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, FileTooLargeError
   9 from allmydata.util import base32, hashutil, mathutil, idlib, log
  10 from allmydata import hashtree, codec
  11 from allmydata.storage.server import si_b2a
  12 from pycryptopp.cipher.aes import AES
  13 from foolscap.eventual import eventually
  14
  15 from common import MODE_WRITE, MODE_CHECK, DictOfSets, \
  16      UncoordinatedWriteError, NotEnoughServersError
  17 from servermap import ServerMap
  18 from layout import pack_prefix, pack_share, unpack_header, pack_checkstring, \
  19      unpack_checkstring, SIGNED_PREFIX
  20
  21 class PublishStatus:
  22     implements(IPublishStatus)
  23     statusid_counter = count(0)
  24     def __init__(self):
  25         self.timings = {}
  26         self.timings["send_per_server"] = {}
  27         self.servermap = None
  28         self.problems = {}
  29         self.active = True
  30         self.storage_index = None
  31         self.helper = False
  32         self.encoding = ("?", "?")
  33         self.size = None
  34         self.status = "Not started"
  35         self.progress = 0.0
  36         self.counter = self.statusid_counter.next()
  37         self.started = time.time()
  38
  39     def add_per_server_time(self, peerid, elapsed):
  40         if peerid not in self.timings["send_per_server"]:
  41             self.timings["send_per_server"][peerid] = []
  42         self.timings["send_per_server"][peerid].append(elapsed)
  43
  44     def get_started(self):
  45         return self.started
  46     def get_storage_index(self):
  47         return self.storage_index
  48     def get_encoding(self):
  49         return self.encoding
  50     def using_helper(self):
  51         return self.helper
  52     def get_servermap(self):
  53         return self.servermap
  54     def get_size(self):
  55         return self.size
  56     def get_status(self):
  57         return self.status
  58     def get_progress(self):
  59         return self.progress
  60     def get_active(self):
  61         return self.active
  62     def get_counter(self):
  63         return self.counter
  64
  65     def set_storage_index(self, si):
  66         self.storage_index = si
  67     def set_helper(self, helper):
  68         self.helper = helper
  69     def set_servermap(self, servermap):
  70         self.servermap = servermap
  71     def set_encoding(self, k, n):
  72         self.encoding = (k, n)
  73     def set_size(self, size):
  74         self.size = size
  75     def set_status(self, status):
  76         self.status = status
  77     def set_progress(self, value):
  78         self.progress = value
  79     def set_active(self, value):
  80         self.active = value
  81
  82 class LoopLimitExceededError(Exception):
  83     pass
  84
  85 class Publish:
  86     """I represent a single act of publishing the mutable file to the grid. I
  87     will only publish my data if the servermap I am using still represents
  88     the current state of the world.
  89
  90     To make the initial publish, set servermap to None.
  91     """
  92
  93     # we limit the segment size as usual to constrain our memory footprint.
  94     # The max segsize is higher for mutable files, because we want to support
  95     # dirnodes with up to 10k children, and each child uses about 330 bytes.
  96     # If you actually put that much into a directory you'll be using a
  97     # footprint of around 14MB, which is higher than we'd like, but it is
  98     # more important right now to support large directories than to make
  99     # memory usage small when you use them. Once we implement MDMF (with
 100     # multiple segments), we will drop this back down, probably to 128KiB.
 101     MAX_SEGMENT_SIZE = 3500000
 102
 103     def __init__(self, filenode, servermap):
 104         self._node = filenode
 105         self._servermap = servermap
 106         self._storage_index = self._node.get_storage_index()
 107         self._log_prefix = prefix = si_b2a(self._storage_index)[:5]
 108         num = self._node._client.log("Publish(%s): starting" % prefix)
 109         self._log_number = num
 110         self._running = True
 111
 112         self._status = PublishStatus()
 113         self._status.set_storage_index(self._storage_index)
 114         self._status.set_helper(False)
 115         self._status.set_progress(0.0)
 116         self._status.set_active(True)
 117
 118     def get_status(self):
 119         return self._status
 120
 121     def log(self, *args, **kwargs):
 122         if 'parent' not in kwargs:
 123             kwargs['parent'] = self._log_number
 124         if "facility" not in kwargs:
 125             kwargs["facility"] = "tahoe.mutable.publish"
 126         return log.msg(*args, **kwargs)
 127
 128     def publish(self, newdata):
 129         """Publish the filenode's current contents.  Returns a Deferred that
 130         fires (with None) when the publish has done as much work as it's ever
 131         going to do, or errbacks with ConsistencyError if it detects a
 132         simultaneous write.
 133         """
 134
 135         # 1: generate shares (SDMF: files are small, so we can do it in RAM)
 136         # 2: perform peer selection, get candidate servers
 137         #  2a: send queries to n+epsilon servers, to determine current shares
 138         #  2b: based upon responses, create target map
 139         # 3: send slot_testv_and_readv_and_writev messages
 140         # 4: as responses return, update share-dispatch table
 141         # 4a: may need to run recovery algorithm
 142         # 5: when enough responses are back, we're done
 143
 144         self.log("starting publish, datalen is %s" % len(newdata))
 145         if len(newdata) > self.MAX_SEGMENT_SIZE:
 146             raise FileTooLargeError("SDMF is limited to one segment, and "
 147                                     "%d > %d" % (len(newdata),
 148                                                  self.MAX_SEGMENT_SIZE))
 149         self._status.set_size(len(newdata))
 150         self._status.set_status("Started")
 151         self._started = time.time()
 152
 153         self.done_deferred = defer.Deferred()
 154
 155         self._writekey = self._node.get_writekey()
 156         assert self._writekey, "need write capability to publish"
 157
 158         # first, which servers will we publish to? We require that the
 159         # servermap was updated in MODE_WRITE, so we can depend upon the
 160         # peerlist computed by that process instead of computing our own.
 161         if self._servermap:
 162             assert self._servermap.last_update_mode in (MODE_WRITE, MODE_CHECK)
 163             # we will push a version that is one larger than anything present
 164             # in the grid, according to the servermap.
 165             self._new_seqnum = self._servermap.highest_seqnum() + 1
 166         else:
 167             # If we don't have a servermap, that's because we're doing the
 168             # initial publish
 169             self._new_seqnum = 1
 170             self._servermap = ServerMap()
 171         self._status.set_servermap(self._servermap)
 172
 173         self.log(format="new seqnum will be %(seqnum)d",
 174                  seqnum=self._new_seqnum, level=log.NOISY)
 175
 176         # having an up-to-date servermap (or using a filenode that was just
 177         # created for the first time) also guarantees that the following
 178         # fields are available
 179         self.readkey = self._node.get_readkey()
 180         self.required_shares = self._node.get_required_shares()
 181         assert self.required_shares is not None
 182         self.total_shares = self._node.get_total_shares()
 183         assert self.total_shares is not None
 184         self._status.set_encoding(self.required_shares, self.total_shares)
 185
 186         self._pubkey = self._node.get_pubkey()
 187         assert self._pubkey
 188         self._privkey = self._node.get_privkey()
 189         assert self._privkey
 190         self._encprivkey = self._node.get_encprivkey()
 191
 192         client = self._node._client
 193         full_peerlist = client.get_permuted_peers("storage",
 194                                                   self._storage_index)
 195         self.full_peerlist = full_peerlist # for use later, immutable
 196         self.bad_peers = set() # peerids who have errbacked/refused requests
 197
 198         self.newdata = newdata
 199         self.salt = os.urandom(16)
 200
 201         self.setup_encoding_parameters()
 202
 203         # if we experience any surprises (writes which were rejected because
 204         # our test vector did not match, or shares which we didn't expect to
 205         # see), we set this flag and report an UncoordinatedWriteError at the
 206         # end of the publish process.
 207         self.surprised = False
 208
 209         # as a failsafe, refuse to iterate through self.loop more than a
 210         # thousand times.
 211         self.looplimit = 1000
 212
 213         # we keep track of three tables. The first is our goal: which share
 214         # we want to see on which servers. This is initially populated by the
 215         # existing servermap.
 216         self.goal = set() # pairs of (peerid, shnum) tuples
 217
 218         # the second table is our list of outstanding queries: those which
 219         # are in flight and may or may not be delivered, accepted, or
 220         # acknowledged. Items are added to this table when the request is
 221         # sent, and removed when the response returns (or errbacks).
 222         self.outstanding = set() # (peerid, shnum) tuples
 223
 224         # the third is a table of successes: share which have actually been
 225         # placed. These are populated when responses come back with success.
 226         # When self.placed == self.goal, we're done.
 227         self.placed = set() # (peerid, shnum) tuples
 228
 229         # we also keep a mapping from peerid to RemoteReference. Each time we
 230         # pull a connection out of the full peerlist, we add it to this for
 231         # use later.
 232         self.connections = {}
 233
 234         self.bad_share_checkstrings = {}
 235
 236         # we use the servermap to populate the initial goal: this way we will
 237         # try to update each existing share in place.
 238         for (peerid, shnum) in self._servermap.servermap:
 239             self.goal.add( (peerid, shnum) )
 240             self.connections[peerid] = self._servermap.connections[peerid]
 241         # then we add in all the shares that were bad (corrupted, bad
 242         # signatures, etc). We want to replace these.
 243         for key, old_checkstring in self._servermap.bad_shares.items():
 244             (peerid, shnum) = key
 245             self.goal.add(key)
 246             self.bad_share_checkstrings[key] = old_checkstring
 247             self.connections[peerid] = self._servermap.connections[peerid]
 248
 249         # create the shares. We'll discard these as they are delivered. SDMF:
 250         # we're allowed to hold everything in memory.
 251
 252         self._status.timings["setup"] = time.time() - self._started
 253         d = self._encrypt_and_encode()
 254         d.addCallback(self._generate_shares)
 255         def _start_pushing(res):
 256             self._started_pushing = time.time()
 257             return res
 258         d.addCallback(_start_pushing)
 259         d.addCallback(self.loop) # trigger delivery
 260         d.addErrback(self._fatal_error)
 261
 262         return self.done_deferred
 263
 264     def setup_encoding_parameters(self):
 265         segment_size = min(self.MAX_SEGMENT_SIZE, len(self.newdata))
 266         # this must be a multiple of self.required_shares
 267         segment_size = mathutil.next_multiple(segment_size,
 268                                               self.required_shares)
 269         self.segment_size = segment_size
 270         if segment_size:
 271             self.num_segments = mathutil.div_ceil(len(self.newdata),
 272                                                   segment_size)
 273         else:
 274             self.num_segments = 0
 275         assert self.num_segments in [0, 1,] # SDMF restrictions
 276
 277     def _fatal_error(self, f):
 278         self.log("error during loop", failure=f, level=log.UNUSUAL)
 279         self._done(f)
 280
 281     def _update_status(self):
 282         self._status.set_status("Sending Shares: %d placed out of %d, "
 283                                 "%d messages outstanding" %
 284                                 (len(self.placed),
 285                                  len(self.goal),
 286                                  len(self.outstanding)))
 287         self._status.set_progress(1.0 * len(self.placed) / len(self.goal))
 288
 289     def loop(self, ignored=None):
 290         self.log("entering loop", level=log.NOISY)
 291         if not self._running:
 292             return
 293
 294         self.looplimit -= 1
 295         if self.looplimit <= 0:
 296             raise LoopLimitExceededError("loop limit exceeded")
 297
 298         if self.surprised:
 299             # don't send out any new shares, just wait for the outstanding
 300             # ones to be retired.
 301             self.log("currently surprised, so don't send any new shares",
 302                      level=log.NOISY)
 303         else:
 304             self.update_goal()
 305             # how far are we from our goal?
 306             needed = self.goal - self.placed - self.outstanding
 307             self._update_status()
 308
 309             if needed:
 310                 # we need to send out new shares
 311                 self.log(format="need to send %(needed)d new shares",
 312                          needed=len(needed), level=log.NOISY)
 313                 self._send_shares(needed)
 314                 return
 315
 316         if self.outstanding:
 317             # queries are still pending, keep waiting
 318             self.log(format="%(outstanding)d queries still outstanding",
 319                      outstanding=len(self.outstanding),
 320                      level=log.NOISY)
 321             return
 322
 323         # no queries outstanding, no placements needed: we're done
 324         self.log("no queries outstanding, no placements needed: done",
 325                  level=log.OPERATIONAL)
 326         now = time.time()
 327         elapsed = now - self._started_pushing
 328         self._status.timings["push"] = elapsed
 329         return self._done(None)
 330
 331     def log_goal(self, goal, message=""):
 332         logmsg = [message]
 333         for (shnum, peerid) in sorted([(s,p) for (p,s) in goal]):
 334             logmsg.append("sh%d to [%s]" % (shnum,
 335                                             idlib.shortnodeid_b2a(peerid)))
 336         self.log("current goal: %s" % (", ".join(logmsg)), level=log.NOISY)
 337         self.log("we are planning to push new seqnum=#%d" % self._new_seqnum,
 338                  level=log.NOISY)
 339
 340     def update_goal(self):
 341         # if log.recording_noisy
 342         if True:
 343             self.log_goal(self.goal, "before update: ")
 344
 345         # first, remove any bad peers from our goal
 346         self.goal = set([ (peerid, shnum)
 347                           for (peerid, shnum) in self.goal
 348                           if peerid not in self.bad_peers ])
 349
 350         # find the homeless shares:
 351         homefull_shares = set([shnum for (peerid, shnum) in self.goal])
 352         homeless_shares = set(range(self.total_shares)) - homefull_shares
 353         homeless_shares = sorted(list(homeless_shares))
 354         # place them somewhere. We prefer unused servers at the beginning of
 355         # the available peer list.
 356
 357         if not homeless_shares:
 358             return
 359
 360         # if an old share X is on a node, put the new share X there too.
 361         # TODO: 1: redistribute shares to achieve one-per-peer, by copying
 362         #       shares from existing peers to new (less-crowded) ones. The
 363         #       old shares must still be updated.
 364         # TODO: 2: move those shares instead of copying them, to reduce future
 365         #       update work
 366
 367         # this is a bit CPU intensive but easy to analyze. We create a sort
 368         # order for each peerid. If the peerid is marked as bad, we don't
 369         # even put them in the list. Then we care about the number of shares
 370         # which have already been assigned to them. After that we care about
 371         # their permutation order.
 372         old_assignments = DictOfSets()
 373         for (peerid, shnum) in self.goal:
 374             old_assignments.add(peerid, shnum)
 375
 376         peerlist = []
 377         for i, (peerid, ss) in enumerate(self.full_peerlist):
 378             if peerid in self.bad_peers:
 379                 continue
 380             entry = (len(old_assignments.get(peerid, [])), i, peerid, ss)
 381             peerlist.append(entry)
 382         peerlist.sort()
 383
 384         if not peerlist:
 385             raise NotEnoughServersError("Ran out of non-bad servers")
 386
 387         new_assignments = []
 388         # we then index this peerlist with an integer, because we may have to
 389         # wrap. We update the goal as we go.
 390         i = 0
 391         for shnum in homeless_shares:
 392             (ignored1, ignored2, peerid, ss) = peerlist[i]
 393             # if we are forced to send a share to a server that already has
 394             # one, we may have two write requests in flight, and the
 395             # servermap (which was computed before either request was sent)
 396             # won't reflect the new shares, so the second response will be
 397             # surprising. There is code in _got_write_answer() to tolerate
 398             # this, otherwise it would cause the publish to fail with an
 399             # UncoordinatedWriteError. See #546 for details of the trouble
 400             # this used to cause.
 401             self.goal.add( (peerid, shnum) )
 402             self.connections[peerid] = ss
 403             i += 1
 404             if i >= len(peerlist):
 405                 i = 0
 406         if True:
 407             self.log_goal(self.goal, "after update: ")
 408
 409
 410
 411     def _encrypt_and_encode(self):
 412         # this returns a Deferred that fires with a list of (sharedata,
 413         # sharenum) tuples. TODO: cache the ciphertext, only produce the
 414         # shares that we care about.
 415         self.log("_encrypt_and_encode")
 416
 417         self._status.set_status("Encrypting")
 418         started = time.time()
 419
 420         key = hashutil.ssk_readkey_data_hash(self.salt, self.readkey)
 421         enc = AES(key)
 422         crypttext = enc.process(self.newdata)
 423         assert len(crypttext) == len(self.newdata)
 424
 425         now = time.time()
 426         self._status.timings["encrypt"] = now - started
 427         started = now
 428
 429         # now apply FEC
 430
 431         self._status.set_status("Encoding")
 432         fec = codec.CRSEncoder()
 433         fec.set_params(self.segment_size,
 434                        self.required_shares, self.total_shares)
 435         piece_size = fec.get_block_size()
 436         crypttext_pieces = [None] * self.required_shares
 437         for i in range(len(crypttext_pieces)):
 438             offset = i * piece_size
 439             piece = crypttext[offset:offset+piece_size]
 440             piece = piece + "\x00"*(piece_size - len(piece)) # padding
 441             crypttext_pieces[i] = piece
 442             assert len(piece) == piece_size
 443
 444         d = fec.encode(crypttext_pieces)
 445         def _done_encoding(res):
 446             elapsed = time.time() - started
 447             self._status.timings["encode"] = elapsed
 448             return res
 449         d.addCallback(_done_encoding)
 450         return d
 451
 452     def _generate_shares(self, shares_and_shareids):
 453         # this sets self.shares and self.root_hash
 454         self.log("_generate_shares")
 455         self._status.set_status("Generating Shares")
 456         started = time.time()
 457
 458         # we should know these by now
 459         privkey = self._privkey
 460         encprivkey = self._encprivkey
 461         pubkey = self._pubkey
 462
 463         (shares, share_ids) = shares_and_shareids
 464
 465         assert len(shares) == len(share_ids)
 466         assert len(shares) == self.total_shares
 467         all_shares = {}
 468         block_hash_trees = {}
 469         share_hash_leaves = [None] * len(shares)
 470         for i in range(len(shares)):
 471             share_data = shares[i]
 472             shnum = share_ids[i]
 473             all_shares[shnum] = share_data
 474
 475             # build the block hash tree. SDMF has only one leaf.
 476             leaves = [hashutil.block_hash(share_data)]
 477             t = hashtree.HashTree(leaves)
 478             block_hash_trees[shnum] = block_hash_tree = list(t)
 479             share_hash_leaves[shnum] = t[0]
 480         for leaf in share_hash_leaves:
 481             assert leaf is not None
 482         share_hash_tree = hashtree.HashTree(share_hash_leaves)
 483         share_hash_chain = {}
 484         for shnum in range(self.total_shares):
 485             needed_hashes = share_hash_tree.needed_hashes(shnum)
 486             share_hash_chain[shnum] = dict( [ (i, share_hash_tree[i])
 487                                               for i in needed_hashes ] )
 488         root_hash = share_hash_tree[0]
 489         assert len(root_hash) == 32
 490         self.log("my new root_hash is %s" % base32.b2a(root_hash))
 491         self._new_version_info = (self._new_seqnum, root_hash, self.salt)
 492
 493         prefix = pack_prefix(self._new_seqnum, root_hash, self.salt,
 494                              self.required_shares, self.total_shares,
 495                              self.segment_size, len(self.newdata))
 496
 497         # now pack the beginning of the share. All shares are the same up
 498         # to the signature, then they have divergent share hash chains,
 499         # then completely different block hash trees + salt + share data,
 500         # then they all share the same encprivkey at the end. The sizes
 501         # of everything are the same for all shares.
 502
 503         sign_started = time.time()
 504         signature = privkey.sign(prefix)
 505         self._status.timings["sign"] = time.time() - sign_started
 506
 507         verification_key = pubkey.serialize()
 508
 509         final_shares = {}
 510         for shnum in range(self.total_shares):
 511             final_share = pack_share(prefix,
 512                                      verification_key,
 513                                      signature,
 514                                      share_hash_chain[shnum],
 515                                      block_hash_trees[shnum],
 516                                      all_shares[shnum],
 517                                      encprivkey)
 518             final_shares[shnum] = final_share
 519         elapsed = time.time() - started
 520         self._status.timings["pack"] = elapsed
 521         self.shares = final_shares
 522         self.root_hash = root_hash
 523
 524         # we also need to build up the version identifier for what we're
 525         # pushing. Extract the offsets from one of our shares.
 526         assert final_shares
 527         offsets = unpack_header(final_shares.values()[0])[-1]
 528         offsets_tuple = tuple( [(key,value) for key,value in offsets.items()] )
 529         verinfo = (self._new_seqnum, root_hash, self.salt,
 530                    self.segment_size, len(self.newdata),
 531                    self.required_shares, self.total_shares,
 532                    prefix, offsets_tuple)
 533         self.versioninfo = verinfo
 534
 535
 536
 537     def _send_shares(self, needed):
 538         self.log("_send_shares")
 539
 540         # we're finally ready to send out our shares. If we encounter any
 541         # surprises here, it's because somebody else is writing at the same
 542         # time. (Note: in the future, when we remove the _query_peers() step
 543         # and instead speculate about [or remember] which shares are where,
 544         # surprises here are *not* indications of UncoordinatedWriteError,
 545         # and we'll need to respond to them more gracefully.)
 546
 547         # needed is a set of (peerid, shnum) tuples. The first thing we do is
 548         # organize it by peerid.
 549
 550         peermap = DictOfSets()
 551         for (peerid, shnum) in needed:
 552             peermap.add(peerid, shnum)
 553
 554         # the next thing is to build up a bunch of test vectors. The
 555         # semantics of Publish are that we perform the operation if the world
 556         # hasn't changed since the ServerMap was constructed (more or less).
 557         # For every share we're trying to place, we create a test vector that
 558         # tests to see if the server*share still corresponds to the
 559         # map.
 560
 561         all_tw_vectors = {} # maps peerid to tw_vectors
 562         sm = self._servermap.servermap
 563
 564         for key in needed:
 565             (peerid, shnum) = key
 566
 567             if key in sm:
 568                 # an old version of that share already exists on the
 569                 # server, according to our servermap. We will create a
 570                 # request that attempts to replace it.
 571                 old_versionid, old_timestamp = sm[key]
 572                 (old_seqnum, old_root_hash, old_salt, old_segsize,
 573                  old_datalength, old_k, old_N, old_prefix,
 574                  old_offsets_tuple) = old_versionid
 575                 old_checkstring = pack_checkstring(old_seqnum,
 576                                                    old_root_hash,
 577                                                    old_salt)
 578                 testv = (0, len(old_checkstring), "eq", old_checkstring)
 579
 580             elif key in self.bad_share_checkstrings:
 581                 old_checkstring = self.bad_share_checkstrings[key]
 582                 testv = (0, len(old_checkstring), "eq", old_checkstring)
 583
 584             else:
 585                 # add a testv that requires the share not exist
 586
 587                 # Unfortunately, foolscap-0.2.5 has a bug in the way inbound
 588                 # constraints are handled. If the same object is referenced
 589                 # multiple times inside the arguments, foolscap emits a
 590                 # 'reference' token instead of a distinct copy of the
 591                 # argument. The bug is that these 'reference' tokens are not
 592                 # accepted by the inbound constraint code. To work around
 593                 # this, we need to prevent python from interning the
 594                 # (constant) tuple, by creating a new copy of this vector
 595                 # each time.
 596
 597                 # This bug is fixed in foolscap-0.2.6, and even though this
 598                 # version of Tahoe requires foolscap-0.3.1 or newer, we are
 599                 # supposed to be able to interoperate with older versions of
 600                 # Tahoe which are allowed to use older versions of foolscap,
 601                 # including foolscap-0.2.5 . In addition, I've seen other
 602                 # foolscap problems triggered by 'reference' tokens (see #541
 603                 # for details). So we must keep this workaround in place.
 604
 605                 #testv = (0, 1, 'eq', "")
 606                 testv = tuple([0, 1, 'eq', ""])
 607
 608             testvs = [testv]
 609             # the write vector is simply the share
 610             writev = [(0, self.shares[shnum])]
 611
 612             if peerid not in all_tw_vectors:
 613                 all_tw_vectors[peerid] = {}
 614                 # maps shnum to (testvs, writevs, new_length)
 615             assert shnum not in all_tw_vectors[peerid]
 616
 617             all_tw_vectors[peerid][shnum] = (testvs, writev, None)
 618
 619         # we read the checkstring back from each share, however we only use
 620         # it to detect whether there was a new share that we didn't know
 621         # about. The success or failure of the write will tell us whether
 622         # there was a collision or not. If there is a collision, the first
 623         # thing we'll do is update the servermap, which will find out what
 624         # happened. We could conceivably reduce a roundtrip by using the
 625         # readv checkstring to populate the servermap, but really we'd have
 626         # to read enough data to validate the signatures too, so it wouldn't
 627         # be an overall win.
 628         read_vector = [(0, struct.calcsize(SIGNED_PREFIX))]
 629
 630         # ok, send the messages!
 631         self.log("sending %d shares" % len(all_tw_vectors), level=log.NOISY)
 632         started = time.time()
 633         for (peerid, tw_vectors) in all_tw_vectors.items():
 634
 635             write_enabler = self._node.get_write_enabler(peerid)
 636             renew_secret = self._node.get_renewal_secret(peerid)
 637             cancel_secret = self._node.get_cancel_secret(peerid)
 638             secrets = (write_enabler, renew_secret, cancel_secret)
 639             shnums = tw_vectors.keys()
 640
 641             for shnum in shnums:
 642                 self.outstanding.add( (peerid, shnum) )
 643
 644             d = self._do_testreadwrite(peerid, secrets,
 645                                        tw_vectors, read_vector)
 646             d.addCallbacks(self._got_write_answer, self._got_write_error,
 647                            callbackArgs=(peerid, shnums, started),
 648                            errbackArgs=(peerid, shnums, started))
 649             d.addCallback(self.loop)
 650             d.addErrback(self._fatal_error)
 651
 652         self._update_status()
 653         self.log("%d shares sent" % len(all_tw_vectors), level=log.NOISY)
 654
 655     def _do_testreadwrite(self, peerid, secrets,
 656                           tw_vectors, read_vector):
 657         storage_index = self._storage_index
 658         ss = self.connections[peerid]
 659
 660         #print "SS[%s] is %s" % (idlib.shortnodeid_b2a(peerid), ss), ss.tracker.interfaceName
 661         d = ss.callRemote("slot_testv_and_readv_and_writev",
 662                           storage_index,
 663                           secrets,
 664                           tw_vectors,
 665                           read_vector)
 666         return d
 667
 668     def _got_write_answer(self, answer, peerid, shnums, started):
 669         lp = self.log("_got_write_answer from %s" %
 670                       idlib.shortnodeid_b2a(peerid))
 671         for shnum in shnums:
 672             self.outstanding.discard( (peerid, shnum) )
 673
 674         now = time.time()
 675         elapsed = now - started
 676         self._status.add_per_server_time(peerid, elapsed)
 677
 678         wrote, read_data = answer
 679
 680         surprise_shares = set(read_data.keys()) - set(shnums)
 681
 682         surprised = False
 683         for shnum in surprise_shares:
 684             # read_data is a dict mapping shnum to checkstring (SIGNED_PREFIX)
 685             checkstring = read_data[shnum][0]
 686             their_version_info = unpack_checkstring(checkstring)
 687             if their_version_info == self._new_version_info:
 688                 # they have the right share, somehow
 689
 690                 if (peerid,shnum) in self.goal:
 691                     # and we want them to have it, so we probably sent them a
 692                     # copy in an earlier write. This is ok, and avoids the
 693                     # #546 problem.
 694                     continue
 695
 696                 # They aren't in our goal, but they are still for the right
 697                 # version. Somebody else wrote them, and it's a convergent
 698                 # uncoordinated write. Pretend this is ok (don't be
 699                 # surprised), since I suspect there's a decent chance that
 700                 # we'll hit this in normal operation.
 701                 continue
 702
 703             else:
 704                 # the new shares are of a different version
 705                 if peerid in self._servermap.reachable_peers:
 706                     # we asked them about their shares, so we had knowledge
 707                     # of what they used to have. Any surprising shares must
 708                     # have come from someone else, so UCW.
 709                     surprised = True
 710                 else:
 711                     # we didn't ask them, and now we've discovered that they
 712                     # have a share we didn't know about. This indicates that
 713                     # mapupdate should have wokred harder and asked more
 714                     # servers before concluding that it knew about them all.
 715
 716                     # signal UCW, but make sure to ask this peer next time,
 717                     # so we'll remember to update it if/when we retry.
 718                     surprised = True
 719                     # TODO: ask this peer next time. I don't yet have a good
 720                     # way to do this. Two insufficient possibilities are:
 721                     #
 722                     # self._servermap.add_new_share(peerid, shnum, verinfo, now)
 723                     #  but that requires fetching/validating/parsing the whole
 724                     #  version string, and all we have is the checkstring
 725                     # self._servermap.mark_bad_share(peerid, shnum, checkstring)
 726                     #  that will make publish overwrite the share next time,
 727                     #  but it won't re-query the server, and it won't make
 728                     #  mapupdate search further
 729
 730                     # TODO later: when publish starts, do
 731                     # servermap.get_best_version(), extract the seqnum,
 732                     # subtract one, and store as highest-replaceable-seqnum.
 733                     # Then, if this surprise-because-we-didn't-ask share is
 734                     # of highest-replaceable-seqnum or lower, we're allowed
 735                     # to replace it: send out a new writev (or rather add it
 736                     # to self.goal and loop).
 737                     pass
 738
 739                 surprised = True
 740
 741         if surprised:
 742             self.log("they had shares %s that we didn't know about" %
 743                      (list(surprise_shares),),
 744                      parent=lp, level=log.WEIRD, umid="un9CSQ")
 745             self.surprised = True
 746
 747         if not wrote:
 748             # TODO: there are two possibilities. The first is that the server
 749             # is full (or just doesn't want to give us any room), which means
 750             # we shouldn't ask them again, but is *not* an indication of an
 751             # uncoordinated write. The second is that our testv failed, which
 752             # *does* indicate an uncoordinated write. We currently don't have
 753             # a way to tell these two apart (in fact, the storage server code
 754             # doesn't have the option of refusing our share).
 755             #
 756             # If the server is full, mark the peer as bad (so we don't ask
 757             # them again), but don't set self.surprised. The loop() will find
 758             # a new server.
 759             #
 760             # If the testv failed, log it, set self.surprised, but don't
 761             # bother adding to self.bad_peers .
 762
 763             self.log("our testv failed, so the write did not happen",
 764                      parent=lp, level=log.WEIRD, umid="8sc26g")
 765             self.surprised = True
 766             self.bad_peers.add(peerid) # don't ask them again
 767             # use the checkstring to add information to the log message
 768             for (shnum,readv) in read_data.items():
 769                 checkstring = readv[0]
 770                 (other_seqnum,
 771                  other_roothash,
 772                  other_salt) = unpack_checkstring(checkstring)
 773                 expected_version = self._servermap.version_on_peer(peerid,
 774                                                                    shnum)
 775                 if expected_version:
 776                     (seqnum, root_hash, IV, segsize, datalength, k, N, prefix,
 777                      offsets_tuple) = expected_version
 778                     self.log("somebody modified the share on us:"
 779                              " shnum=%d: I thought they had #%d:R=%s,"
 780                              " but testv reported #%d:R=%s" %
 781                              (shnum,
 782                               seqnum, base32.b2a(root_hash)[:4],
 783                               other_seqnum, base32.b2a(other_roothash)[:4]),
 784                              parent=lp, level=log.NOISY)
 785                 # if expected_version==None, then we didn't expect to see a
 786                 # share on that peer, and the 'surprise_shares' clause above
 787                 # will have logged it.
 788             # self.loop() will take care of finding new homes
 789             return
 790
 791         for shnum in shnums:
 792             self.placed.add( (peerid, shnum) )
 793             # and update the servermap
 794             self._servermap.add_new_share(peerid, shnum,
 795                                           self.versioninfo, started)
 796
 797         # self.loop() will take care of checking to see if we're done
 798         return
 799
 800     def _got_write_error(self, f, peerid, shnums, started):
 801         for shnum in shnums:
 802             self.outstanding.discard( (peerid, shnum) )
 803         self.bad_peers.add(peerid)
 804         self.log(format="error while writing shares %(shnums)s to peerid %(peerid)s",
 805                  shnums=list(shnums), peerid=idlib.shortnodeid_b2a(peerid),
 806                  failure=f,
 807                  level=log.UNUSUAL)
 808         # self.loop() will take care of checking to see if we're done
 809         return
 810
 811
 812     def _done(self, res):
 813         if not self._running:
 814             return
 815         self._running = False
 816         now = time.time()
 817         self._status.timings["total"] = now - self._started
 818         self._status.set_active(False)
 819         if isinstance(res, failure.Failure):
 820             self.log("Publish done, with failure", failure=res,
 821                      level=log.WEIRD, umid="nRsR9Q")
 822             self._status.set_status("Failed")
 823         elif self.surprised:
 824             self.log("Publish done, UncoordinatedWriteError", level=log.UNUSUAL)
 825             self._status.set_status("UncoordinatedWriteError")
 826             # deliver a failure
 827             res = failure.Failure(UncoordinatedWriteError())
 828             # TODO: recovery
 829         else:
 830             self.log("Publish done, success")
 831             self._status.set_status("Done")
 832             self._status.set_progress(1.0)
 833         eventually(self.done_deferred.callback, res)
 834
 835