src/allmydata/mutable/checker.py

   1
   2 from twisted.internet import defer
   3 from twisted.python import failure
   4 from allmydata import hashtree
   5 from allmydata.uri import from_string
   6 from allmydata.util import hashutil, base32, idlib, log
   7 from allmydata.check_results import CheckAndRepairResults, CheckResults
   8
   9 from common import MODE_CHECK, CorruptShareError
  10 from servermap import ServerMap, ServermapUpdater
  11 from layout import unpack_share, SIGNED_PREFIX_LENGTH
  12
  13 class MutableChecker:
  14
  15     def __init__(self, node, monitor):
  16         self._node = node
  17         self._monitor = monitor
  18         self.bad_shares = [] # list of (nodeid,shnum,failure)
  19         self._storage_index = self._node.get_storage_index()
  20         self.results = CheckResults(from_string(node.get_uri()), self._storage_index)
  21         self.need_repair = False
  22         self.responded = set() # set of (binary) nodeids
  23
  24     def check(self, verify=False):
  25         servermap = ServerMap()
  26         u = ServermapUpdater(self._node, self._monitor, servermap, MODE_CHECK)
  27         history = self._node._client.get_history()
  28         if history:
  29             history.notify_mapupdate(u.get_status())
  30         d = u.update()
  31         d.addCallback(self._got_mapupdate_results)
  32         if verify:
  33             d.addCallback(self._verify_all_shares)
  34         d.addCallback(lambda res: servermap)
  35         d.addCallback(self._fill_checker_results, self.results)
  36         d.addCallback(lambda res: self.results)
  37         return d
  38
  39     def _got_mapupdate_results(self, servermap):
  40         # the file is healthy if there is exactly one recoverable version, it
  41         # has at least N distinct shares, and there are no unrecoverable
  42         # versions: all existing shares will be for the same version.
  43         self._monitor.raise_if_cancelled()
  44         self.best_version = None
  45         num_recoverable = len(servermap.recoverable_versions())
  46         if num_recoverable:
  47             self.best_version = servermap.best_recoverable_version()
  48
  49         if servermap.unrecoverable_versions():
  50             self.need_repair = True
  51         if num_recoverable != 1:
  52             self.need_repair = True
  53         if self.best_version:
  54             available_shares = servermap.shares_available()
  55             (num_distinct_shares, k, N) = available_shares[self.best_version]
  56             if num_distinct_shares < N:
  57                 self.need_repair = True
  58
  59         return servermap
  60
  61     def _verify_all_shares(self, servermap):
  62         # read every byte of each share
  63         if not self.best_version:
  64             return
  65         versionmap = servermap.make_versionmap()
  66         shares = versionmap[self.best_version]
  67         (seqnum, root_hash, IV, segsize, datalength, k, N, prefix,
  68          offsets_tuple) = self.best_version
  69         offsets = dict(offsets_tuple)
  70         readv = [ (0, offsets["EOF"]) ]
  71         dl = []
  72         for (shnum, peerid, timestamp) in shares:
  73             ss = servermap.connections[peerid]
  74             d = self._do_read(ss, peerid, self._storage_index, [shnum], readv)
  75             d.addCallback(self._got_answer, peerid, servermap)
  76             dl.append(d)
  77         return defer.DeferredList(dl, fireOnOneErrback=True, consumeErrors=True)
  78
  79     def _do_read(self, ss, peerid, storage_index, shnums, readv):
  80         # isolate the callRemote to a separate method, so tests can subclass
  81         # Publish and override it
  82         d = ss.callRemote("slot_readv", storage_index, shnums, readv)
  83         return d
  84
  85     def _got_answer(self, datavs, peerid, servermap):
  86         for shnum,datav in datavs.items():
  87             data = datav[0]
  88             try:
  89                 self._got_results_one_share(shnum, peerid, data)
  90             except CorruptShareError:
  91                 f = failure.Failure()
  92                 self.need_repair = True
  93                 self.bad_shares.append( (peerid, shnum, f) )
  94                 prefix = data[:SIGNED_PREFIX_LENGTH]
  95                 servermap.mark_bad_share(peerid, shnum, prefix)
  96                 ss = servermap.connections[peerid]
  97                 self.notify_server_corruption(ss, shnum, str(f.value))
  98
  99     def check_prefix(self, peerid, shnum, data):
 100         (seqnum, root_hash, IV, segsize, datalength, k, N, prefix,
 101          offsets_tuple) = self.best_version
 102         got_prefix = data[:SIGNED_PREFIX_LENGTH]
 103         if got_prefix != prefix:
 104             raise CorruptShareError(peerid, shnum,
 105                                     "prefix mismatch: share changed while we were reading it")
 106
 107     def _got_results_one_share(self, shnum, peerid, data):
 108         self.check_prefix(peerid, shnum, data)
 109
 110         # the [seqnum:signature] pieces are validated by _compare_prefix,
 111         # which checks their signature against the pubkey known to be
 112         # associated with this file.
 113
 114         (seqnum, root_hash, IV, k, N, segsize, datalen, pubkey, signature,
 115          share_hash_chain, block_hash_tree, share_data,
 116          enc_privkey) = unpack_share(data)
 117
 118         # validate [share_hash_chain,block_hash_tree,share_data]
 119
 120         leaves = [hashutil.block_hash(share_data)]
 121         t = hashtree.HashTree(leaves)
 122         if list(t) != block_hash_tree:
 123             raise CorruptShareError(peerid, shnum, "block hash tree failure")
 124         share_hash_leaf = t[0]
 125         t2 = hashtree.IncompleteHashTree(N)
 126         # root_hash was checked by the signature
 127         t2.set_hashes({0: root_hash})
 128         try:
 129             t2.set_hashes(hashes=share_hash_chain,
 130                           leaves={shnum: share_hash_leaf})
 131         except (hashtree.BadHashError, hashtree.NotEnoughHashesError,
 132                 IndexError), e:
 133             msg = "corrupt hashes: %s" % (e,)
 134             raise CorruptShareError(peerid, shnum, msg)
 135
 136         # validate enc_privkey: only possible if we have a write-cap
 137         if not self._node.is_readonly():
 138             alleged_privkey_s = self._node._decrypt_privkey(enc_privkey)
 139             alleged_writekey = hashutil.ssk_writekey_hash(alleged_privkey_s)
 140             if alleged_writekey != self._node.get_writekey():
 141                 raise CorruptShareError(peerid, shnum, "invalid privkey")
 142
 143     def notify_server_corruption(self, ss, shnum, reason):
 144         ss.callRemoteOnly("advise_corrupt_share",
 145                           "mutable", self._storage_index, shnum, reason)
 146
 147     def _count_shares(self, smap, version):
 148         available_shares = smap.shares_available()
 149         (num_distinct_shares, k, N) = available_shares[version]
 150         counters = {}
 151         counters["count-shares-good"] = num_distinct_shares
 152         counters["count-shares-needed"] = k
 153         counters["count-shares-expected"] = N
 154         good_hosts = smap.all_peers_for_version(version)
 155         counters["count-good-share-hosts"] = len(good_hosts)
 156         vmap = smap.make_versionmap()
 157         counters["count-wrong-shares"] = sum([len(shares)
 158                                           for verinfo,shares in vmap.items()
 159                                           if verinfo != version])
 160
 161         return counters
 162
 163     def _fill_checker_results(self, smap, r):
 164         self._monitor.raise_if_cancelled()
 165         r.set_servermap(smap.copy())
 166         healthy = True
 167         data = {}
 168         report = []
 169         summary = []
 170         vmap = smap.make_versionmap()
 171         recoverable = smap.recoverable_versions()
 172         unrecoverable = smap.unrecoverable_versions()
 173         data["count-recoverable-versions"] = len(recoverable)
 174         data["count-unrecoverable-versions"] = len(unrecoverable)
 175
 176         if recoverable:
 177             report.append("Recoverable Versions: " +
 178                           "/".join(["%d*%s" % (len(vmap[v]),
 179                                                smap.summarize_version(v))
 180                                     for v in recoverable]))
 181         if unrecoverable:
 182             report.append("Unrecoverable Versions: " +
 183                           "/".join(["%d*%s" % (len(vmap[v]),
 184                                                smap.summarize_version(v))
 185                                     for v in unrecoverable]))
 186         if smap.unrecoverable_versions():
 187             healthy = False
 188             summary.append("some versions are unrecoverable")
 189             report.append("Unhealthy: some versions are unrecoverable")
 190         if len(recoverable) == 0:
 191             healthy = False
 192             summary.append("no versions are recoverable")
 193             report.append("Unhealthy: no versions are recoverable")
 194         if len(recoverable) > 1:
 195             healthy = False
 196             summary.append("multiple versions are recoverable")
 197             report.append("Unhealthy: there are multiple recoverable versions")
 198
 199         needs_rebalancing = False
 200         if recoverable:
 201             best_version = smap.best_recoverable_version()
 202             report.append("Best Recoverable Version: " +
 203                           smap.summarize_version(best_version))
 204             counters = self._count_shares(smap, best_version)
 205             data.update(counters)
 206             s = counters["count-shares-good"]
 207             k = counters["count-shares-needed"]
 208             N = counters["count-shares-expected"]
 209             if s < N:
 210                 healthy = False
 211                 report.append("Unhealthy: best version has only %d shares "
 212                               "(encoding is %d-of-%d)" % (s, k, N))
 213                 summary.append("%d shares (enc %d-of-%d)" % (s, k, N))
 214             hosts = smap.all_peers_for_version(best_version)
 215             needs_rebalancing = bool( len(hosts) < N )
 216         elif unrecoverable:
 217             healthy = False
 218             # find a k and N from somewhere
 219             first = list(unrecoverable)[0]
 220             # not exactly the best version, but that doesn't matter too much
 221             data.update(self._count_shares(smap, first))
 222             # leave needs_rebalancing=False: the file being unrecoverable is
 223             # the bigger problem
 224         else:
 225             # couldn't find anything at all
 226             data["count-shares-good"] = 0
 227             data["count-shares-needed"] = 3 # arbitrary defaults
 228             data["count-shares-expected"] = 10
 229             data["count-good-share-hosts"] = 0
 230             data["count-wrong-shares"] = 0
 231
 232         if self.bad_shares:
 233             data["count-corrupt-shares"] = len(self.bad_shares)
 234             data["list-corrupt-shares"] = locators = []
 235             report.append("Corrupt Shares:")
 236             summary.append("Corrupt Shares:")
 237             for (peerid, shnum, f) in sorted(self.bad_shares):
 238                 locators.append( (peerid, self._storage_index, shnum) )
 239                 s = "%s-sh%d" % (idlib.shortnodeid_b2a(peerid), shnum)
 240                 if f.check(CorruptShareError):
 241                     ft = f.value.reason
 242                 else:
 243                     ft = str(f)
 244                 report.append(" %s: %s" % (s, ft))
 245                 summary.append(s)
 246                 p = (peerid, self._storage_index, shnum, f)
 247                 r.problems.append(p)
 248                 msg = ("CorruptShareError during mutable verify, "
 249                        "peerid=%(peerid)s, si=%(si)s, shnum=%(shnum)d, "
 250                        "where=%(where)s")
 251                 log.msg(format=msg, peerid=idlib.nodeid_b2a(peerid),
 252                         si=base32.b2a(self._storage_index),
 253                         shnum=shnum,
 254                         where=ft,
 255                         level=log.WEIRD, umid="EkK8QA")
 256         else:
 257             data["count-corrupt-shares"] = 0
 258             data["list-corrupt-shares"] = []
 259
 260         sharemap = {}
 261         for verinfo in vmap:
 262             for (shnum, peerid, timestamp) in vmap[verinfo]:
 263                 shareid = "%s-sh%d" % (smap.summarize_version(verinfo), shnum)
 264                 if shareid not in sharemap:
 265                     sharemap[shareid] = []
 266                 sharemap[shareid].append(peerid)
 267         data["sharemap"] = sharemap
 268         data["servers-responding"] = list(smap.reachable_peers)
 269
 270         r.set_healthy(healthy)
 271         r.set_recoverable(bool(recoverable))
 272         r.set_needs_rebalancing(needs_rebalancing)
 273         r.set_data(data)
 274         if healthy:
 275             r.set_summary("Healthy")
 276         else:
 277             r.set_summary("Unhealthy: " + " ".join(summary))
 278         r.set_report(report)
 279
 280
 281 class MutableCheckAndRepairer(MutableChecker):
 282     def __init__(self, node, monitor):
 283         MutableChecker.__init__(self, node, monitor)
 284         self.cr_results = CheckAndRepairResults(self._storage_index)
 285         self.cr_results.pre_repair_results = self.results
 286         self.need_repair = False
 287
 288     def check(self, verify=False):
 289         d = MutableChecker.check(self, verify)
 290         d.addCallback(self._maybe_repair)
 291         d.addCallback(lambda res: self.cr_results)
 292         return d
 293
 294     def _maybe_repair(self, res):
 295         self._monitor.raise_if_cancelled()
 296         if not self.need_repair:
 297             self.cr_results.post_repair_results = self.results
 298             return
 299         self.cr_results.repair_attempted = True
 300         d = self._node.repair(self.results)
 301         def _repair_finished(repair_results):
 302             self.cr_results.repair_successful = True
 303             r = CheckResults(from_string(self._node.get_uri()), self._storage_index)
 304             self.cr_results.post_repair_results = r
 305             self._fill_checker_results(repair_results.servermap, r)
 306             self.cr_results.repair_results = repair_results # TODO?
 307         def _repair_error(f):
 308             # I'm not sure if I want to pass through a failure or not.
 309             self.cr_results.repair_successful = False
 310             self.cr_results.repair_failure = f # TODO?
 311             #self.cr_results.post_repair_results = ??
 312             return f
 313         d.addCallbacks(_repair_finished, _repair_error)
 314         return d