src/allmydata/mutable/publish.py

   1
   2
   3 import os, struct, time
   4 from itertools import count
   5 from zope.interface import implements
   6 from twisted.internet import defer
   7 from twisted.python import failure
   8 from allmydata.interfaces import IPublishStatus
   9 from allmydata.util import base32, hashutil, mathutil, idlib, log
  10 from allmydata.util.dictutil import DictOfSets
  11 from allmydata import hashtree, codec
  12 from allmydata.storage.server import si_b2a
  13 from pycryptopp.cipher.aes import AES
  14 from foolscap.api import eventually, fireEventually
  15
  16 from allmydata.mutable.common import MODE_WRITE, MODE_CHECK, \
  17      UncoordinatedWriteError, NotEnoughServersError
  18 from allmydata.mutable.servermap import ServerMap
  19 from allmydata.mutable.layout import pack_prefix, pack_share, unpack_header, pack_checkstring, \
  20      unpack_checkstring, SIGNED_PREFIX
  21
  22 class PublishStatus:
  23     implements(IPublishStatus)
  24     statusid_counter = count(0)
  25     def __init__(self):
  26         self.timings = {}
  27         self.timings["send_per_server"] = {}
  28         self.servermap = None
  29         self.problems = {}
  30         self.active = True
  31         self.storage_index = None
  32         self.helper = False
  33         self.encoding = ("?", "?")
  34         self.size = None
  35         self.status = "Not started"
  36         self.progress = 0.0
  37         self.counter = self.statusid_counter.next()
  38         self.started = time.time()
  39
  40     def add_per_server_time(self, peerid, elapsed):
  41         if peerid not in self.timings["send_per_server"]:
  42             self.timings["send_per_server"][peerid] = []
  43         self.timings["send_per_server"][peerid].append(elapsed)
  44
  45     def get_started(self):
  46         return self.started
  47     def get_storage_index(self):
  48         return self.storage_index
  49     def get_encoding(self):
  50         return self.encoding
  51     def using_helper(self):
  52         return self.helper
  53     def get_servermap(self):
  54         return self.servermap
  55     def get_size(self):
  56         return self.size
  57     def get_status(self):
  58         return self.status
  59     def get_progress(self):
  60         return self.progress
  61     def get_active(self):
  62         return self.active
  63     def get_counter(self):
  64         return self.counter
  65
  66     def set_storage_index(self, si):
  67         self.storage_index = si
  68     def set_helper(self, helper):
  69         self.helper = helper
  70     def set_servermap(self, servermap):
  71         self.servermap = servermap
  72     def set_encoding(self, k, n):
  73         self.encoding = (k, n)
  74     def set_size(self, size):
  75         self.size = size
  76     def set_status(self, status):
  77         self.status = status
  78     def set_progress(self, value):
  79         self.progress = value
  80     def set_active(self, value):
  81         self.active = value
  82
  83 class LoopLimitExceededError(Exception):
  84     pass
  85
  86 class Publish:
  87     """I represent a single act of publishing the mutable file to the grid. I
  88     will only publish my data if the servermap I am using still represents
  89     the current state of the world.
  90
  91     To make the initial publish, set servermap to None.
  92     """
  93
  94     def __init__(self, filenode, storage_broker, servermap):
  95         self._node = filenode
  96         self._storage_broker = storage_broker
  97         self._servermap = servermap
  98         self._storage_index = self._node.get_storage_index()
  99         self._log_prefix = prefix = si_b2a(self._storage_index)[:5]
 100         num = self.log("Publish(%s): starting" % prefix, parent=None)
 101         self._log_number = num
 102         self._running = True
 103         self._first_write_error = None
 104
 105         self._status = PublishStatus()
 106         self._status.set_storage_index(self._storage_index)
 107         self._status.set_helper(False)
 108         self._status.set_progress(0.0)
 109         self._status.set_active(True)
 110
 111     def get_status(self):
 112         return self._status
 113
 114     def log(self, *args, **kwargs):
 115         if 'parent' not in kwargs:
 116             kwargs['parent'] = self._log_number
 117         if "facility" not in kwargs:
 118             kwargs["facility"] = "tahoe.mutable.publish"
 119         return log.msg(*args, **kwargs)
 120
 121     def publish(self, newdata):
 122         """Publish the filenode's current contents.  Returns a Deferred that
 123         fires (with None) when the publish has done as much work as it's ever
 124         going to do, or errbacks with ConsistencyError if it detects a
 125         simultaneous write.
 126         """
 127
 128         # 1: generate shares (SDMF: files are small, so we can do it in RAM)
 129         # 2: perform peer selection, get candidate servers
 130         #  2a: send queries to n+epsilon servers, to determine current shares
 131         #  2b: based upon responses, create target map
 132         # 3: send slot_testv_and_readv_and_writev messages
 133         # 4: as responses return, update share-dispatch table
 134         # 4a: may need to run recovery algorithm
 135         # 5: when enough responses are back, we're done
 136
 137         self.log("starting publish, datalen is %s" % len(newdata))
 138         self._status.set_size(len(newdata))
 139         self._status.set_status("Started")
 140         self._started = time.time()
 141
 142         self.done_deferred = defer.Deferred()
 143
 144         self._writekey = self._node.get_writekey()
 145         assert self._writekey, "need write capability to publish"
 146
 147         # first, which servers will we publish to? We require that the
 148         # servermap was updated in MODE_WRITE, so we can depend upon the
 149         # peerlist computed by that process instead of computing our own.
 150         if self._servermap:
 151             assert self._servermap.last_update_mode in (MODE_WRITE, MODE_CHECK)
 152             # we will push a version that is one larger than anything present
 153             # in the grid, according to the servermap.
 154             self._new_seqnum = self._servermap.highest_seqnum() + 1
 155         else:
 156             # If we don't have a servermap, that's because we're doing the
 157             # initial publish
 158             self._new_seqnum = 1
 159             self._servermap = ServerMap()
 160         self._status.set_servermap(self._servermap)
 161
 162         self.log(format="new seqnum will be %(seqnum)d",
 163                  seqnum=self._new_seqnum, level=log.NOISY)
 164
 165         # having an up-to-date servermap (or using a filenode that was just
 166         # created for the first time) also guarantees that the following
 167         # fields are available
 168         self.readkey = self._node.get_readkey()
 169         self.required_shares = self._node.get_required_shares()
 170         assert self.required_shares is not None
 171         self.total_shares = self._node.get_total_shares()
 172         assert self.total_shares is not None
 173         self._status.set_encoding(self.required_shares, self.total_shares)
 174
 175         self._pubkey = self._node.get_pubkey()
 176         assert self._pubkey
 177         self._privkey = self._node.get_privkey()
 178         assert self._privkey
 179         self._encprivkey = self._node.get_encprivkey()
 180
 181         sb = self._storage_broker
 182         full_peerlist = sb.get_servers_for_index(self._storage_index)
 183         self.full_peerlist = full_peerlist # for use later, immutable
 184         self.bad_peers = set() # peerids who have errbacked/refused requests
 185
 186         self.newdata = newdata
 187         self.salt = os.urandom(16)
 188
 189         self.setup_encoding_parameters()
 190
 191         # if we experience any surprises (writes which were rejected because
 192         # our test vector did not match, or shares which we didn't expect to
 193         # see), we set this flag and report an UncoordinatedWriteError at the
 194         # end of the publish process.
 195         self.surprised = False
 196
 197         # as a failsafe, refuse to iterate through self.loop more than a
 198         # thousand times.
 199         self.looplimit = 1000
 200
 201         # we keep track of three tables. The first is our goal: which share
 202         # we want to see on which servers. This is initially populated by the
 203         # existing servermap.
 204         self.goal = set() # pairs of (peerid, shnum) tuples
 205
 206         # the second table is our list of outstanding queries: those which
 207         # are in flight and may or may not be delivered, accepted, or
 208         # acknowledged. Items are added to this table when the request is
 209         # sent, and removed when the response returns (or errbacks).
 210         self.outstanding = set() # (peerid, shnum) tuples
 211
 212         # the third is a table of successes: share which have actually been
 213         # placed. These are populated when responses come back with success.
 214         # When self.placed == self.goal, we're done.
 215         self.placed = set() # (peerid, shnum) tuples
 216
 217         # we also keep a mapping from peerid to RemoteReference. Each time we
 218         # pull a connection out of the full peerlist, we add it to this for
 219         # use later.
 220         self.connections = {}
 221
 222         self.bad_share_checkstrings = {}
 223
 224         # we use the servermap to populate the initial goal: this way we will
 225         # try to update each existing share in place.
 226         for (peerid, shnum) in self._servermap.servermap:
 227             self.goal.add( (peerid, shnum) )
 228             self.connections[peerid] = self._servermap.connections[peerid]
 229         # then we add in all the shares that were bad (corrupted, bad
 230         # signatures, etc). We want to replace these.
 231         for key, old_checkstring in self._servermap.bad_shares.items():
 232             (peerid, shnum) = key
 233             self.goal.add(key)
 234             self.bad_share_checkstrings[key] = old_checkstring
 235             self.connections[peerid] = self._servermap.connections[peerid]
 236
 237         # create the shares. We'll discard these as they are delivered. SDMF:
 238         # we're allowed to hold everything in memory.
 239
 240         self._status.timings["setup"] = time.time() - self._started
 241         d = self._encrypt_and_encode()
 242         d.addCallback(self._generate_shares)
 243         def _start_pushing(res):
 244             self._started_pushing = time.time()
 245             return res
 246         d.addCallback(_start_pushing)
 247         d.addCallback(self.loop) # trigger delivery
 248         d.addErrback(self._fatal_error)
 249
 250         return self.done_deferred
 251
 252     def setup_encoding_parameters(self):
 253         segment_size = len(self.newdata)
 254         # this must be a multiple of self.required_shares
 255         segment_size = mathutil.next_multiple(segment_size,
 256                                               self.required_shares)
 257         self.segment_size = segment_size
 258         if segment_size:
 259             self.num_segments = mathutil.div_ceil(len(self.newdata),
 260                                                   segment_size)
 261         else:
 262             self.num_segments = 0
 263         assert self.num_segments in [0, 1,] # SDMF restrictions
 264
 265     def _fatal_error(self, f):
 266         self.log("error during loop", failure=f, level=log.UNUSUAL)
 267         self._done(f)
 268
 269     def _update_status(self):
 270         self._status.set_status("Sending Shares: %d placed out of %d, "
 271                                 "%d messages outstanding" %
 272                                 (len(self.placed),
 273                                  len(self.goal),
 274                                  len(self.outstanding)))
 275         self._status.set_progress(1.0 * len(self.placed) / len(self.goal))
 276
 277     def loop(self, ignored=None):
 278         self.log("entering loop", level=log.NOISY)
 279         if not self._running:
 280             return
 281
 282         self.looplimit -= 1
 283         if self.looplimit <= 0:
 284             raise LoopLimitExceededError("loop limit exceeded")
 285
 286         if self.surprised:
 287             # don't send out any new shares, just wait for the outstanding
 288             # ones to be retired.
 289             self.log("currently surprised, so don't send any new shares",
 290                      level=log.NOISY)
 291         else:
 292             self.update_goal()
 293             # how far are we from our goal?
 294             needed = self.goal - self.placed - self.outstanding
 295             self._update_status()
 296
 297             if needed:
 298                 # we need to send out new shares
 299                 self.log(format="need to send %(needed)d new shares",
 300                          needed=len(needed), level=log.NOISY)
 301                 self._send_shares(needed)
 302                 return
 303
 304         if self.outstanding:
 305             # queries are still pending, keep waiting
 306             self.log(format="%(outstanding)d queries still outstanding",
 307                      outstanding=len(self.outstanding),
 308                      level=log.NOISY)
 309             return
 310
 311         # no queries outstanding, no placements needed: we're done
 312         self.log("no queries outstanding, no placements needed: done",
 313                  level=log.OPERATIONAL)
 314         now = time.time()
 315         elapsed = now - self._started_pushing
 316         self._status.timings["push"] = elapsed
 317         return self._done(None)
 318
 319     def log_goal(self, goal, message=""):
 320         logmsg = [message]
 321         for (shnum, peerid) in sorted([(s,p) for (p,s) in goal]):
 322             logmsg.append("sh%d to [%s]" % (shnum,
 323                                             idlib.shortnodeid_b2a(peerid)))
 324         self.log("current goal: %s" % (", ".join(logmsg)), level=log.NOISY)
 325         self.log("we are planning to push new seqnum=#%d" % self._new_seqnum,
 326                  level=log.NOISY)
 327
 328     def update_goal(self):
 329         # if log.recording_noisy
 330         if True:
 331             self.log_goal(self.goal, "before update: ")
 332
 333         # first, remove any bad peers from our goal
 334         self.goal = set([ (peerid, shnum)
 335                           for (peerid, shnum) in self.goal
 336                           if peerid not in self.bad_peers ])
 337
 338         # find the homeless shares:
 339         homefull_shares = set([shnum for (peerid, shnum) in self.goal])
 340         homeless_shares = set(range(self.total_shares)) - homefull_shares
 341         homeless_shares = sorted(list(homeless_shares))
 342         # place them somewhere. We prefer unused servers at the beginning of
 343         # the available peer list.
 344
 345         if not homeless_shares:
 346             return
 347
 348         # if an old share X is on a node, put the new share X there too.
 349         # TODO: 1: redistribute shares to achieve one-per-peer, by copying
 350         #       shares from existing peers to new (less-crowded) ones. The
 351         #       old shares must still be updated.
 352         # TODO: 2: move those shares instead of copying them, to reduce future
 353         #       update work
 354
 355         # this is a bit CPU intensive but easy to analyze. We create a sort
 356         # order for each peerid. If the peerid is marked as bad, we don't
 357         # even put them in the list. Then we care about the number of shares
 358         # which have already been assigned to them. After that we care about
 359         # their permutation order.
 360         old_assignments = DictOfSets()
 361         for (peerid, shnum) in self.goal:
 362             old_assignments.add(peerid, shnum)
 363
 364         peerlist = []
 365         for i, (peerid, ss) in enumerate(self.full_peerlist):
 366             if peerid in self.bad_peers:
 367                 continue
 368             entry = (len(old_assignments.get(peerid, [])), i, peerid, ss)
 369             peerlist.append(entry)
 370         peerlist.sort()
 371
 372         if not peerlist:
 373             raise NotEnoughServersError("Ran out of non-bad servers, "
 374                                         "first_error=%s" %
 375                                         str(self._first_write_error),
 376                                         self._first_write_error)
 377
 378         # we then index this peerlist with an integer, because we may have to
 379         # wrap. We update the goal as we go.
 380         i = 0
 381         for shnum in homeless_shares:
 382             (ignored1, ignored2, peerid, ss) = peerlist[i]
 383             # if we are forced to send a share to a server that already has
 384             # one, we may have two write requests in flight, and the
 385             # servermap (which was computed before either request was sent)
 386             # won't reflect the new shares, so the second response will be
 387             # surprising. There is code in _got_write_answer() to tolerate
 388             # this, otherwise it would cause the publish to fail with an
 389             # UncoordinatedWriteError. See #546 for details of the trouble
 390             # this used to cause.
 391             self.goal.add( (peerid, shnum) )
 392             self.connections[peerid] = ss
 393             i += 1
 394             if i >= len(peerlist):
 395                 i = 0
 396         if True:
 397             self.log_goal(self.goal, "after update: ")
 398
 399
 400
 401     def _encrypt_and_encode(self):
 402         # this returns a Deferred that fires with a list of (sharedata,
 403         # sharenum) tuples. TODO: cache the ciphertext, only produce the
 404         # shares that we care about.
 405         self.log("_encrypt_and_encode")
 406
 407         self._status.set_status("Encrypting")
 408         started = time.time()
 409
 410         key = hashutil.ssk_readkey_data_hash(self.salt, self.readkey)
 411         enc = AES(key)
 412         crypttext = enc.process(self.newdata)
 413         assert len(crypttext) == len(self.newdata)
 414
 415         now = time.time()
 416         self._status.timings["encrypt"] = now - started
 417         started = now
 418
 419         # now apply FEC
 420
 421         self._status.set_status("Encoding")
 422         fec = codec.CRSEncoder()
 423         fec.set_params(self.segment_size,
 424                        self.required_shares, self.total_shares)
 425         piece_size = fec.get_block_size()
 426         crypttext_pieces = [None] * self.required_shares
 427         for i in range(len(crypttext_pieces)):
 428             offset = i * piece_size
 429             piece = crypttext[offset:offset+piece_size]
 430             piece = piece + "\x00"*(piece_size - len(piece)) # padding
 431             crypttext_pieces[i] = piece
 432             assert len(piece) == piece_size
 433
 434         d = fec.encode(crypttext_pieces)
 435         def _done_encoding(res):
 436             elapsed = time.time() - started
 437             self._status.timings["encode"] = elapsed
 438             return res
 439         d.addCallback(_done_encoding)
 440         return d
 441
 442     def _generate_shares(self, shares_and_shareids):
 443         # this sets self.shares and self.root_hash
 444         self.log("_generate_shares")
 445         self._status.set_status("Generating Shares")
 446         started = time.time()
 447
 448         # we should know these by now
 449         privkey = self._privkey
 450         encprivkey = self._encprivkey
 451         pubkey = self._pubkey
 452
 453         (shares, share_ids) = shares_and_shareids
 454
 455         assert len(shares) == len(share_ids)
 456         assert len(shares) == self.total_shares
 457         all_shares = {}
 458         block_hash_trees = {}
 459         share_hash_leaves = [None] * len(shares)
 460         for i in range(len(shares)):
 461             share_data = shares[i]
 462             shnum = share_ids[i]
 463             all_shares[shnum] = share_data
 464
 465             # build the block hash tree. SDMF has only one leaf.
 466             leaves = [hashutil.block_hash(share_data)]
 467             t = hashtree.HashTree(leaves)
 468             block_hash_trees[shnum] = list(t)
 469             share_hash_leaves[shnum] = t[0]
 470         for leaf in share_hash_leaves:
 471             assert leaf is not None
 472         share_hash_tree = hashtree.HashTree(share_hash_leaves)
 473         share_hash_chain = {}
 474         for shnum in range(self.total_shares):
 475             needed_hashes = share_hash_tree.needed_hashes(shnum)
 476             share_hash_chain[shnum] = dict( [ (i, share_hash_tree[i])
 477                                               for i in needed_hashes ] )
 478         root_hash = share_hash_tree[0]
 479         assert len(root_hash) == 32
 480         self.log("my new root_hash is %s" % base32.b2a(root_hash))
 481         self._new_version_info = (self._new_seqnum, root_hash, self.salt)
 482
 483         prefix = pack_prefix(self._new_seqnum, root_hash, self.salt,
 484                              self.required_shares, self.total_shares,
 485                              self.segment_size, len(self.newdata))
 486
 487         # now pack the beginning of the share. All shares are the same up
 488         # to the signature, then they have divergent share hash chains,
 489         # then completely different block hash trees + salt + share data,
 490         # then they all share the same encprivkey at the end. The sizes
 491         # of everything are the same for all shares.
 492
 493         sign_started = time.time()
 494         signature = privkey.sign(prefix)
 495         self._status.timings["sign"] = time.time() - sign_started
 496
 497         verification_key = pubkey.serialize()
 498
 499         final_shares = {}
 500         for shnum in range(self.total_shares):
 501             final_share = pack_share(prefix,
 502                                      verification_key,
 503                                      signature,
 504                                      share_hash_chain[shnum],
 505                                      block_hash_trees[shnum],
 506                                      all_shares[shnum],
 507                                      encprivkey)
 508             final_shares[shnum] = final_share
 509         elapsed = time.time() - started
 510         self._status.timings["pack"] = elapsed
 511         self.shares = final_shares
 512         self.root_hash = root_hash
 513
 514         # we also need to build up the version identifier for what we're
 515         # pushing. Extract the offsets from one of our shares.
 516         assert final_shares
 517         offsets = unpack_header(final_shares.values()[0])[-1]
 518         offsets_tuple = tuple( [(key,value) for key,value in offsets.items()] )
 519         verinfo = (self._new_seqnum, root_hash, self.salt,
 520                    self.segment_size, len(self.newdata),
 521                    self.required_shares, self.total_shares,
 522                    prefix, offsets_tuple)
 523         self.versioninfo = verinfo
 524
 525
 526
 527     def _send_shares(self, needed):
 528         self.log("_send_shares")
 529
 530         # we're finally ready to send out our shares. If we encounter any
 531         # surprises here, it's because somebody else is writing at the same
 532         # time. (Note: in the future, when we remove the _query_peers() step
 533         # and instead speculate about [or remember] which shares are where,
 534         # surprises here are *not* indications of UncoordinatedWriteError,
 535         # and we'll need to respond to them more gracefully.)
 536
 537         # needed is a set of (peerid, shnum) tuples. The first thing we do is
 538         # organize it by peerid.
 539
 540         peermap = DictOfSets()
 541         for (peerid, shnum) in needed:
 542             peermap.add(peerid, shnum)
 543
 544         # the next thing is to build up a bunch of test vectors. The
 545         # semantics of Publish are that we perform the operation if the world
 546         # hasn't changed since the ServerMap was constructed (more or less).
 547         # For every share we're trying to place, we create a test vector that
 548         # tests to see if the server*share still corresponds to the
 549         # map.
 550
 551         all_tw_vectors = {} # maps peerid to tw_vectors
 552         sm = self._servermap.servermap
 553
 554         for key in needed:
 555             (peerid, shnum) = key
 556
 557             if key in sm:
 558                 # an old version of that share already exists on the
 559                 # server, according to our servermap. We will create a
 560                 # request that attempts to replace it.
 561                 old_versionid, old_timestamp = sm[key]
 562                 (old_seqnum, old_root_hash, old_salt, old_segsize,
 563                  old_datalength, old_k, old_N, old_prefix,
 564                  old_offsets_tuple) = old_versionid
 565                 old_checkstring = pack_checkstring(old_seqnum,
 566                                                    old_root_hash,
 567                                                    old_salt)
 568                 testv = (0, len(old_checkstring), "eq", old_checkstring)
 569
 570             elif key in self.bad_share_checkstrings:
 571                 old_checkstring = self.bad_share_checkstrings[key]
 572                 testv = (0, len(old_checkstring), "eq", old_checkstring)
 573
 574             else:
 575                 # add a testv that requires the share not exist
 576
 577                 # Unfortunately, foolscap-0.2.5 has a bug in the way inbound
 578                 # constraints are handled. If the same object is referenced
 579                 # multiple times inside the arguments, foolscap emits a
 580                 # 'reference' token instead of a distinct copy of the
 581                 # argument. The bug is that these 'reference' tokens are not
 582                 # accepted by the inbound constraint code. To work around
 583                 # this, we need to prevent python from interning the
 584                 # (constant) tuple, by creating a new copy of this vector
 585                 # each time.
 586
 587                 # This bug is fixed in foolscap-0.2.6, and even though this
 588                 # version of Tahoe requires foolscap-0.3.1 or newer, we are
 589                 # supposed to be able to interoperate with older versions of
 590                 # Tahoe which are allowed to use older versions of foolscap,
 591                 # including foolscap-0.2.5 . In addition, I've seen other
 592                 # foolscap problems triggered by 'reference' tokens (see #541
 593                 # for details). So we must keep this workaround in place.
 594
 595                 #testv = (0, 1, 'eq', "")
 596                 testv = tuple([0, 1, 'eq', ""])
 597
 598             testvs = [testv]
 599             # the write vector is simply the share
 600             writev = [(0, self.shares[shnum])]
 601
 602             if peerid not in all_tw_vectors:
 603                 all_tw_vectors[peerid] = {}
 604                 # maps shnum to (testvs, writevs, new_length)
 605             assert shnum not in all_tw_vectors[peerid]
 606
 607             all_tw_vectors[peerid][shnum] = (testvs, writev, None)
 608
 609         # we read the checkstring back from each share, however we only use
 610         # it to detect whether there was a new share that we didn't know
 611         # about. The success or failure of the write will tell us whether
 612         # there was a collision or not. If there is a collision, the first
 613         # thing we'll do is update the servermap, which will find out what
 614         # happened. We could conceivably reduce a roundtrip by using the
 615         # readv checkstring to populate the servermap, but really we'd have
 616         # to read enough data to validate the signatures too, so it wouldn't
 617         # be an overall win.
 618         read_vector = [(0, struct.calcsize(SIGNED_PREFIX))]
 619
 620         # ok, send the messages!
 621         self.log("sending %d shares" % len(all_tw_vectors), level=log.NOISY)
 622         started = time.time()
 623         for (peerid, tw_vectors) in all_tw_vectors.items():
 624
 625             write_enabler = self._node.get_write_enabler(peerid)
 626             renew_secret = self._node.get_renewal_secret(peerid)
 627             cancel_secret = self._node.get_cancel_secret(peerid)
 628             secrets = (write_enabler, renew_secret, cancel_secret)
 629             shnums = tw_vectors.keys()
 630
 631             for shnum in shnums:
 632                 self.outstanding.add( (peerid, shnum) )
 633
 634             d = self._do_testreadwrite(peerid, secrets,
 635                                        tw_vectors, read_vector)
 636             d.addCallbacks(self._got_write_answer, self._got_write_error,
 637                            callbackArgs=(peerid, shnums, started),
 638                            errbackArgs=(peerid, shnums, started))
 639             # tolerate immediate errback, like with DeadReferenceError
 640             d.addBoth(fireEventually)
 641             d.addCallback(self.loop)
 642             d.addErrback(self._fatal_error)
 643
 644         self._update_status()
 645         self.log("%d shares sent" % len(all_tw_vectors), level=log.NOISY)
 646
 647     def _do_testreadwrite(self, peerid, secrets,
 648                           tw_vectors, read_vector):
 649         storage_index = self._storage_index
 650         ss = self.connections[peerid]
 651
 652         #print "SS[%s] is %s" % (idlib.shortnodeid_b2a(peerid), ss), ss.tracker.interfaceName
 653         d = ss.callRemote("slot_testv_and_readv_and_writev",
 654                           storage_index,
 655                           secrets,
 656                           tw_vectors,
 657                           read_vector)
 658         return d
 659
 660     def _got_write_answer(self, answer, peerid, shnums, started):
 661         lp = self.log("_got_write_answer from %s" %
 662                       idlib.shortnodeid_b2a(peerid))
 663         for shnum in shnums:
 664             self.outstanding.discard( (peerid, shnum) )
 665
 666         now = time.time()
 667         elapsed = now - started
 668         self._status.add_per_server_time(peerid, elapsed)
 669
 670         wrote, read_data = answer
 671
 672         surprise_shares = set(read_data.keys()) - set(shnums)
 673
 674         surprised = False
 675         for shnum in surprise_shares:
 676             # read_data is a dict mapping shnum to checkstring (SIGNED_PREFIX)
 677             checkstring = read_data[shnum][0]
 678             their_version_info = unpack_checkstring(checkstring)
 679             if their_version_info == self._new_version_info:
 680                 # they have the right share, somehow
 681
 682                 if (peerid,shnum) in self.goal:
 683                     # and we want them to have it, so we probably sent them a
 684                     # copy in an earlier write. This is ok, and avoids the
 685                     # #546 problem.
 686                     continue
 687
 688                 # They aren't in our goal, but they are still for the right
 689                 # version. Somebody else wrote them, and it's a convergent
 690                 # uncoordinated write. Pretend this is ok (don't be
 691                 # surprised), since I suspect there's a decent chance that
 692                 # we'll hit this in normal operation.
 693                 continue
 694
 695             else:
 696                 # the new shares are of a different version
 697                 if peerid in self._servermap.reachable_peers:
 698                     # we asked them about their shares, so we had knowledge
 699                     # of what they used to have. Any surprising shares must
 700                     # have come from someone else, so UCW.
 701                     surprised = True
 702                 else:
 703                     # we didn't ask them, and now we've discovered that they
 704                     # have a share we didn't know about. This indicates that
 705                     # mapupdate should have wokred harder and asked more
 706                     # servers before concluding that it knew about them all.
 707
 708                     # signal UCW, but make sure to ask this peer next time,
 709                     # so we'll remember to update it if/when we retry.
 710                     surprised = True
 711                     # TODO: ask this peer next time. I don't yet have a good
 712                     # way to do this. Two insufficient possibilities are:
 713                     #
 714                     # self._servermap.add_new_share(peerid, shnum, verinfo, now)
 715                     #  but that requires fetching/validating/parsing the whole
 716                     #  version string, and all we have is the checkstring
 717                     # self._servermap.mark_bad_share(peerid, shnum, checkstring)
 718                     #  that will make publish overwrite the share next time,
 719                     #  but it won't re-query the server, and it won't make
 720                     #  mapupdate search further
 721
 722                     # TODO later: when publish starts, do
 723                     # servermap.get_best_version(), extract the seqnum,
 724                     # subtract one, and store as highest-replaceable-seqnum.
 725                     # Then, if this surprise-because-we-didn't-ask share is
 726                     # of highest-replaceable-seqnum or lower, we're allowed
 727                     # to replace it: send out a new writev (or rather add it
 728                     # to self.goal and loop).
 729                     pass
 730
 731                 surprised = True
 732
 733         if surprised:
 734             self.log("they had shares %s that we didn't know about" %
 735                      (list(surprise_shares),),
 736                      parent=lp, level=log.WEIRD, umid="un9CSQ")
 737             self.surprised = True
 738
 739         if not wrote:
 740             # TODO: there are two possibilities. The first is that the server
 741             # is full (or just doesn't want to give us any room), which means
 742             # we shouldn't ask them again, but is *not* an indication of an
 743             # uncoordinated write. The second is that our testv failed, which
 744             # *does* indicate an uncoordinated write. We currently don't have
 745             # a way to tell these two apart (in fact, the storage server code
 746             # doesn't have the option of refusing our share).
 747             #
 748             # If the server is full, mark the peer as bad (so we don't ask
 749             # them again), but don't set self.surprised. The loop() will find
 750             # a new server.
 751             #
 752             # If the testv failed, log it, set self.surprised, but don't
 753             # bother adding to self.bad_peers .
 754
 755             self.log("our testv failed, so the write did not happen",
 756                      parent=lp, level=log.WEIRD, umid="8sc26g")
 757             self.surprised = True
 758             self.bad_peers.add(peerid) # don't ask them again
 759             # use the checkstring to add information to the log message
 760             for (shnum,readv) in read_data.items():
 761                 checkstring = readv[0]
 762                 (other_seqnum,
 763                  other_roothash,
 764                  other_salt) = unpack_checkstring(checkstring)
 765                 expected_version = self._servermap.version_on_peer(peerid,
 766                                                                    shnum)
 767                 if expected_version:
 768                     (seqnum, root_hash, IV, segsize, datalength, k, N, prefix,
 769                      offsets_tuple) = expected_version
 770                     self.log("somebody modified the share on us:"
 771                              " shnum=%d: I thought they had #%d:R=%s,"
 772                              " but testv reported #%d:R=%s" %
 773                              (shnum,
 774                               seqnum, base32.b2a(root_hash)[:4],
 775                               other_seqnum, base32.b2a(other_roothash)[:4]),
 776                              parent=lp, level=log.NOISY)
 777                 # if expected_version==None, then we didn't expect to see a
 778                 # share on that peer, and the 'surprise_shares' clause above
 779                 # will have logged it.
 780             # self.loop() will take care of finding new homes
 781             return
 782
 783         for shnum in shnums:
 784             self.placed.add( (peerid, shnum) )
 785             # and update the servermap
 786             self._servermap.add_new_share(peerid, shnum,
 787                                           self.versioninfo, started)
 788
 789         # self.loop() will take care of checking to see if we're done
 790         return
 791
 792     def _got_write_error(self, f, peerid, shnums, started):
 793         for shnum in shnums:
 794             self.outstanding.discard( (peerid, shnum) )
 795         self.bad_peers.add(peerid)
 796         if self._first_write_error is None:
 797             self._first_write_error = f
 798         self.log(format="error while writing shares %(shnums)s to peerid %(peerid)s",
 799                  shnums=list(shnums), peerid=idlib.shortnodeid_b2a(peerid),
 800                  failure=f,
 801                  level=log.UNUSUAL)
 802         # self.loop() will take care of checking to see if we're done
 803         return
 804
 805
 806     def _done(self, res):
 807         if not self._running:
 808             return
 809         self._running = False
 810         now = time.time()
 811         self._status.timings["total"] = now - self._started
 812         self._status.set_active(False)
 813         if isinstance(res, failure.Failure):
 814             self.log("Publish done, with failure", failure=res,
 815                      level=log.WEIRD, umid="nRsR9Q")
 816             self._status.set_status("Failed")
 817         elif self.surprised:
 818             self.log("Publish done, UncoordinatedWriteError", level=log.UNUSUAL)
 819             self._status.set_status("UncoordinatedWriteError")
 820             # deliver a failure
 821             res = failure.Failure(UncoordinatedWriteError())
 822             # TODO: recovery
 823         else:
 824             self.log("Publish done, success")
 825             self._status.set_status("Finished")
 826             self._status.set_progress(1.0)
 827         eventually(self.done_deferred.callback, res)
 828
 829