frag_index is not None:
# primary node should not 409 if it has a non-primary fragment
subreq.headers['X-Backend-Ssync-Frag-Index'] = self.frag_index
if replication_headers:
subreq.headers['X-Backend-Replication-Headers'] = \
' '.join(replication_headers)
# Route subrequest and translate response.
resp = subreq.get_response(self.app)
if http.is_success(resp.status_int) or \
resp.status_int == http.HTTP_NOT_FOUND:
successes += 1
else:
self.app.logger.warning(
'ssync subrequest failed with %s: %s %s (%s)' %
(resp.status_int, method, subreq.path, resp.body))
failures += 1
if failures >= self.app.replication_failure_threshold and (
not successes or
float(failures) / successes >
self.app.replication_failure_ratio):
raise Exception(
'Too many %d failures to %d successes' %
(failures, successes))
# The subreq may have failed, but we want to read the rest of the
# body from the remote side so we can continue on with the next
# subreq.
for junk in subreq.environ['wsgi.input']:
pass
if updates % 5 == 0:
sleep() # Gives a chance for other greenthreads to run
updates += 1
if failures:
raise swob.HTTPInternalServerError(
'ERROR: With :UPDATES: %d failures to %d successes' %
(failures, successes))
yield b':UPDATES: START\r\n'
yield b':UPDATES: END\r\n'
View Git Blame Copy Permalink