Commit 723ddd40 authored by Jim Fulton's avatar Jim Fulton

Refactored cache verification to fix threading bugs during connection.

Changed connections to work with unset (None) clients.  Messages
aren't forwarded until the client is set.  This is to prevent sending
spurious invalidation messages until a client is ready to recieve them.
parent 1d61dc56
......@@ -491,6 +491,7 @@ class ClientStorage(object):
# If we are upgrading from a read-only fallback connection,
# we must close the old connection to prevent it from being
# used while the cache is verified against the new connection.
self._connection.register_object(None) # Don't call me!
self._connection.close()
self._connection = None
self._ready.clear()
......@@ -558,54 +559,6 @@ class ClientStorage(object):
else:
return '%s:%s' % (self._storage, self._server_addr)
def verify_cache(self, server):
"""Internal routine called to verify the cache.
The return value (indicating which path we took) is used by
the test suite.
"""
# If verify_cache() finishes the cache verification process,
# it should set self._server. If it goes through full cache
# verification, then endVerify() should self._server.
last_inval_tid = self._cache.getLastTid()
if last_inval_tid is not None:
ltid = server.lastTransaction()
if ltid == last_inval_tid:
log2("No verification necessary (last_inval_tid up-to-date)")
self._server = server
self._ready.set()
return "no verification"
# log some hints about last transaction
log2("last inval tid: %r %s\n"
% (last_inval_tid, tid2time(last_inval_tid)))
log2("last transaction: %r %s" %
(ltid, ltid and tid2time(ltid)))
pair = server.getInvalidations(last_inval_tid)
if pair is not None:
log2("Recovering %d invalidations" % len(pair[1]))
self.invalidateTransaction(*pair)
self._server = server
self._ready.set()
return "quick verification"
log2("Verifying cache")
# setup tempfile to hold zeoVerify results
self._tfile = tempfile.TemporaryFile(suffix=".inv")
self._pickler = cPickle.Pickler(self._tfile, 1)
self._pickler.fast = 1 # Don't use the memo
# TODO: should batch these operations for efficiency; would need
# to acquire lock ...
for oid, tid, version in self._cache.contents():
server.verify(oid, version, tid)
self._pending_server = server
server.endZeoVerify()
return "full verification"
### Is there a race condition between notifyConnected and
### notifyDisconnected? In Particular, what if we get
### notifyDisconnected in the middle of notifyConnected?
......@@ -1162,7 +1115,7 @@ class ClientStorage(object):
return
for oid, version, data in self._tbuf:
self._cache.invalidate(oid, version, tid)
self._cache.invalidate(oid, version, tid, False)
# If data is None, we just invalidate.
if data is not None:
s = self._seriald[oid]
......@@ -1224,8 +1177,6 @@ class ClientStorage(object):
"""Storage API: return a sequence of versions in the storage."""
return self._server.versions(max)
# Below are methods invoked by the StorageServer
def serialnos(self, args):
"""Server callback to pass a list of changed (oid, serial) pairs."""
self._serials.extend(args)
......@@ -1234,6 +1185,57 @@ class ClientStorage(object):
"""Server callback to update the info dictionary."""
self._info.update(dict)
def verify_cache(self, server):
"""Internal routine called to verify the cache.
The return value (indicating which path we took) is used by
the test suite.
"""
self._pending_server = server
# setup tempfile to hold zeoVerify results and interim
# invalidation results
self._tfile = tempfile.TemporaryFile(suffix=".inv")
self._pickler = cPickle.Pickler(self._tfile, 1)
self._pickler.fast = 1 # Don't use the memo
# allow incoming invalidations:
self._connection.register_object(self)
# If verify_cache() finishes the cache verification process,
# it should set self._server. If it goes through full cache
# verification, then endVerify() should self._server.
last_inval_tid = self._cache.getLastTid()
if last_inval_tid is not None:
ltid = server.lastTransaction()
if ltid == last_inval_tid:
log2("No verification necessary (last_inval_tid up-to-date)")
self.finish_verification()
return "no verification"
# log some hints about last transaction
log2("last inval tid: %r %s\n"
% (last_inval_tid, tid2time(last_inval_tid)))
log2("last transaction: %r %s" %
(ltid, ltid and tid2time(ltid)))
pair = server.getInvalidations(last_inval_tid)
if pair is not None:
log2("Recovering %d invalidations" % len(pair[1]))
self.finish_verification(pair)
return "quick verification"
log2("Verifying cache")
# TODO: should batch these operations for efficiency; would need
# to acquire lock ...
for oid, tid, version in self._cache.contents():
server.verify(oid, version, tid)
server.endZeoVerify()
return "full verification"
def invalidateVerify(self, args):
"""Server callback to invalidate an (oid, version) pair.
......@@ -1245,67 +1247,92 @@ class ClientStorage(object):
# This should never happen. TODO: assert it doesn't, or log
# if it does.
return
self._pickler.dump(args)
oid, version = args
self._pickler.dump((oid, version, None))
def _process_invalidations(self, invs):
# Invalidations are sent by the ZEO server as a sequence of
# oid, version pairs. The DB's invalidate() method expects a
# dictionary of oids.
def endVerify(self):
"""Server callback to signal end of cache validation."""
log2("endVerify finishing")
self.finish_verification()
log2("endVerify finished")
def finish_verification(self, catch_up=None):
self._lock.acquire()
try:
# versions maps version names to dictionary of invalidations
versions = {}
for oid, version, tid in invs:
if oid == self._load_oid:
self._load_status = 0
self._cache.invalidate(oid, version, tid)
oids = versions.get((version, tid))
if not oids:
versions[(version, tid)] = [oid]
else:
oids.append(oid)
if self._db is not None:
for (version, tid), d in versions.items():
self._db.invalidate(tid, d, version=version)
finally:
self._lock.release()
if catch_up:
# process catch-up invalidations
tid, invalidations = catch_up
self._process_invalidations(
(oid, version, tid)
for oid, version in invalidations
)
def endVerify(self):
"""Server callback to signal end of cache validation."""
if self._pickler is None:
return
# write end-of-data marker
self._pickler.dump((None, None))
self._pickler.dump((None, None, None))
self._pickler = None
self._tfile.seek(0)
f = self._tfile
unpickler = cPickle.Unpickler(self._tfile)
min_tid = self._cache.getLastTid()
def InvalidationLogIterator():
while 1:
oid, version, tid = unpickler.load()
if oid is None:
break
if ((tid is None)
or (min_tid is None)
or (tid > min_tid)
):
yield oid, version, tid
self._process_invalidations(InvalidationLogIterator())
self._tfile.close()
self._tfile = None
self._process_invalidations(InvalidationLogIterator(f))
f.close()
finally:
self._lock.release()
log2("endVerify finishing")
self._server = self._pending_server
self._ready.set()
self._pending_conn = None
log2("endVerify finished")
self._pending_server = None
def invalidateTransaction(self, tid, args):
"""Invalidate objects modified by tid."""
"""Server callback: Invalidate objects modified by tid."""
self._lock.acquire()
try:
self._cache.setLastTid(tid)
finally:
self._lock.release()
if self._pickler is not None:
log2("Transactional invalidation during cache verification",
level=BLATHER)
for t in args:
self._pickler.dump(t)
for oid, version in args:
self._pickler.dump((oid, version, tid))
return
self._process_invalidations([(oid, version, tid)
for oid, version in args])
finally:
self._lock.release()
def _process_invalidations(self, invs):
# Invalidations are sent by the ZEO server as a sequence of
# oid, version, tid triples. The DB's invalidate() method expects a
# dictionary of oids.
# versions maps version names to dictionary of invalidations
versions = {}
for oid, version, tid in invs:
if oid == self._load_oid:
self._load_status = 0
self._cache.invalidate(oid, version, tid)
oids = versions.get((version, tid))
if not oids:
versions[(version, tid)] = [oid]
else:
oids.append(oid)
if self._db is not None:
for (version, tid), d in versions.items():
self._db.invalidate(tid, d, version=version)
# The following are for compatibility with protocol version 2.0.0
......@@ -1315,11 +1342,3 @@ class ClientStorage(object):
invalidate = invalidateVerify
end = endVerify
Invalidate = invalidateTrans
def InvalidationLogIterator(fileobj):
unpickler = cPickle.Unpickler(fileobj)
while 1:
oid, version = unpickler.load()
if oid is None:
break
yield oid, version, None
Invalidations while connecting
==============================
As soon as a client registers with a server, it will recieve
invalidations from the server. The client must be careful to queue
these invalidations until it is ready to deal with them. At the time
of the writing of this test, clients weren't careful enogh about
queing invalidations. This led to cache corruption in the form of
both low-level file corruption as well as out-of-date records marked
as current.
This tests tries to provoke this bug by:
- starting a server
>>> import ZEO.tests.testZEO, ZEO.tests.forker
>>> addr = 'localhost', ZEO.tests.testZEO.get_port()
>>> zconf = ZEO.tests.forker.ZEOConfig(addr)
>>> sconf = '<filestorage 1>\npath Data.fs\n</filestorage>\n'
>>> _, adminaddr, pid, conf_path = ZEO.tests.forker.start_zeo_server(
... sconf, zconf, addr[1])
- opening a client to the server that writes some objects, filling
it's cache at the same time,
>>> import ZEO.ClientStorage, ZODB.tests.MinPO, transaction
>>> db = ZODB.DB(ZEO.ClientStorage.ClientStorage(addr, client='x'))
>>> conn = db.open()
>>> nobs = 1000
>>> for i in range(nobs):
... conn.root()[i] = ZODB.tests.MinPO.MinPO(0)
>>> transaction.commit()
- disconnecting the first client (closing it with a persistent cache),
>>> db.close()
- starting a second client that writes objects more or less
constantly,
>>> import random, threading
>>> stop = False
>>> db2 = ZODB.DB(ZEO.ClientStorage.ClientStorage(addr))
>>> tm = transaction.TransactionManager()
>>> conn2 = db2.open(transaction_manager=tm)
>>> random = random.Random(0)
>>> lock = threading.Lock()
>>> def run():
... while 1:
... i = random.randint(0, nobs-1)
... if stop:
... return
... lock.acquire()
... try:
... conn2.root()[i].value += 1
... tm.commit()
... finally:
... lock.release()
... time.sleep(0)
>>> thread = threading.Thread(target=run)
>>> thread.start()
- restarting the first client, and
- testing for cache validity.
>>> import zope.testing.loggingsupport, logging
>>> handler = zope.testing.loggingsupport.InstalledHandler(
... 'ZEO', level=logging.ERROR)
>>> import time
>>> for c in range(10):
... time.sleep(.1)
... db = ZODB.DB(ZEO.ClientStorage.ClientStorage(addr, client='x'))
... _ = lock.acquire()
... try:
... time.sleep(.1)
... assert (db._storage.lastTransaction()
... == db._storage._server.lastTransaction()), (
... db._storage.lastTransaction(),
... db._storage._server.lastTransactiion())
... conn = db.open()
... for i in range(1000):
... if conn.root()[i].value != conn2.root()[i].value:
... print 'bad', c, i, conn.root()[i].value,
... print conn2.root()[i].value
... finally:
... _ = lock.release()
... db.close()
>>> stop = True
>>> thread.join(10)
>>> thread.isAlive()
False
>>> for record in handler.records:
... print record.name, record.levelname
... print handler.format(record)
>>> handler.uninstall()
>>> db.close()
>>> db2.close()
......@@ -21,7 +21,7 @@ platform-dependent scaffolding.
import unittest
# Import the actual test class
from ZEO.tests import ConnectionTests, InvalidationTests
from zope.testing import doctest, setupstack
class FileStorageConfig:
def getConfig(self, path, create, read_only):
......@@ -135,6 +135,10 @@ def test_suite():
for klass in test_classes:
sub = unittest.makeSuite(klass, 'check')
suite.addTest(sub)
suite.addTest(doctest.DocFileSuite(
'invalidations_while_connecting.test',
setUp=setupstack.setUpDirectory, tearDown=setupstack.tearDown,
))
return suite
......
......@@ -447,8 +447,7 @@ class ConnectWrapper:
Call the client's testConnection(), giving the client a chance
to do app-level check of the connection.
"""
self.conn = ManagedClientConnection(self.sock, self.addr,
self.client, self.mgr)
self.conn = ManagedClientConnection(self.sock, self.addr, self.mgr)
self.sock = None # The socket is now owned by the connection
try:
self.preferred = self.client.testConnection(self.conn)
......
......@@ -555,14 +555,23 @@ class Connection(smac.SizedMessageAsyncConnection, object):
self.replies_cond.release()
def handle_request(self, msgid, flags, name, args):
if not self.check_method(name):
msg = "Invalid method name: %s on %s" % (name, repr(self.obj))
obj = self.obj
if name.startswith('_') or not hasattr(obj, name):
if obj is None:
if __debug__:
self.log("no object calling %s%s"
% (name, short_repr(args)),
level=logging.DEBUG)
return
msg = "Invalid method name: %s on %s" % (name, repr(obj))
raise ZRPCError(msg)
if __debug__:
self.log("calling %s%s" % (name, short_repr(args)),
level=logging.DEBUG)
meth = getattr(self.obj, name)
meth = getattr(obj, name)
try:
self.waiting_for_reply = True
try:
......@@ -601,12 +610,6 @@ class Connection(smac.SizedMessageAsyncConnection, object):
level=logging.ERROR, exc_info=True)
self.close()
def check_method(self, name):
# TODO: This is hardly "secure".
if name.startswith('_'):
return None
return hasattr(self.obj, name)
def send_reply(self, msgid, ret):
# encode() can pass on a wide variety of exceptions from cPickle.
# While a bare `except` is generally poor practice, in this case
......@@ -897,7 +900,7 @@ class ManagedClientConnection(Connection):
__super_close = Connection.close
base_message_output = Connection.message_output
def __init__(self, sock, addr, obj, mgr):
def __init__(self, sock, addr, mgr):
self.mgr = mgr
# We can't use the base smac's message_output directly because the
......@@ -914,7 +917,7 @@ class ManagedClientConnection(Connection):
self.queue_output = True
self.queued_messages = []
self.__super_init(sock, addr, obj, tag='C', map=client_map)
self.__super_init(sock, addr, None, tag='C', map=client_map)
self.thr_async = True
self.trigger = client_trigger
client_trigger.pull_trigger()
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment