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