import sys, time from twisted.internet import defer from twisted.python import failure from foolscap.eventual import eventually from allmydata.util import base32, hashutil, idlib, log from allmydata import storage from pycryptopp.publickey import rsa from common import MODE_CHECK, MODE_ANYTHING, MODE_WRITE, MODE_READ, \ DictOfSets, CorruptShareError, NeedMoreDataError from layout import unpack_prefix_and_signature, unpack_header, unpack_share class ServerMap: """I record the placement of mutable shares. This object records which shares (of various versions) are located on which servers. One purpose I serve is to inform callers about which versions of the mutable file are recoverable and 'current'. A second purpose is to serve as a state marker for test-and-set operations. I am passed out of retrieval operations and back into publish operations, which means 'publish this new version, but only if nothing has changed since I last retrieved this data'. This reduces the chances of clobbering a simultaneous (uncoordinated) write. @ivar servermap: a dictionary, mapping a (peerid, shnum) tuple to a (versionid, timestamp) tuple. Each 'versionid' is a tuple of (seqnum, root_hash, IV, segsize, datalength, k, N, signed_prefix, offsets) @ivar connections: maps peerid to a RemoteReference @ivar bad_shares: a sequence of (peerid, shnum) tuples, describing shares that I should ignore (because a previous user of the servermap determined that they were invalid). The updater only locates a certain number of shares: if some of these turn out to have integrity problems and are unusable, the caller will need to mark those shares as bad, then re-update the servermap, then try again. """ def __init__(self): self.servermap = {} self.connections = {} self.unreachable_peers = set() # peerids that didn't respond to queries self.problems = [] # mostly for debugging self.bad_shares = set() self.last_update_mode = None self.last_update_time = 0 def mark_bad_share(self, peerid, shnum): """This share was found to be bad, not in the checkstring or signature, but deeper in the share, detected at retrieve time. Remove it from our list of useful shares, and remember that it is bad so we don't add it back again later. """ key = (peerid, shnum) self.bad_shares.add(key) self.servermap.pop(key, None) def add_new_share(self, peerid, shnum, verinfo, timestamp): """We've written a new share out, replacing any that was there before.""" key = (peerid, shnum) self.bad_shares.discard(key) self.servermap[key] = (verinfo, timestamp) def dump(self, out=sys.stdout): print >>out, "servermap:" for ( (peerid, shnum), (verinfo, timestamp) ) in self.servermap.items(): (seqnum, root_hash, IV, segsize, datalength, k, N, prefix, offsets_tuple) = verinfo print >>out, ("[%s]: sh#%d seq%d-%s %d-of-%d len%d" % (idlib.shortnodeid_b2a(peerid), shnum, seqnum, base32.b2a(root_hash)[:4], k, N, datalength)) return out def all_peers(self): return set([peerid for (peerid, shnum) in self.servermap]) def make_versionmap(self): """Return a dict that maps versionid to sets of (shnum, peerid, timestamp) tuples.""" versionmap = DictOfSets() for ( (peerid, shnum), (verinfo, timestamp) ) in self.servermap.items(): versionmap.add(verinfo, (shnum, peerid, timestamp)) return versionmap def shares_on_peer(self, peerid): return set([shnum for (s_peerid, shnum) in self.servermap if s_peerid == peerid]) def version_on_peer(self, peerid, shnum): key = (peerid, shnum) if key in self.servermap: (verinfo, timestamp) = self.servermap[key] return verinfo return None return None def shares_available(self): """Return a dict that maps verinfo to tuples of (num_distinct_shares, k) tuples.""" versionmap = self.make_versionmap() all_shares = {} for verinfo, shares in versionmap.items(): s = set() for (shnum, peerid, timestamp) in shares: s.add(shnum) (seqnum, root_hash, IV, segsize, datalength, k, N, prefix, offsets_tuple) = verinfo all_shares[verinfo] = (len(s), k) return all_shares def highest_seqnum(self): available = self.shares_available() seqnums = [verinfo[0] for verinfo in available.keys()] seqnums.append(0) return max(seqnums) def summarize_versions(self): """Return a string describing which versions we know about.""" versionmap = self.make_versionmap() bits = [] for (verinfo, shares) in versionmap.items(): (seqnum, root_hash, IV, segsize, datalength, k, N, prefix, offsets_tuple) = verinfo shnums = set([shnum for (shnum, peerid, timestamp) in shares]) bits.append("%d*seq%d-%s" % (len(shnums), seqnum, base32.b2a(root_hash)[:4])) return "/".join(bits) def recoverable_versions(self): """Return a set of versionids, one for each version that is currently recoverable.""" versionmap = self.make_versionmap() recoverable_versions = set() for (verinfo, shares) in versionmap.items(): (seqnum, root_hash, IV, segsize, datalength, k, N, prefix, offsets_tuple) = verinfo shnums = set([shnum for (shnum, peerid, timestamp) in shares]) if len(shnums) >= k: # this one is recoverable recoverable_versions.add(verinfo) return recoverable_versions def unrecoverable_versions(self): """Return a set of versionids, one for each version that is currently unrecoverable.""" versionmap = self.make_versionmap() unrecoverable_versions = set() for (verinfo, shares) in versionmap.items(): (seqnum, root_hash, IV, segsize, datalength, k, N, prefix, offsets_tuple) = verinfo shnums = set([shnum for (shnum, peerid, timestamp) in shares]) if len(shnums) < k: unrecoverable_versions.add(verinfo) return unrecoverable_versions def best_recoverable_version(self): """Return a single versionid, for the so-called 'best' recoverable version. Sequence number is the primary sort criteria, followed by root hash. Returns None if there are no recoverable versions.""" recoverable = list(self.recoverable_versions()) recoverable.sort() if recoverable: return recoverable[-1] return None def unrecoverable_newer_versions(self): # Return a dict of versionid -> health, for versions that are # unrecoverable and have later seqnums than any recoverable versions. # These indicate that a write will lose data. pass def needs_merge(self): # return True if there are multiple recoverable versions with the # same seqnum, meaning that MutableFileNode.read_best_version is not # giving you the whole story, and that using its data to do a # subsequent publish will lose information. pass class ServermapUpdater: def __init__(self, filenode, servermap, mode=MODE_READ): """I update a servermap, locating a sufficient number of useful shares and remembering where they are located. """ self._node = filenode self._servermap = servermap self.mode = mode self._running = True self._storage_index = filenode.get_storage_index() self._last_failure = None # how much data should we read? # * if we only need the checkstring, then [0:75] # * if we need to validate the checkstring sig, then [543ish:799ish] # * if we need the verification key, then [107:436ish] # * the offset table at [75:107] tells us about the 'ish' # * if we need the encrypted private key, we want [-1216ish:] # * but we can't read from negative offsets # * the offset table tells us the 'ish', also the positive offset # A future version of the SMDF slot format should consider using # fixed-size slots so we can retrieve less data. For now, we'll just # read 2000 bytes, which also happens to read enough actual data to # pre-fetch a 9-entry dirnode. self._read_size = 2000 if mode == MODE_CHECK: # we use unpack_prefix_and_signature, so we need 1k self._read_size = 1000 self._need_privkey = False if mode == MODE_WRITE and not self._node._privkey: self._need_privkey = True prefix = storage.si_b2a(self._storage_index)[:5] self._log_number = log.msg(format="SharemapUpdater(%(si)s): starting (%(mode)s)", si=prefix, mode=mode) def log(self, *args, **kwargs): if "parent" not in kwargs: kwargs["parent"] = self._log_number return log.msg(*args, **kwargs) def update(self): """Update the servermap to reflect current conditions. Returns a Deferred that fires with the servermap once the update has finished.""" # self._valid_versions is a set of validated verinfo tuples. We just # use it to remember which versions had valid signatures, so we can # avoid re-checking the signatures for each share. self._valid_versions = set() # self.versionmap maps verinfo tuples to sets of (shnum, peerid, # timestamp) tuples. This is used to figure out which versions might # be retrievable, and to make the eventual data download faster. self.versionmap = DictOfSets() self._started = time.time() self._done_deferred = defer.Deferred() # first, which peers should be talk to? Any that were in our old # servermap, plus "enough" others. self._queries_completed = 0 client = self._node._client full_peerlist = client.get_permuted_peers("storage", self._node._storage_index) self.full_peerlist = full_peerlist # for use later, immutable self.extra_peers = full_peerlist[:] # peers are removed as we use them self._good_peers = set() # peers who had some shares self._empty_peers = set() # peers who don't have any shares self._bad_peers = set() # peers to whom our queries failed k = self._node.get_required_shares() if k is None: # make a guess k = 3 N = self._node.get_required_shares() if N is None: N = 10 self.EPSILON = k # we want to send queries to at least this many peers (although we # might not wait for all of their answers to come back) self.num_peers_to_query = k + self.EPSILON if self.mode == MODE_CHECK: initial_peers_to_query = dict(full_peerlist) must_query = set(initial_peers_to_query.keys()) self.extra_peers = [] elif self.mode == MODE_WRITE: # we're planning to replace all the shares, so we want a good # chance of finding them all. We will keep searching until we've # seen epsilon that don't have a share. self.num_peers_to_query = N + self.EPSILON initial_peers_to_query, must_query = self._build_initial_querylist() self.required_num_empty_peers = self.EPSILON # TODO: arrange to read lots of data from k-ish servers, to avoid # the extra round trip required to read large directories. This # might also avoid the round trip required to read the encrypted # private key. else: initial_peers_to_query, must_query = self._build_initial_querylist() # this is a set of peers that we are required to get responses from: # they are peers who used to have a share, so we need to know where # they currently stand, even if that means we have to wait for a # silently-lost TCP connection to time out. We remove peers from this # set as we get responses. self._must_query = must_query # now initial_peers_to_query contains the peers that we should ask, # self.must_query contains the peers that we must have heard from # before we can consider ourselves finished, and self.extra_peers # contains the overflow (peers that we should tap if we don't get # enough responses) self._send_initial_requests(initial_peers_to_query) return self._done_deferred def _build_initial_querylist(self): initial_peers_to_query = {} must_query = set() for peerid in self._servermap.all_peers(): ss = self._servermap.connections[peerid] # we send queries to everyone who was already in the sharemap initial_peers_to_query[peerid] = ss # and we must wait for responses from them must_query.add(peerid) while ((self.num_peers_to_query > len(initial_peers_to_query)) and self.extra_peers): (peerid, ss) = self.extra_peers.pop(0) initial_peers_to_query[peerid] = ss return initial_peers_to_query, must_query def _send_initial_requests(self, peerlist): self._queries_outstanding = set() self._sharemap = DictOfSets() # shnum -> [(peerid, seqnum, R)..] dl = [] for (peerid, ss) in peerlist.items(): self._queries_outstanding.add(peerid) self._do_query(ss, peerid, self._storage_index, self._read_size) # control flow beyond this point: state machine. Receiving responses # from queries is the input. We might send out more queries, or we # might produce a result. return None def _do_query(self, ss, peerid, storage_index, readsize): self.log(format="sending query to [%(peerid)s], readsize=%(readsize)d", peerid=idlib.shortnodeid_b2a(peerid), readsize=readsize, level=log.NOISY) self._servermap.connections[peerid] = ss started = time.time() self._queries_outstanding.add(peerid) d = self._do_read(ss, peerid, storage_index, [], [(0, readsize)]) d.addCallback(self._got_results, peerid, readsize, (ss, storage_index), started) d.addErrback(self._query_failed, peerid) # errors that aren't handled by _query_failed (and errors caused by # _query_failed) get logged, but we still want to check for doneness. d.addErrback(log.err) d.addBoth(self._check_for_done) d.addErrback(self._fatal_error) return d def _do_read(self, ss, peerid, storage_index, shnums, readv): d = ss.callRemote("slot_readv", storage_index, shnums, readv) return d def _got_results(self, datavs, peerid, readsize, stuff, started): lp = self.log(format="got result from [%(peerid)s], %(numshares)d shares", peerid=idlib.shortnodeid_b2a(peerid), numshares=len(datavs), level=log.NOISY) self._queries_outstanding.discard(peerid) self._must_query.discard(peerid) self._queries_completed += 1 if not self._running: self.log("but we're not running, so we'll ignore it", parent=lp) return if datavs: self._good_peers.add(peerid) else: self._empty_peers.add(peerid) last_verinfo = None last_shnum = None for shnum,datav in datavs.items(): data = datav[0] try: verinfo = self._got_results_one_share(shnum, data, peerid) last_verinfo = verinfo last_shnum = shnum except CorruptShareError, e: # log it and give the other shares a chance to be processed f = failure.Failure() self.log("bad share: %s %s" % (f, f.value), parent=lp, level=log.WEIRD) self._bad_peers.add(peerid) self._last_failure = f self._servermap.problems.append(f) pass if self._need_privkey and last_verinfo: # send them a request for the privkey. We send one request per # server. (seqnum, root_hash, IV, segsize, datalength, k, N, prefix, offsets_tuple) = last_verinfo o = dict(offsets_tuple) self._queries_outstanding.add(peerid) readv = [ (o['enc_privkey'], (o['EOF'] - o['enc_privkey'])) ] ss = self._servermap.connections[peerid] d = self._do_read(ss, peerid, self._storage_index, [last_shnum], readv) d.addCallback(self._got_privkey_results, peerid, last_shnum) d.addErrback(self._privkey_query_failed, peerid, last_shnum) d.addErrback(log.err) d.addCallback(self._check_for_done) d.addErrback(self._fatal_error) # all done! self.log("_got_results done", parent=lp) def _got_results_one_share(self, shnum, data, peerid): lp = self.log(format="_got_results: got shnum #%(shnum)d from peerid %(peerid)s", shnum=shnum, peerid=idlib.shortnodeid_b2a(peerid)) # this might raise NeedMoreDataError, if the pubkey and signature # live at some weird offset. That shouldn't happen, so I'm going to # treat it as a bad share. (seqnum, root_hash, IV, k, N, segsize, datalength, pubkey_s, signature, prefix) = unpack_prefix_and_signature(data) if not self._node._pubkey: fingerprint = hashutil.ssk_pubkey_fingerprint_hash(pubkey_s) assert len(fingerprint) == 32 if fingerprint != self._node._fingerprint: raise CorruptShareError(peerid, shnum, "pubkey doesn't match fingerprint") self._node._pubkey = self._deserialize_pubkey(pubkey_s) if self._need_privkey: self._try_to_extract_privkey(data, peerid, shnum) (ig_version, ig_seqnum, ig_root_hash, ig_IV, ig_k, ig_N, ig_segsize, ig_datalen, offsets) = unpack_header(data) offsets_tuple = tuple( [(key,value) for key,value in offsets.items()] ) verinfo = (seqnum, root_hash, IV, segsize, datalength, k, N, prefix, offsets_tuple) if verinfo not in self._valid_versions: # it's a new pair. Verify the signature. valid = self._node._pubkey.verify(prefix, signature) if not valid: raise CorruptShareError(peerid, shnum, "signature is invalid") # ok, it's a valid verinfo. Add it to the list of validated # versions. self.log(" found valid version %d-%s from %s-sh%d: %d-%d/%d/%d" % (seqnum, base32.b2a(root_hash)[:4], idlib.shortnodeid_b2a(peerid), shnum, k, N, segsize, datalength), parent=lp) self._valid_versions.add(verinfo) # We now know that this is a valid candidate verinfo. if (peerid, shnum) in self._servermap.bad_shares: # we've been told that the rest of the data in this share is # unusable, so don't add it to the servermap. self.log("but we've been told this is a bad share", parent=lp, level=log.UNUSUAL) return verinfo # Add the info to our servermap. timestamp = time.time() self._servermap.add_new_share(peerid, shnum, verinfo, timestamp) # and the versionmap self.versionmap.add(verinfo, (shnum, peerid, timestamp)) return verinfo def _deserialize_pubkey(self, pubkey_s): verifier = rsa.create_verifying_key_from_string(pubkey_s) return verifier def _try_to_extract_privkey(self, data, peerid, shnum): try: r = unpack_share(data) except NeedMoreDataError, e: # this share won't help us. oh well. offset = e.encprivkey_offset length = e.encprivkey_length self.log("shnum %d on peerid %s: share was too short (%dB) " "to get the encprivkey; [%d:%d] ought to hold it" % (shnum, idlib.shortnodeid_b2a(peerid), len(data), offset, offset+length)) # NOTE: if uncoordinated writes are taking place, someone might # change the share (and most probably move the encprivkey) before # we get a chance to do one of these reads and fetch it. This # will cause us to see a NotEnoughSharesError(unable to fetch # privkey) instead of an UncoordinatedWriteError . This is a # nuisance, but it will go away when we move to DSA-based mutable # files (since the privkey will be small enough to fit in the # write cap). return (seqnum, root_hash, IV, k, N, segsize, datalen, pubkey, signature, share_hash_chain, block_hash_tree, share_data, enc_privkey) = r return self._try_to_validate_privkey(self, enc_privkey, peerid, shnum) def _try_to_validate_privkey(self, enc_privkey, peerid, shnum): alleged_privkey_s = self._node._decrypt_privkey(enc_privkey) alleged_writekey = hashutil.ssk_writekey_hash(alleged_privkey_s) if alleged_writekey != self._node.get_writekey(): self.log("invalid privkey from %s shnum %d" % (idlib.nodeid_b2a(peerid)[:8], shnum), level=log.WEIRD) return # it's good self.log("got valid privkey from shnum %d on peerid %s" % (shnum, idlib.shortnodeid_b2a(peerid))) privkey = rsa.create_signing_key_from_string(alleged_privkey_s) self._node._populate_encprivkey(enc_privkey) self._node._populate_privkey(privkey) self._need_privkey = False def _query_failed(self, f, peerid): self.log("error during query: %s %s" % (f, f.value), level=log.WEIRD) if not self._running: return self._must_query.discard(peerid) self._queries_outstanding.discard(peerid) self._bad_peers.add(peerid) self._servermap.problems.append(f) self._servermap.unreachable_peers.add(peerid) # TODO: overkill? self._queries_completed += 1 self._last_failure = f def _got_privkey_results(self, datavs, peerid, shnum): self._queries_outstanding.discard(peerid) if not self._need_privkey: return if shnum not in datavs: self.log("privkey wasn't there when we asked it", level=log.WEIRD) return datav = datavs[shnum] enc_privkey = datav[0] self._try_to_validate_privkey(enc_privkey, peerid, shnum) def _privkey_query_failed(self, f, peerid, shnum): self._queries_outstanding.discard(peerid) self.log("error during privkey query: %s %s" % (f, f.value), level=log.WEIRD) if not self._running: return self._queries_outstanding.discard(peerid) self._servermap.problems.append(f) self._last_failure = f def _check_for_done(self, res): # exit paths: # return self._send_more_queries(outstanding) : send some more queries # return self._done() : all done # return : keep waiting, no new queries lp = self.log(format=("_check_for_done, mode is '%(mode)s', " "%(outstanding)d queries outstanding, " "%(extra)d extra peers available, " "%(must)d 'must query' peers left" ), mode=self.mode, outstanding=len(self._queries_outstanding), extra=len(self.extra_peers), must=len(self._must_query), level=log.NOISY, ) if not self._running: self.log("but we're not running", parent=lp, level=log.NOISY) return if self._must_query: # we are still waiting for responses from peers that used to have # a share, so we must continue to wait. No additional queries are # required at this time. self.log("%d 'must query' peers left" % len(self._must_query), parent=lp) return if (not self._queries_outstanding and not self.extra_peers): # all queries have retired, and we have no peers left to ask. No # more progress can be made, therefore we are done. self.log("all queries are retired, no extra peers: done", parent=lp) return self._done() recoverable_versions = self._servermap.recoverable_versions() unrecoverable_versions = self._servermap.unrecoverable_versions() # what is our completion policy? how hard should we work? if self.mode == MODE_ANYTHING: if recoverable_versions: self.log("%d recoverable versions: done" % len(recoverable_versions), parent=lp) return self._done() if self.mode == MODE_CHECK: # we used self._must_query, and we know there aren't any # responses still waiting, so that means we must be done self.log("done", parent=lp) return self._done() MAX_IN_FLIGHT = 5 if self.mode == MODE_READ: # if we've queried k+epsilon servers, and we see a recoverable # version, and we haven't seen any unrecoverable higher-seqnum'ed # versions, then we're done. if self._queries_completed < self.num_peers_to_query: self.log(format="%(completed)d completed, %(query)d to query: need more", completed=self._queries_completed, query=self.num_peers_to_query, parent=lp) return self._send_more_queries(MAX_IN_FLIGHT) if not recoverable_versions: self.log("no recoverable versions: need more", parent=lp) return self._send_more_queries(MAX_IN_FLIGHT) highest_recoverable = max(recoverable_versions) highest_recoverable_seqnum = highest_recoverable[0] for unrec_verinfo in unrecoverable_versions: if unrec_verinfo[0] > highest_recoverable_seqnum: # there is evidence of a higher-seqnum version, but we # don't yet see enough shares to recover it. Try harder. # TODO: consider sending more queries. # TODO: consider limiting the search distance self.log("evidence of higher seqnum: need more") return self._send_more_queries(MAX_IN_FLIGHT) # all the unrecoverable versions were old or concurrent with a # recoverable version. Good enough. self.log("no higher-seqnum: done", parent=lp) return self._done() if self.mode == MODE_WRITE: # we want to keep querying until we've seen a few that don't have # any shares, to be sufficiently confident that we've seen all # the shares. This is still less work than MODE_CHECK, which asks # every server in the world. if not recoverable_versions: self.log("no recoverable versions: need more", parent=lp) return self._send_more_queries(MAX_IN_FLIGHT) last_found = -1 last_not_responded = -1 num_not_responded = 0 num_not_found = 0 states = [] found_boundary = False for i,(peerid,ss) in enumerate(self.full_peerlist): if peerid in self._bad_peers: # query failed states.append("x") #self.log("loop [%s]: x" % idlib.shortnodeid_b2a(peerid)) elif peerid in self._empty_peers: # no shares states.append("0") #self.log("loop [%s]: 0" % idlib.shortnodeid_b2a(peerid)) if last_found != -1: num_not_found += 1 if num_not_found >= self.EPSILON: self.log("found our boundary, %s" % "".join(states), parent=lp) found_boundary = True break elif peerid in self._good_peers: # yes shares states.append("1") #self.log("loop [%s]: 1" % idlib.shortnodeid_b2a(peerid)) last_found = i num_not_found = 0 else: # not responded yet states.append("?") #self.log("loop [%s]: ?" % idlib.shortnodeid_b2a(peerid)) last_not_responded = i num_not_responded += 1 if found_boundary: # we need to know that we've gotten answers from # everybody to the left of here if last_not_responded == -1: # we're done self.log("have all our answers", parent=lp) # .. unless we're still waiting on the privkey if self._need_privkey: self.log("but we're still waiting for the privkey", parent=lp) # if we found the boundary but we haven't yet found # the privkey, we may need to look further. If # somehow all the privkeys were corrupted (but the # shares were readable), then this is likely to do an # exhaustive search. return self._send_more_queries(MAX_IN_FLIGHT) return self._done() # still waiting for somebody return self._send_more_queries(num_not_responded) # if we hit here, we didn't find our boundary, so we're still # waiting for peers self.log("no boundary yet, %s" % "".join(states), parent=lp) return self._send_more_queries(MAX_IN_FLIGHT) # otherwise, keep up to 5 queries in flight. TODO: this is pretty # arbitrary, really I want this to be something like k - # max(known_version_sharecounts) + some extra self.log("catchall: need more", parent=lp) return self._send_more_queries(MAX_IN_FLIGHT) def _send_more_queries(self, num_outstanding): more_queries = [] while True: self.log(format=" there are %(outstanding)d queries outstanding", outstanding=len(self._queries_outstanding), level=log.NOISY) active_queries = len(self._queries_outstanding) + len(more_queries) if active_queries >= num_outstanding: break if not self.extra_peers: break more_queries.append(self.extra_peers.pop(0)) self.log(format="sending %(more)d more queries: %(who)s", more=len(more_queries), who=" ".join(["[%s]" % idlib.shortnodeid_b2a(peerid) for (peerid,ss) in more_queries]), level=log.NOISY) for (peerid, ss) in more_queries: self._do_query(ss, peerid, self._storage_index, self._read_size) # we'll retrigger when those queries come back def _done(self): if not self._running: return self._running = False self._servermap.last_update_mode = self.mode self._servermap.last_update_time = self._started # the servermap will not be touched after this self.log("servermap: %s" % self._servermap.summarize_versions()) eventually(self._done_deferred.callback, self._servermap) def _fatal_error(self, f): self.log("fatal error", failure=f, level=log.WEIRD) self._done_deferred.errback(f)