Ticket #393: 393status10.dpatch

Thu Jun 24 16:46:37 PDT 2010  Kevan Carstensen <kevan@isnotajoke.com>
  * Misc. changes to support the work I'm doing
  
      - Add a notion of file version number to interfaces.py
      - Alter mutable file node interfaces to have a notion of version,
        though this may be changed later.
      - Alter mutable/filenode.py to conform to these changes.
      - Add a salt hasher to util/hashutil.py

Thu Jun 24 16:48:33 PDT 2010  Kevan Carstensen <kevan@isnotajoke.com>
  * nodemaker.py: create MDMF files when asked to

Thu Jun 24 16:49:05 PDT 2010  Kevan Carstensen <kevan@isnotajoke.com>
  * storage/server.py: minor code cleanup

Thu Jun 24 16:49:24 PDT 2010  Kevan Carstensen <kevan@isnotajoke.com>
  * test/test_mutable.py: alter some tests that were failing due to MDMF; minor code cleanup.

Fri Jun 25 17:35:20 PDT 2010  Kevan Carstensen <kevan@isnotajoke.com>
  * test/test_mutable.py: change the definition of corrupt() to work with MDMF as well as SDMF files, change users of corrupt to use the new definition

Sat Jun 26 16:40:33 PDT 2010  Kevan Carstensen <kevan@isnotajoke.com>
  * Add MDMF reader and writer
  
  The MDMF/SDMF reader and MDMF writer are similar to the object proxies
  that exist for immutable files. They abstract away details of
  connection, state, and caching from their callers (in this case, the
  download, servermap updater, and uploader), and expose methods to get
  and set information on the remote server.
  
  MDMFSlotReadProxy reads a mutable file from the server, doing the right
  thing (in most cases) regardless of whether the file is MDMF or SDMF. It
  allows callers to tell it how to batch and flush reads.
  
  MDMFSlotWriteProxy writes an MDMF mutable file to a server.
  
  This patch also includes tests for MDMFSlotReadProxy and
  MDMFSlotWriteProxy.

Sat Jun 26 16:41:18 PDT 2010  Kevan Carstensen <kevan@isnotajoke.com>
  * Alter the ServermapUpdater to find MDMF files
  
  The servermapupdater should find MDMF files on a grid in the same way
  that it finds SDMF files. This patch makes it do that.

Sat Jun 26 16:42:04 PDT 2010  Kevan Carstensen <kevan@isnotajoke.com>
  * Make a segmented mutable uploader
  
  The mutable file uploader should be able to publish files with one
  segment and files with multiple segments. This patch makes it do that.
  This is still incomplete, and rather ugly -- I need to flesh out error
  handling, I need to write tests, and I need to remove some of the uglier
  kludges in the process before I can call this done.

Sat Jun 26 16:43:14 PDT 2010  Kevan Carstensen <kevan@isnotajoke.com>
  * Write a segmented mutable downloader
  
  The segmented mutable downloader can deal with MDMF files (files with
  one or more segments in MDMF format) and SDMF files (files with one
  segment in SDMF format). It is backwards compatible with the old
  file format.
  
  This patch also contains tests for the segmented mutable downloader.

New patches:

[Misc. changes to support the work I'm doing
Kevan Carstensen <kevan@isnotajoke.com>**20100624234637
 Ignore-this: fdd18fa8cc05f4b4b15ff53ee24a1819
 
     - Add a notion of file version number to interfaces.py
     - Alter mutable file node interfaces to have a notion of version,
       though this may be changed later.
     - Alter mutable/filenode.py to conform to these changes.
     - Add a salt hasher to util/hashutil.py
] {
hunk ./src/allmydata/interfaces.py 7
      ChoiceOf, IntegerConstraint, Any, RemoteInterface, Referenceable
 
 HASH_SIZE=32
+SALT_SIZE=16
+
+SDMF_VERSION=0
+MDMF_VERSION=1
 
 Hash = StringConstraint(maxLength=HASH_SIZE,
                         minLength=HASH_SIZE)# binary format 32-byte SHA256 hash
hunk ./src/allmydata/interfaces.py 811
         writer-visible data using this writekey.
         """
 
+    def set_version(version):
+        """Tahoe-LAFS supports SDMF and MDMF mutable files. By default,
+        we upload in SDMF for reasons of compatibility. If you want to
+        change this, set_version will let you do that.
+
+        To say that this file should be uploaded in SDMF, pass in a 0. To
+        say that the file should be uploaded as MDMF, pass in a 1.
+        """
+
+    def get_version():
+        """Returns the mutable file protocol version."""
+
 class NotEnoughSharesError(Exception):
     """Download was unable to get enough shares"""
 
hunk ./src/allmydata/mutable/filenode.py 8
 from twisted.internet import defer, reactor
 from foolscap.api import eventually
 from allmydata.interfaces import IMutableFileNode, \
-     ICheckable, ICheckResults, NotEnoughSharesError
+     ICheckable, ICheckResults, NotEnoughSharesError, MDMF_VERSION, SDMF_VERSION
 from allmydata.util import hashutil, log
 from allmydata.util.assertutil import precondition
 from allmydata.uri import WriteableSSKFileURI, ReadonlySSKFileURI
hunk ./src/allmydata/mutable/filenode.py 67
         self._sharemap = {} # known shares, shnum-to-[nodeids]
         self._cache = ResponseCache()
         self._most_recent_size = None
+        # filled in after __init__ if we're being created for the first time;
+        # filled in by the servermap updater before publishing, otherwise.
+        # set to this default value in case neither of those things happen,
+        # or in case the servermap can't find any shares to tell us what
+        # to publish as.
+        # TODO: Set this back to None, and find out why the tests fail
+        #       with it set to None.
+        self._protocol_version = SDMF_VERSION
 
         # all users of this MutableFileNode go through the serializer. This
         # takes advantage of the fact that Deferreds discard the callbacks
hunk ./src/allmydata/mutable/filenode.py 472
     def _did_upload(self, res, size):
         self._most_recent_size = size
         return res
+
+
+    def set_version(self, version):
+        # I can be set in two ways:
+        #  1. When the node is created.
+        #  2. (for an existing share) when the Servermap is updated 
+        #     before I am read.
+        assert version in (MDMF_VERSION, SDMF_VERSION)
+        self._protocol_version = version
+
+
+    def get_version(self):
+        return self._protocol_version
hunk ./src/allmydata/util/hashutil.py 90
 MUTABLE_READKEY_TAG = "allmydata_mutable_writekey_to_readkey_v1"
 MUTABLE_DATAKEY_TAG = "allmydata_mutable_readkey_to_datakey_v1"
 MUTABLE_STORAGEINDEX_TAG = "allmydata_mutable_readkey_to_storage_index_v1"
+MUTABLE_SALT_TAG = "allmydata_mutable_segment_salt_v1"
 
 # dirnodes
 DIRNODE_CHILD_WRITECAP_TAG = "allmydata_mutable_writekey_and_salt_to_dirnode_child_capkey_v1"
hunk ./src/allmydata/util/hashutil.py 134
 def plaintext_segment_hasher():
     return tagged_hasher(PLAINTEXT_SEGMENT_TAG)
 
+def mutable_salt_hash(data):
+    return tagged_hash(MUTABLE_SALT_TAG, data)
+def mutable_salt_hasher():
+    return tagged_hasher(MUTABLE_SALT_TAG)
+
 KEYLEN = 16
 IVLEN = 16
 
}
[nodemaker.py: create MDMF files when asked to
Kevan Carstensen <kevan@isnotajoke.com>**20100624234833
 Ignore-this: 26c16aaca9ddab7a7ce37a4530bc970
] {
hunk ./src/allmydata/nodemaker.py 3
 import weakref
 from zope.interface import implements
-from allmydata.interfaces import INodeMaker
+from allmydata.util.assertutil import precondition
+from allmydata.interfaces import INodeMaker, MustBeDeepImmutableError, \
+                                 SDMF_VERSION, MDMF_VERSION
 from allmydata.immutable.filenode import ImmutableFileNode, LiteralFileNode
 from allmydata.immutable.upload import Data
 from allmydata.mutable.filenode import MutableFileNode
hunk ./src/allmydata/nodemaker.py 92
             return self._create_dirnode(filenode)
         return None
 
-    def create_mutable_file(self, contents=None, keysize=None):
+    def create_mutable_file(self, contents=None, keysize=None,
+                            version=SDMF_VERSION):
         n = MutableFileNode(self.storage_broker, self.secret_holder,
                             self.default_encoding_parameters, self.history)
hunk ./src/allmydata/nodemaker.py 96
+        n.set_version(version)
         d = self.key_generator.generate(keysize)
         d.addCallback(n.create_with_keys, contents)
         d.addCallback(lambda res: n)
hunk ./src/allmydata/nodemaker.py 102
         return d
 
-    def create_new_mutable_directory(self, initial_children={}):
+    def create_new_mutable_directory(self, initial_children={},
+                                     version=SDMF_VERSION):
+        # initial_children must have metadata (i.e. {} instead of None)
+        for (name, (node, metadata)) in initial_children.iteritems():
+            precondition(isinstance(metadata, dict),
+                         "create_new_mutable_directory requires metadata to be a dict, not None", metadata)
+            node.raise_error()
         d = self.create_mutable_file(lambda n:
hunk ./src/allmydata/nodemaker.py 110
-                                     pack_children(n, initial_children))
+                                     pack_children(n, initial_children),
+                                     version)
         d.addCallback(self._create_dirnode)
         return d
 
}
[storage/server.py: minor code cleanup
Kevan Carstensen <kevan@isnotajoke.com>**20100624234905
 Ignore-this: 2358c531c39e48d3c8e56b62b5768228
] {
hunk ./src/allmydata/storage/server.py 569
                                          self)
         return share
 
-    def remote_slot_readv(self, storage_index, shares, readv):
+    def remote_slot_readv(self, storage_index, shares, readvs):
         start = time.time()
         self.count("readv")
         si_s = si_b2a(storage_index)
hunk ./src/allmydata/storage/server.py 590
             if sharenum in shares or not shares:
                 filename = os.path.join(bucketdir, sharenum_s)
                 msf = MutableShareFile(filename, self)
-                datavs[sharenum] = msf.readv(readv)
+                datavs[sharenum] = msf.readv(readvs)
         log.msg("returning shares %s" % (datavs.keys(),),
                 facility="tahoe.storage", level=log.NOISY, parent=lp)
         self.add_latency("readv", time.time() - start)
}
[test/test_mutable.py: alter some tests that were failing due to MDMF; minor code cleanup.
Kevan Carstensen <kevan@isnotajoke.com>**20100624234924
 Ignore-this: afb86ec1fbdbfe1a5ef6f46f350273c0
] {
hunk ./src/allmydata/test/test_mutable.py 151
             chr(ord(original[byte_offset]) ^ 0x01) +
             original[byte_offset+1:])
 
+def add_two(original, byte_offset):
+    # It isn't enough to simply flip the bit for the version number,
+    # because 1 is a valid version number. So we add two instead.
+    return (original[:byte_offset] +
+            chr(ord(original[byte_offset]) ^ 0x02) +
+            original[byte_offset+1:])
+
 def corrupt(res, s, offset, shnums_to_corrupt=None, offset_offset=0):
     # if shnums_to_corrupt is None, corrupt all shares. Otherwise it is a
     # list of shnums to corrupt.
hunk ./src/allmydata/test/test_mutable.py 187
                 real_offset = offset1
             real_offset = int(real_offset) + offset2 + offset_offset
             assert isinstance(real_offset, int), offset
-            shares[shnum] = flip_bit(data, real_offset)
+            if offset1 == 0: # verbyte
+                f = add_two
+            else:
+                f = flip_bit
+            shares[shnum] = f(data, real_offset)
     return res
 
 def make_storagebroker(s=None, num_peers=10):
hunk ./src/allmydata/test/test_mutable.py 423
         d.addCallback(_created)
         return d
 
+
     def test_modify_backoffer(self):
         def _modifier(old_contents, servermap, first_time):
             return old_contents + "line2"
hunk ./src/allmydata/test/test_mutable.py 658
         d.addCallback(_created)
         return d
 
+
     def _copy_shares(self, ignored, index):
         shares = self._storage._peers
         # we need a deep copy
}
[test/test_mutable.py: change the definition of corrupt() to work with MDMF as well as SDMF files, change users of corrupt to use the new definition
Kevan Carstensen <kevan@isnotajoke.com>**20100626003520
 Ignore-this: 836e59e2fde0535f6b4bea3468dc8244
] {
hunk ./src/allmydata/test/test_mutable.py 168
                 and shnum not in shnums_to_corrupt):
                 continue
             data = shares[shnum]
-            (version,
-             seqnum,
-             root_hash,
-             IV,
-             k, N, segsize, datalen,
-             o) = unpack_header(data)
-            if isinstance(offset, tuple):
-                offset1, offset2 = offset
-            else:
-                offset1 = offset
-                offset2 = 0
-            if offset1 == "pubkey":
-                real_offset = 107
-            elif offset1 in o:
-                real_offset = o[offset1]
-            else:
-                real_offset = offset1
-            real_offset = int(real_offset) + offset2 + offset_offset
-            assert isinstance(real_offset, int), offset
-            if offset1 == 0: # verbyte
-                f = add_two
-            else:
-                f = flip_bit
-            shares[shnum] = f(data, real_offset)
-    return res
+            # We're feeding the reader all of the share data, so it
+            # won't need to use the rref that we didn't provide, nor the
+            # storage index that we didn't provide. We do this because
+            # the reader will work for both MDMF and SDMF.
+            reader = MDMFSlotReadProxy(None, None, shnum, data)
+            # We need to get the offsets for the next part.
+            d = reader.get_verinfo()
+            def _do_corruption(verinfo, data, shnum):
+                (seqnum,
+                 root_hash,
+                 IV,
+                 segsize,
+                 datalen,
+                 k, n, prefix, o) = verinfo
+                if isinstance(offset, tuple):
+                    offset1, offset2 = offset
+                else:
+                    offset1 = offset
+                    offset2 = 0
+                if offset1 == "pubkey":
+                    real_offset = 107
+                elif offset1 in o:
+                    real_offset = o[offset1]
+                else:
+                    real_offset = offset1
+                real_offset = int(real_offset) + offset2 + offset_offset
+                assert isinstance(real_offset, int), offset
+                if offset1 == 0: # verbyte
+                    f = add_two
+                else:
+                    f = flip_bit
+                shares[shnum] = f(data, real_offset)
+            d.addCallback(_do_corruption, data, shnum)
+            ds.append(d)
+    dl = defer.DeferredList(ds)
+    dl.addCallback(lambda ignored: res)
+    return dl
 
 def make_storagebroker(s=None, num_peers=10):
     if not s:
hunk ./src/allmydata/test/test_mutable.py 1177
         return d
 
     def test_download_fails(self):
-        corrupt(None, self._storage, "signature")
-        d = self.shouldFail(UnrecoverableFileError, "test_download_anyway",
+        d = corrupt(None, self._storage, "signature")
+        d.addCallback(lambda ignored:
+            self.shouldFail(UnrecoverableFileError, "test_download_anyway",
                             "no recoverable versions",
                             self._fn.download_best_version)
         return d
hunk ./src/allmydata/test/test_mutable.py 1232
         return d
 
     def test_check_all_bad_sig(self):
-        corrupt(None, self._storage, 1) # bad sig
-        d = self._fn.check(Monitor())
+        d = corrupt(None, self._storage, 1) # bad sig
+        d.addCallback(lambda ignored:
+            self._fn.check(Monitor()))
         d.addCallback(self.check_bad, "test_check_all_bad_sig")
         return d
 
hunk ./src/allmydata/test/test_mutable.py 1239
     def test_check_all_bad_blocks(self):
-        corrupt(None, self._storage, "share_data", [9]) # bad blocks
+        d = corrupt(None, self._storage, "share_data", [9]) # bad blocks
         # the Checker won't notice this.. it doesn't look at actual data
hunk ./src/allmydata/test/test_mutable.py 1241
-        d = self._fn.check(Monitor())
+        d.addCallback(lambda ignored:
+            self._fn.check(Monitor()))
         d.addCallback(self.check_good, "test_check_all_bad_blocks")
         return d
 
hunk ./src/allmydata/test/test_mutable.py 1252
         return d
 
     def test_verify_all_bad_sig(self):
-        corrupt(None, self._storage, 1) # bad sig
-        d = self._fn.check(Monitor(), verify=True)
+        d = corrupt(None, self._storage, 1) # bad sig
+        d.addCallback(lambda ignored:
+            self._fn.check(Monitor(), verify=True))
         d.addCallback(self.check_bad, "test_verify_all_bad_sig")
         return d
 
hunk ./src/allmydata/test/test_mutable.py 1259
     def test_verify_one_bad_sig(self):
-        corrupt(None, self._storage, 1, [9]) # bad sig
-        d = self._fn.check(Monitor(), verify=True)
+        d = corrupt(None, self._storage, 1, [9]) # bad sig
+        d.addCallback(lambda ignored:
+            self._fn.check(Monitor(), verify=True))
         d.addCallback(self.check_bad, "test_verify_one_bad_sig")
         return d
 
hunk ./src/allmydata/test/test_mutable.py 1266
     def test_verify_one_bad_block(self):
-        corrupt(None, self._storage, "share_data", [9]) # bad blocks
+        d = corrupt(None, self._storage, "share_data", [9]) # bad blocks
         # the Verifier *will* notice this, since it examines every byte
hunk ./src/allmydata/test/test_mutable.py 1268
-        d = self._fn.check(Monitor(), verify=True)
+        d.addCallback(lambda ignored:
+            self._fn.check(Monitor(), verify=True))
         d.addCallback(self.check_bad, "test_verify_one_bad_block")
         d.addCallback(self.check_expected_failure,
                       CorruptShareError, "block hash tree failure",
hunk ./src/allmydata/test/test_mutable.py 1277
         return d
 
     def test_verify_one_bad_sharehash(self):
-        corrupt(None, self._storage, "share_hash_chain", [9], 5)
-        d = self._fn.check(Monitor(), verify=True)
+        d = corrupt(None, self._storage, "share_hash_chain", [9], 5)
+        d.addCallback(lambda ignored:
+            self._fn.check(Monitor(), verify=True))
         d.addCallback(self.check_bad, "test_verify_one_bad_sharehash")
         d.addCallback(self.check_expected_failure,
                       CorruptShareError, "corrupt hashes",
hunk ./src/allmydata/test/test_mutable.py 1287
         return d
 
     def test_verify_one_bad_encprivkey(self):
-        corrupt(None, self._storage, "enc_privkey", [9]) # bad privkey
-        d = self._fn.check(Monitor(), verify=True)
+        d = corrupt(None, self._storage, "enc_privkey", [9]) # bad privkey
+        d.addCallback(lambda ignored:
+            self._fn.check(Monitor(), verify=True))
         d.addCallback(self.check_bad, "test_verify_one_bad_encprivkey")
         d.addCallback(self.check_expected_failure,
                       CorruptShareError, "invalid privkey",
hunk ./src/allmydata/test/test_mutable.py 1297
         return d
 
     def test_verify_one_bad_encprivkey_uncheckable(self):
-        corrupt(None, self._storage, "enc_privkey", [9]) # bad privkey
+        d = corrupt(None, self._storage, "enc_privkey", [9]) # bad privkey
         readonly_fn = self._fn.get_readonly()
         # a read-only node has no way to validate the privkey
hunk ./src/allmydata/test/test_mutable.py 1300
-        d = readonly_fn.check(Monitor(), verify=True)
+        d.addCallback(lambda ignored:
+            readonly_fn.check(Monitor(), verify=True))
         d.addCallback(self.check_good,
                       "test_verify_one_bad_encprivkey_uncheckable")
         return d
}
[Add MDMF reader and writer
Kevan Carstensen <kevan@isnotajoke.com>**20100626234033
 Ignore-this: 30bb93141aacc4201e87302a462b3075
 
 The MDMF/SDMF reader and MDMF writer are similar to the object proxies
 that exist for immutable files. They abstract away details of
 connection, state, and caching from their callers (in this case, the
 download, servermap updater, and uploader), and expose methods to get
 and set information on the remote server.
 
 MDMFSlotReadProxy reads a mutable file from the server, doing the right
 thing (in most cases) regardless of whether the file is MDMF or SDMF. It
 allows callers to tell it how to batch and flush reads.
 
 MDMFSlotWriteProxy writes an MDMF mutable file to a server.
 
 This patch also includes tests for MDMFSlotReadProxy and
 MDMFSlotWriteProxy.
] {
hunk ./src/allmydata/mutable/layout.py 4
 
 import struct
 from allmydata.mutable.common import NeedMoreDataError, UnknownVersionError
+from allmydata.interfaces import HASH_SIZE, SALT_SIZE, SDMF_VERSION, \
+                                 MDMF_VERSION
+from allmydata.util import mathutil, observer
+from twisted.python import failure
+from twisted.internet import defer
+
+
+# These strings describe the format of the packed structs they help process
+# Here's what they mean:
+#
+#  PREFIX:
+#    >: Big-endian byte order; the most significant byte is first (leftmost).
+#    B: The version information; an 8 bit version identifier. Stored as
+#       an unsigned char. This is currently 00 00 00 00; our modifications
+#       will turn it into 00 00 00 01.
+#    Q: The sequence number; this is sort of like a revision history for
+#       mutable files; they start at 1 and increase as they are changed after
+#       being uploaded. Stored as an unsigned long long, which is 8 bytes in
+#       length.
+#  32s: The root hash of the share hash tree. We use sha-256d, so we use 32 
+#       characters = 32 bytes to store the value.
+#  16s: The salt for the readkey. This is a 16-byte random value, stored as
+#       16 characters.
+#
+#  SIGNED_PREFIX additions, things that are covered by the signature:
+#    B: The "k" encoding parameter. We store this as an 8-bit character, 
+#       which is convenient because our erasure coding scheme cannot 
+#       encode if you ask for more than 255 pieces.
+#    B: The "N" encoding parameter. Stored as an 8-bit character for the 
+#       same reasons as above.
+#    Q: The segment size of the uploaded file. This will essentially be the
+#       length of the file in SDMF. An unsigned long long, so we can store 
+#       files of quite large size.
+#    Q: The data length of the uploaded file. Modulo padding, this will be
+#       the same of the data length field. Like the data length field, it is
+#       an unsigned long long and can be quite large.
+#
+#   HEADER additions:
+#     L: The offset of the signature of this. An unsigned long.
+#     L: The offset of the share hash chain. An unsigned long.
+#     L: The offset of the block hash tree. An unsigned long.
+#     L: The offset of the share data. An unsigned long.
+#     Q: The offset of the encrypted private key. An unsigned long long, to
+#        account for the possibility of a lot of share data.
+#     Q: The offset of the EOF. An unsigned long long, to account for the
+#        possibility of a lot of share data.
+# 
+#  After all of these, we have the following:
+#    - The verification key: Occupies the space between the end of the header
+#      and the start of the signature (i.e.: data[HEADER_LENGTH:o['signature']].
+#    - The signature, which goes from the signature offset to the share hash
+#      chain offset.
+#    - The share hash chain, which goes from the share hash chain offset to
+#      the block hash tree offset.
+#    - The share data, which goes from the share data offset to the encrypted
+#      private key offset.
+#    - The encrypted private key offset, which goes until the end of the file.
+# 
+#  The block hash tree in this encoding has only one share, so the offset of
+#  the share data will be 32 bits more than the offset of the block hash tree.
+#  Given this, we may need to check to see how many bytes a reasonably sized
+#  block hash tree will take up.
 
 PREFIX = ">BQ32s16s" # each version has a different prefix
 SIGNED_PREFIX = ">BQ32s16s BBQQ" # this is covered by the signature
hunk ./src/allmydata/mutable/layout.py 72
 SIGNED_PREFIX_LENGTH = struct.calcsize(SIGNED_PREFIX)
 HEADER = ">BQ32s16s BBQQ LLLLQQ" # includes offsets
 HEADER_LENGTH = struct.calcsize(HEADER)
+OFFSETS = ">LLLLQQ"
+OFFSETS_LENGTH = struct.calcsize(OFFSETS)
 
 def unpack_header(data):
     o = {}
hunk ./src/allmydata/mutable/layout.py 193
     return (share_hash_chain, block_hash_tree, share_data)
 
 
-def pack_checkstring(seqnum, root_hash, IV):
+def pack_checkstring(seqnum, root_hash, IV, version=0):
     return struct.pack(PREFIX,
hunk ./src/allmydata/mutable/layout.py 195
-                       0, # version,
+                       version,
                        seqnum,
                        root_hash,
                        IV)
hunk ./src/allmydata/mutable/layout.py 268
                            encprivkey])
     return final_share
 
+def pack_prefix(seqnum, root_hash, IV,
+                required_shares, total_shares,
+                segment_size, data_length):
+    prefix = struct.pack(SIGNED_PREFIX,
+                         0, # version,
+                         seqnum,
+                         root_hash,
+                         IV,
+                         required_shares,
+                         total_shares,
+                         segment_size,
+                         data_length,
+                         )
+    return prefix
+
+
+MDMFHEADER = ">BQ32sBBQQ QQQQQQ"
+MDMFHEADERWITHOUTOFFSETS = ">BQ32sBBQQ"
+MDMFHEADERSIZE = struct.calcsize(MDMFHEADER)
+MDMFHEADERWITHOUTOFFSETSSIZE = struct.calcsize(MDMFHEADERWITHOUTOFFSETS)
+MDMFCHECKSTRING = ">BQ32s"
+MDMFSIGNABLEHEADER = ">BQ32sBBQQ"
+MDMFOFFSETS = ">QQQQQQ"
+MDMFOFFSETS_LENGTH = struct.calcsize(MDMFOFFSETS)
+
+class MDMFSlotWriteProxy:
+    #implements(IMutableSlotWriter) TODO
+
+    """
+    I represent a remote write slot for an MDMF mutable file.
+
+    I abstract away from my caller the details of block and salt
+    management, and the implementation of the on-disk format for MDMF
+    shares.
+    """
+    # We want to get rid of the salt tree root hash for MDMF. It is
+    # easier to check validity if we store the salts next to the block
+    # data for each segment. We then make the block hash tree leaves
+    # made up of hashes of block + salt for a particular block/segment.
+    # In this way, we get salt validity checking for free, and make for
+    # simpler upload and download logic.
+    # 
+    # Problems?
+    # 
+    # Well, the meaning of the datalength field, per
+    # docs/specifications/mutable.txt, is to give the datalength of the
+    # original plaintext -- not the ciphertext.
+    #
+    # We don't really care about this issue when we're uploading. But
+    # the downloader needs to know how many segments there are, and
+    # calculates this by checking the data length against the segment
+    # size. Not a problem when the only thing in segments are the 
+    # Wait, this is stupid. Leave the data length and the segment size
+    # alone -- this will allow the code to calculate the number of
+    # segments appropriately. Then tweak the read code to read more from
+    # the remote server. Done.
+
+    # Expected layout, MDMF:
+    # offset:     size:       name:
+    #-- signed part --
+    # 0           1           version number (01)
+    # 1           8           sequence number 
+    # 9           32          share tree root hash
+    # 41          1           The "k" encoding parameter
+    # 42          1           The "N" encoding parameter
+    # 43          8           The segment size of the uploaded file
+    # 51          8           The data length of the original plaintext
+    #-- end signed part --
+    # 59          8           The offset of the encrypted private key
+    # 67          8           The offset of the block hash tree
+    # 75          8           The offset of the share hash chain 
+    # 83          8           The offset of the signature
+    # 91          8           The offset of the verification key
+    # 99          8           The offset of the EOF
+    #
+    # followed by salts and share data, the encrypted private key, the
+    # block hash tree, the salt hash tree, the share hash chain, a
+    # signature over the first eight fields, and a verification key.
+    # 
+    # The checkstring is the first three fields -- the version number,
+    # sequence number, root hash and root salt hash. This is consistent
+    # in meaning to what we have with SDMF files, except now instead of
+    # using the literal salt, we use a value derived from all of the
+    # salts -- the share hash root.
+    # 
+    # The salt is stored before the block for each segment. The block
+    # hash tree is computed over the combination of block and salt for
+    # each segment. In this way, we get integrity checking for both
+    # block and salt with the current block hash tree arrangement.
+    # 
+    # The ordering of the offsets is different to reflect the dependencies
+    # that we'll run into with an MDMF file. The expected write flow is
+    # something like this:
+    #
+    #   0: Initialize with the sequence number, encoding parameters and
+    #      data length. From this, we can deduce the number of segments,
+    #      and where they should go.. We can also figure out where the
+    #      encrypted private key should go, because we can figure out how
+    #      big the share data will be.
+    # 
+    #   1: Encrypt, encode, and upload the file in chunks. Do something
+    #      like 
+    #
+    #       put_block(data, segnum, salt)
+    #
+    #      to write a block and a salt to the disk. We can do both of
+    #      these operations now because we have enough of the offsets to
+    #      know where to put them.
+    # 
+    #   2: Put the encrypted private key. Use:
+    #
+    #        put_encprivkey(encprivkey)
+    #
+    #      Now that we know the length of the private key, we can fill
+    #      in the offset for the block hash tree.
+    #
+    #   3: We're now in a position to upload the block hash tree for
+    #      a share. Put that using something like:
+    #       
+    #        put_blockhashes(block_hash_tree)
+    #
+    #      Note that block_hash_tree is a list of hashes -- we'll take
+    #      care of the details of serializing that appropriately. When
+    #      we get the block hash tree, we are also in a position to
+    #      calculate the offset for the share hash chain, and fill that
+    #      into the offsets table.
+    #
+    #   4: At the same time, we're in a position to upload the salt hash
+    #      tree. This is a Merkle tree over all of the salts. We use a
+    #      Merkle tree so that we can validate each block,salt pair as
+    #      we download them later. We do this using 
+    #
+    #        put_salthashes(salt_hash_tree)
+    #
+    #      When you do this, I automatically put the root of the tree
+    #      (the hash at index 0 of the list) in its appropriate slot in
+    #      the signed prefix of the share.
+    #
+    #   5: We're now in a position to upload the share hash chain for
+    #      a share. Do that with something like:
+    #      
+    #        put_sharehashes(share_hash_chain) 
+    #
+    #      share_hash_chain should be a dictionary mapping shnums to 
+    #      32-byte hashes -- the wrapper handles serialization.
+    #      We'll know where to put the signature at this point, also.
+    #      The root of this tree will be put explicitly in the next
+    #      step.
+    # 
+    #      TODO: Why? Why not just include it in the tree here?
+    #
+    #   6: Before putting the signature, we must first put the
+    #      root_hash. Do this with:
+    # 
+    #        put_root_hash(root_hash).
+    #      
+    #      In terms of knowing where to put this value, it was always
+    #      possible to place it, but it makes sense semantically to
+    #      place it after the share hash tree, so that's why you do it
+    #      in this order.
+    #
+    #   6: With the root hash put, we can now sign the header. Use:
+    #
+    #        get_signable()
+    #
+    #      to get the part of the header that you want to sign, and use:
+    #       
+    #        put_signature(signature)
+    #
+    #      to write your signature to the remote server.
+    #
+    #   6: Add the verification key, and finish. Do:
+    #
+    #        put_verification_key(key) 
+    #
+    #      and 
+    #
+    #        finish_publish()
+    #
+    # Checkstring management:
+    # 
+    # To write to a mutable slot, we have to provide test vectors to ensure
+    # that we are writing to the same data that we think we are. These
+    # vectors allow us to detect uncoordinated writes; that is, writes
+    # where both we and some other shareholder are writing to the
+    # mutable slot, and to report those back to the parts of the program
+    # doing the writing. 
+    #
+    # With SDMF, this was easy -- all of the share data was written in
+    # one go, so it was easy to detect uncoordinated writes, and we only
+    # had to do it once. With MDMF, not all of the file is written at
+    # once.
+    #
+    # If a share is new, we write out as much of the header as we can
+    # before writing out anything else. This gives other writers a
+    # canary that they can use to detect uncoordinated writes, and, if
+    # they do the same thing, gives us the same canary. We them update
+    # the share. We won't be able to write out two fields of the header
+    # -- the share tree hash and the salt hash -- until we finish
+    # writing out the share. We only require the writer to provide the
+    # initial checkstring, and keep track of what it should be after
+    # updates ourselves.
+    #
+    # If we haven't written anything yet, then on the first write (which
+    # will probably be a block + salt of a share), we'll also write out
+    # the header. On subsequent passes, we'll expect to see the header.
+    # This changes in two places:
+    #
+    #   - When we write out the salt hash
+    #   - When we write out the root of the share hash tree
+    #
+    # since these values will change the header. It is possible that we 
+    # can just make those be written in one operation to minimize
+    # disruption.
+    def __init__(self,
+                 shnum,
+                 rref, # a remote reference to a storage server
+                 storage_index,
+                 secrets, # (write_enabler, renew_secret, cancel_secret)
+                 seqnum, # the sequence number of the mutable file
+                 required_shares,
+                 total_shares,
+                 segment_size,
+                 data_length): # the length of the original file
+        self._shnum = shnum
+        self._rref = rref
+        self._storage_index = storage_index
+        self._seqnum = seqnum
+        self._required_shares = required_shares
+        assert self._shnum >= 0 and self._shnum < total_shares
+        self._total_shares = total_shares
+        # We build up the offset table as we write things. It is the
+        # last thing we write to the remote server. 
+        self._offsets = {}
+        self._testvs = []
+        self._secrets = secrets
+        # The segment size needs to be a multiple of the k parameter --
+        # any padding should have been carried out by the publisher
+        # already.
+        assert segment_size % required_shares == 0
+        self._segment_size = segment_size
+        self._data_length = data_length
+
+        # These are set later -- we define them here so that we can
+        # check for their existence easily
+
+        # This is the root of the share hash tree -- the Merkle tree
+        # over the roots of the block hash trees computed for shares in
+        # this upload.
+        self._root_hash = None
+
+        # We haven't yet written anything to the remote bucket. By
+        # setting this, we tell the _write method as much. The write
+        # method will then know that it also needs to add a write vector
+        # for the checkstring (or what we have of it) to the first write
+        # request. We'll then record that value for future use.  If
+        # we're expecting something to be there already, we need to call
+        # set_checkstring before we write anything to tell the first
+        # write about that.
+        self._written = False
+
+        # When writing data to the storage servers, we get a read vector
+        # for free. We'll read the checkstring, which will help us
+        # figure out what's gone wrong if a write fails.
+        self._readv = [(0, struct.calcsize(MDMFCHECKSTRING))]
+
+        # We calculate the number of segments because it tells us
+        # where the salt part of the file ends/share segment begins,
+        # and also because it provides a useful amount of bounds checking.
+        self._num_segments = mathutil.div_ceil(self._data_length,
+                                               self._segment_size)
+        self._block_size = self._segment_size / self._required_shares
+        # We also calculate the share size, to help us with block
+        # constraints later.
+        tail_size = self._data_length % self._segment_size
+        if not tail_size:
+            self._tail_block_size = self._block_size
+        else:
+            self._tail_block_size = mathutil.next_multiple(tail_size,
+                                                           self._required_shares)
+            self._tail_block_size /= self._required_shares
+
+        # We already know where the sharedata starts; right after the end
+        # of the header (which is defined as the signable part + the offsets)
+        # We can also calculate where the encrypted private key begins
+        # from what we know know.
+        self._actual_block_size = self._block_size + SALT_SIZE
+        data_size = self._actual_block_size * (self._num_segments - 1)
+        data_size += self._tail_block_size
+        data_size += SALT_SIZE
+        self._offsets['enc_privkey'] = MDMFHEADERSIZE
+        self._offsets['enc_privkey'] += data_size
+        # We'll wait for the rest. Callers can now call my "put_block" and
+        # "set_checkstring" methods.
+
+
+    def set_checkstring(self, seqnum_or_checkstring, root_hash=None):
+        """
+        Set checkstring checkstring for the given shnum.
+
+        This can be invoked in one of two ways.
+
+        With one argument, I assume that you are giving me a literal
+        checkstring -- e.g., the output of get_checkstring. I will then
+        set that checkstring as it is. This form is used by unit tests.
+
+        With two arguments, I assume that you are giving me a sequence
+        number and root hash to make a checkstring from. In that case, I
+        will build a checkstring and set it for you. This form is used
+        by the publisher.
+
+        By default, I assume that I am writing new shares to the grid.
+        If you don't explcitly set your own checkstring, I will use
+        one that requires that the remote share not exist. You will want
+        to use this method if you are updating a share in-place;
+        otherwise, writes will fail.
+        """
+        # You're allowed to overwrite checkstrings with this method;
+        # I assume that users know what they are doing when they call
+        # it.
+        if root_hash:
+            checkstring = struct.pack(MDMFCHECKSTRING,
+                                      1,
+                                      seqnum_or_checkstring,
+                                      root_hash)
+        else:
+            checkstring = seqnum_or_checkstring
+
+        if checkstring == "":
+            # We special-case this, since len("") = 0, but we need
+            # length of 1 for the case of an empty share to work on the
+            # storage server, which is what a checkstring that is the
+            # empty string means.
+            self._testvs = []
+        else:
+            self._testvs = []
+            self._testvs.append((0, len(checkstring), "eq", checkstring))
+
+
+    def __repr__(self):
+        return "MDMFSlotWriteProxy for share %d" % self._shnum
+
+
+    def get_checkstring(self):
+        """
+        Given a share number, I return a representation of what the
+        checkstring for that share on the server will look like.
+
+        I am mostly used for tests.
+        """
+        if self._root_hash:
+            roothash = self._root_hash
+        else:
+            roothash = "\x00" * 32
+        return struct.pack(MDMFCHECKSTRING,
+                           1,
+                           self._seqnum,
+                           roothash)
+
+
+    def put_block(self, data, segnum, salt):
+        """
+        Put the encrypted-and-encoded data segment in the slot, along
+        with the salt.
+        """
+        if segnum >= self._num_segments:
+            raise LayoutInvalid("I won't overwrite the private key")
+        if len(salt) != SALT_SIZE:
+            raise LayoutInvalid("I was given a salt of size %d, but "
+                                "I wanted a salt of size %d")
+        if segnum + 1 == self._num_segments:
+            if len(data) != self._tail_block_size:
+                raise LayoutInvalid("I was given the wrong size block to write")
+        elif len(data) != self._block_size:
+            raise LayoutInvalid("I was given the wrong size block to write")
+
+        # We want to write at len(MDMFHEADER) + segnum * block_size.
+
+        offset = MDMFHEADERSIZE + (self._actual_block_size * segnum)
+        data = salt + data
+
+        datavs = [tuple([offset, data])]
+        return self._write(datavs)
+
+
+    def put_encprivkey(self, encprivkey):
+        """
+        Put the encrypted private key in the remote slot.
+        """
+        assert self._offsets
+        assert self._offsets['enc_privkey']
+        # You shouldn't re-write the encprivkey after the block hash
+        # tree is written, since that could cause the private key to run
+        # into the block hash tree. Before it writes the block hash
+        # tree, the block hash tree writing method writes the offset of
+        # the salt hash tree. So that's a good indicator of whether or
+        # not the block hash tree has been written.
+        if "share_hash_chain" in self._offsets:
+            raise LayoutInvalid("You must write this before the block hash tree")
+
+        self._offsets['block_hash_tree'] = self._offsets['enc_privkey'] + len(encprivkey)
+        datavs = [(tuple([self._offsets['enc_privkey'], encprivkey]))]
+        def _on_failure():
+            del(self._offsets['block_hash_tree'])
+        return self._write(datavs, on_failure=_on_failure)
+
+
+    def put_blockhashes(self, blockhashes):
+        """
+        Put the block hash tree in the remote slot.
+
+        The encrypted private key must be put before the block hash
+        tree, since we need to know how large it is to know where the
+        block hash tree should go. The block hash tree must be put
+        before the salt hash tree, since its size determines the
+        offset of the share hash chain.
+        """
+        assert self._offsets
+        assert isinstance(blockhashes, list)
+        if "block_hash_tree" not in self._offsets:
+            raise LayoutInvalid("You must put the encrypted private key "
+                                "before you put the block hash tree")
+        # If written, the share hash chain causes the signature offset
+        # to be defined.
+        if "signature" in self._offsets:
+            raise LayoutInvalid("You must put the block hash tree before "
+                                "you put the share hash chain")
+        blockhashes_s = "".join(blockhashes)
+        self._offsets['share_hash_chain'] = self._offsets['block_hash_tree'] + len(blockhashes_s)
+        datavs = []
+        datavs.append(tuple([self._offsets['block_hash_tree'], blockhashes_s]))
+        def _on_failure():
+            del(self._offsets['share_hash_chain'])
+        return self._write(datavs, on_failure=_on_failure)
+
+
+    def put_sharehashes(self, sharehashes):
+        """
+        Put the share hash chain in the remote slot.
+
+        The salt hash tree must be put before the share hash chain,
+        since we need to know where the salt hash tree ends before we
+        can know where the share hash chain starts. The share hash chain
+        must be put before the signature, since the length of the packed
+        share hash chain determines the offset of the signature. Also,
+        semantically, you must know what the root of the salt hash tree
+        is before you can generate a valid signature.
+        """
+        assert isinstance(sharehashes, dict)
+        if "share_hash_chain" not in self._offsets:
+            raise LayoutInvalid("You need to put the salt hash tree before "
+                                "you can put the share hash chain")
+        # The signature comes after the share hash chain. If the
+        # signature has already been written, we must not write another
+        # share hash chain. The signature writes the verification key
+        # offset when it gets sent to the remote server, so we look for
+        # that.
+        if "verification_key" in self._offsets:
+            raise LayoutInvalid("You must write the share hash chain "
+                                "before you write the signature")
+        datavs = []
+        sharehashes_s = "".join([struct.pack(">H32s", i, sharehashes[i])
+                                  for i in sorted(sharehashes.keys())])
+        self._offsets['signature'] = self._offsets['share_hash_chain'] + len(sharehashes_s)
+        datavs.append(tuple([self._offsets['share_hash_chain'], sharehashes_s]))
+        def _on_failure():
+            del(self._offsets['signature'])
+        return self._write(datavs, on_failure=_on_failure)
+
+
+    def put_root_hash(self, roothash):
+        """
+        Put the root hash (the root of the share hash tree) in the
+        remote slot.
+        """
+        # It does not make sense to be able to put the root 
+        # hash without first putting the share hashes, since you need
+        # the share hashes to generate the root hash.
+        #
+        # Signature is defined by the routine that places the share hash
+        # chain, so it's a good thing to look for in finding out whether
+        # or not the share hash chain exists on the remote server.
+        if "signature" not in self._offsets:
+            raise LayoutInvalid("You need to put the share hash chain "
+                                "before you can put the root share hash")
+        if len(roothash) != HASH_SIZE:
+            raise LayoutInvalid("hashes and salts must be exactly %d bytes"
+                                 % HASH_SIZE)
+        datavs = []
+        self._root_hash = roothash
+        # To write both of these values, we update the checkstring on
+        # the remote server, which includes them
+        checkstring = self.get_checkstring()
+        datavs.append(tuple([0, checkstring]))
+        # This write, if successful, changes the checkstring, so we need
+        # to update our internal checkstring to be consistent with the
+        # one on the server.
+        def _on_success():
+            self._testvs = [(0, len(checkstring), "eq", checkstring)]
+        def _on_failure():
+            self._root_hash = None
+        return self._write(datavs,
+                           on_success=_on_success,
+                           on_failure=_on_failure)
+
+
+    def get_signable(self):
+        """
+        Get the first seven fields of the mutable file; the parts that
+        are signed.
+        """
+        if not self._root_hash:
+            raise LayoutInvalid("You need to set the root hash "
+                                "before getting something to "
+                                "sign")
+        return struct.pack(MDMFSIGNABLEHEADER,
+                           1,
+                           self._seqnum,
+                           self._root_hash,
+                           self._required_shares,
+                           self._total_shares,
+                           self._segment_size,
+                           self._data_length)
+
+
+    def put_signature(self, signature):
+        """
+        Put the signature field to the remote slot.
+
+        I require that the root hash and share hash chain have been put
+        to the grid before I will write the signature to the grid.
+        """
+        if "signature" not in self._offsets:
+            raise LayoutInvalid("You must put the share hash chain "
+        # It does not make sense to put a signature without first
+        # putting the root hash and the salt hash (since otherwise
+        # the signature would be incomplete), so we don't allow that.
+                       "before putting the signature")
+        if not self._root_hash:
+            raise LayoutInvalid("You must complete the signed prefix "
+                                "before computing a signature")
+        # If we put the signature after we put the verification key, we
+        # could end up running into the verification key, and will
+        # probably screw up the offsets as well. So we don't allow that.
+        # The method that writes the verification key defines the EOF
+        # offset before writing the verification key, so look for that.
+        if "EOF" in self._offsets:
+            raise LayoutInvalid("You must write the signature before the verification key")
+
+        self._offsets['verification_key'] = self._offsets['signature'] + len(signature)
+        datavs = []
+        datavs.append(tuple([self._offsets['signature'], signature]))
+        def _on_failure():
+            del(self._offsets['verification_key'])
+        return self._write(datavs, on_failure=_on_failure)
+
+
+    def put_verification_key(self, verification_key):
+        """
+        Put the verification key into the remote slot.
+
+        I require that the signature have been written to the storage
+        server before I allow the verification key to be written to the
+        remote server.
+        """
+        if "verification_key" not in self._offsets:
+            raise LayoutInvalid("You must put the signature before you "
+                                "can put the verification key")
+        self._offsets['EOF'] = self._offsets['verification_key'] + len(verification_key)
+        datavs = []
+        datavs.append(tuple([self._offsets['verification_key'], verification_key]))
+        def _on_failure():
+            del(self._offsets['EOF'])
+        return self._write(datavs, on_failure=_on_failure)
+
+
+    def finish_publishing(self):
+        """
+        Write the offset table and encoding parameters to the remote
+        slot, since that's the only thing we have yet to publish at this
+        point.
+        """
+        if "EOF" not in self._offsets:
+            raise LayoutInvalid("You must put the verification key before "
+                                "you can publish the offsets")
+        offsets_offset = struct.calcsize(MDMFHEADERWITHOUTOFFSETS)
+        offsets = struct.pack(MDMFOFFSETS,
+                              self._offsets['enc_privkey'],
+                              self._offsets['block_hash_tree'],
+                              self._offsets['share_hash_chain'],
+                              self._offsets['signature'],
+                              self._offsets['verification_key'],
+                              self._offsets['EOF'])
+        datavs = []
+        datavs.append(tuple([offsets_offset, offsets]))
+        encoding_parameters_offset = struct.calcsize(MDMFCHECKSTRING)
+        params = struct.pack(">BBQQ",
+                             self._required_shares,
+                             self._total_shares,
+                             self._segment_size,
+                             self._data_length)
+        datavs.append(tuple([encoding_parameters_offset, params]))
+        return self._write(datavs)
+
+
+    def _write(self, datavs, on_failure=None, on_success=None):
+        """I write the data vectors in datavs to the remote slot."""
+        tw_vectors = {}
+        new_share = False
+        if not self._testvs:
+            self._testvs = []
+            self._testvs.append(tuple([0, 1, "eq", ""]))
+            new_share = True
+        if not self._written:
+            # Write a new checkstring to the share when we write it, so
+            # that we have something to check later.
+            new_checkstring = self.get_checkstring()
+            datavs.append((0, new_checkstring))
+            def _first_write():
+                self._written = True
+                self._testvs = [(0, len(new_checkstring), "eq", new_checkstring)]
+            on_success = _first_write
+        tw_vectors[self._shnum] = (self._testvs, datavs, None)
+        datalength = sum([len(x[1]) for x in datavs])
+        d = self._rref.callRemote("slot_testv_and_readv_and_writev",
+                                  self._storage_index,
+                                  self._secrets,
+                                  tw_vectors,
+                                  self._readv)
+        def _result(results):
+            if isinstance(results, failure.Failure) or not results[0]:
+                # Do nothing; the write was unsuccessful.
+                if on_failure: on_failure()
+            else:
+                if on_success: on_success()
+            return results
+        d.addCallback(_result)
+        return d
+
+
+class MDMFSlotReadProxy:
+    """
+    I read from a mutable slot filled with data written in the MDMF data
+    format (which is described above).
+
+    I can be initialized with some amount of data, which I will use (if
+    it is valid) to eliminate some of the need to fetch it from servers.
+    """
+    def __init__(self,
+                 rref,
+                 storage_index,
+                 shnum,
+                 data=""):
+        # Start the initialization process.
+        self._rref = rref
+        self._storage_index = storage_index
+        self.shnum = shnum
+
+        # Before doing anything, the reader is probably going to want to
+        # verify that the signature is correct. To do that, they'll need
+        # the verification key, and the signature. To get those, we'll
+        # need the offset table. So fetch the offset table on the
+        # assumption that that will be the first thing that a reader is
+        # going to do.
+
+        # The fact that these encoding parameters are None tells us
+        # that we haven't yet fetched them from the remote share, so we
+        # should. We could just not set them, but the checks will be
+        # easier to read if we don't have to use hasattr.
+        self._version_number = None
+        self._sequence_number = None
+        self._root_hash = None
+        # Filled in if we're dealing with an SDMF file. Unused
+        # otherwise.
+        self._salt = None
+        self._required_shares = None
+        self._total_shares = None
+        self._segment_size = None
+        self._data_length = None
+        self._offsets = None
+
+        # If the user has chosen to initialize us with some data, we'll
+        # try to satisfy subsequent data requests with that data before
+        # asking the storage server for it. If 
+        self._data = data
+        # The way callers interact with cache in the filenode returns
+        # None if there isn't any cached data, but the way we index the
+        # cached data requires a string, so convert None to "".
+        if self._data == None:
+            self._data = ""
+
+        self._queue_observers = observer.ObserverList()
+        self._queue_errbacks = observer.ObserverList()
+        self._readvs = []
+
+
+    def _maybe_fetch_offsets_and_header(self, force_remote=False):
+        """
+        I fetch the offset table and the header from the remote slot if
+        I don't already have them. If I do have them, I do nothing and
+        return an empty Deferred.
+        """
+        if self._offsets:
+            return defer.succeed(None)
+        # At this point, we may be either SDMF or MDMF. Fetching 107 
+        # bytes will be enough to get header and offsets for both SDMF and
+        # MDMF, though we'll be left with 4 more bytes than we
+        # need if this ends up being MDMF. This is probably less
+        # expensive than the cost of a second roundtrip.
+        readvs = [(0, 107)]
+        d = self._read(readvs, force_remote)
+        d.addCallback(self._process_encoding_parameters)
+        d.addCallback(self._process_offsets)
+        return d
+
+
+    def _process_encoding_parameters(self, encoding_parameters):
+        assert self.shnum in encoding_parameters
+        encoding_parameters = encoding_parameters[self.shnum][0]
+        # The first byte is the version number. It will tell us what
+        # to do next.
+        (verno,) = struct.unpack(">B", encoding_parameters[:1])
+        if verno == MDMF_VERSION:
+            read_size = MDMFHEADERWITHOUTOFFSETSSIZE
+            (verno,
+             seqnum,
+             root_hash,
+             k,
+             n,
+             segsize,
+             datalen) = struct.unpack(MDMFHEADERWITHOUTOFFSETS,
+                                      encoding_parameters[:read_size])
+            if segsize == 0 and datalen == 0:
+                # Empty file, no segments.
+                self._num_segments = 0
+            else:
+                self._num_segments = mathutil.div_ceil(datalen, segsize)
+
+        elif verno == SDMF_VERSION:
+            read_size = SIGNED_PREFIX_LENGTH
+            (verno,
+             seqnum,
+             root_hash,
+             salt,
+             k,
+             n,
+             segsize,
+             datalen) = struct.unpack(">BQ32s16s BBQQ",
+                                encoding_parameters[:SIGNED_PREFIX_LENGTH])
+            self._salt = salt
+            if segsize == 0 and datalen == 0:
+                # empty file
+                self._num_segments = 0
+            else:
+                # non-empty SDMF files have one segment.
+                self._num_segments = 1
+        else:
+            raise UnknownVersionError("You asked me to read mutable file "
+                                      "version %d, but I only understand "
+                                      "%d and %d" % (verno, SDMF_VERSION,
+                                                     MDMF_VERSION))
+
+        self._version_number = verno
+        self._sequence_number = seqnum
+        self._root_hash = root_hash
+        self._required_shares = k
+        self._total_shares = n
+        self._segment_size = segsize
+        self._data_length = datalen
+
+        self._block_size = self._segment_size / self._required_shares
+        # We can upload empty files, and need to account for this fact
+        # so as to avoid zero-division and zero-modulo errors.
+        if datalen > 0:
+            tail_size = self._data_length % self._segment_size
+        else:
+            tail_size = 0
+        if not tail_size:
+            self._tail_block_size = self._block_size
+        else:
+            self._tail_block_size = mathutil.next_multiple(tail_size,
+                                                    self._required_shares)
+            self._tail_block_size /= self._required_shares
+
+        return encoding_parameters
+
+
+    def _process_offsets(self, offsets):
+        if self._version_number == 0:
+            read_size = OFFSETS_LENGTH
+            read_offset = SIGNED_PREFIX_LENGTH
+            end = read_size + read_offset
+            (signature,
+             share_hash_chain,
+             block_hash_tree,
+             share_data,
+             enc_privkey,
+             EOF) = struct.unpack(">LLLLQQ",
+                                  offsets[read_offset:end])
+            self._offsets = {}
+            self._offsets['signature'] = signature
+            self._offsets['share_data'] = share_data
+            self._offsets['block_hash_tree'] = block_hash_tree
+            self._offsets['share_hash_chain'] = share_hash_chain
+            self._offsets['enc_privkey'] = enc_privkey
+            self._offsets['EOF'] = EOF
+
+        elif self._version_number == 1:
+            read_offset = MDMFHEADERWITHOUTOFFSETSSIZE
+            read_length = MDMFOFFSETS_LENGTH
+            end = read_offset + read_length
+            (encprivkey,
+             blockhashes,
+             sharehashes,
+             signature,
+             verification_key,
+             eof) = struct.unpack(MDMFOFFSETS,
+                                  offsets[read_offset:end])
+            self._offsets = {}
+            self._offsets['enc_privkey'] = encprivkey
+            self._offsets['block_hash_tree'] = blockhashes
+            self._offsets['share_hash_chain'] = sharehashes
+            self._offsets['signature'] = signature
+            self._offsets['verification_key'] = verification_key
+            self._offsets['EOF'] = eof
+
+
+    def get_block_and_salt(self, segnum, queue=False):
+        """
+        I return (block, salt), where block is the block data and
+        salt is the salt used to encrypt that segment.
+        """
+        d = self._maybe_fetch_offsets_and_header()
+        def _then(ignored):
+            if self._version_number == 1:
+                base_share_offset = MDMFHEADERSIZE
+            else:
+                base_share_offset = self._offsets['share_data']
+
+            if segnum + 1 > self._num_segments:
+                raise LayoutInvalid("Not a valid segment number")
+
+            if self._version_number == 0:
+                share_offset = base_share_offset + self._block_size * segnum
+            else:
+                share_offset = base_share_offset + (self._block_size + \
+                                                    SALT_SIZE) * segnum
+            if segnum + 1 == self._num_segments:
+                data = self._tail_block_size
+            else:
+                data = self._block_size
+
+            if self._version_number == 1:
+                data += SALT_SIZE
+
+            readvs = [(share_offset, data)]
+            return readvs
+        d.addCallback(_then)
+        d.addCallback(lambda readvs:
+            self._read(readvs, queue=queue))
+        def _process_results(results):
+            assert self.shnum in results
+            if self._version_number == 0:
+                # We only read the share data, but we know the salt from
+                # when we fetched the header
+                data = results[self.shnum]
+                if not data:
+                    data = ""
+                else:
+                    assert len(data) == 1
+                    data = data[0]
+                salt = self._salt
+            else:
+                data = results[self.shnum]
+                if not data:
+                    salt = data = ""
+                else:
+                    salt_and_data = results[self.shnum][0]
+                    salt = salt_and_data[:SALT_SIZE]
+                    data = salt_and_data[SALT_SIZE:]
+            return data, salt
+        d.addCallback(_process_results)
+        return d
+
+
+    def get_blockhashes(self, needed=None, queue=False, force_remote=False):
+        """
+        I return the block hash tree
+
+        I take an optional argument, needed, which is a set of indices
+        correspond to hashes that I should fetch. If this argument is
+        missing, I will fetch the entire block hash tree; otherwise, I
+        may attempt to fetch fewer hashes, based on what needed says
+        that I should do. Note that I may fetch as many hashes as I
+        want, so long as the set of hashes that I do fetch is a superset
+        of the ones that I am asked for, so callers should be prepared
+        to tolerate additional hashes.
+        """
+        # TODO: Return only the parts of the block hash tree necessary
+        # to validate the blocknum provided?
+        # This is a good idea, but it is hard to implement correctly. It
+        # is bad to fetch any one block hash more than once, so we
+        # probably just want to fetch the whole thing at once and then
+        # serve it.
+        if needed == set([]):
+            return defer.succeed([])
+        d = self._maybe_fetch_offsets_and_header()
+        def _then(ignored):
+            blockhashes_offset = self._offsets['block_hash_tree']
+            if self._version_number == 1:
+                blockhashes_length = self._offsets['share_hash_chain'] - blockhashes_offset
+            else:
+                blockhashes_length = self._offsets['share_data'] - blockhashes_offset
+            readvs = [(blockhashes_offset, blockhashes_length)]
+            return readvs
+        d.addCallback(_then)
+        d.addCallback(lambda readvs:
+            self._read(readvs, queue=queue, force_remote=force_remote))
+        def _build_block_hash_tree(results):
+            assert self.shnum in results
+
+            rawhashes = results[self.shnum][0]
+            results = [rawhashes[i:i+HASH_SIZE]
+                       for i in range(0, len(rawhashes), HASH_SIZE)]
+            return results
+        d.addCallback(_build_block_hash_tree)
+        return d
+
+
+    def get_sharehashes(self, needed=None, queue=False, force_remote=False):
+        """
+        I return the part of the share hash chain placed to validate
+        this share.
+
+        I take an optional argument, needed. Needed is a set of indices
+        that correspond to the hashes that I should fetch. If needed is
+        not present, I will fetch and return the entire share hash
+        chain. Otherwise, I may fetch and return any part of the share
+        hash chain that is a superset of the part that I am asked to
+        fetch. Callers should be prepared to deal with more hashes than
+        they've asked for.
+        """
+        if needed == set([]):
+            return defer.succeed([])
+        d = self._maybe_fetch_offsets_and_header()
+
+        def _make_readvs(ignored):
+            sharehashes_offset = self._offsets['share_hash_chain']
+            if self._version_number == 0:
+                sharehashes_length = self._offsets['block_hash_tree'] - sharehashes_offset
+            else:
+                sharehashes_length = self._offsets['signature'] - sharehashes_offset
+            readvs = [(sharehashes_offset, sharehashes_length)]
+            return readvs
+        d.addCallback(_make_readvs)
+        d.addCallback(lambda readvs:
+            self._read(readvs, queue=queue, force_remote=force_remote))
+        def _build_share_hash_chain(results):
+            assert self.shnum in results
+
+            sharehashes = results[self.shnum][0]
+            results = [sharehashes[i:i+(HASH_SIZE + 2)]
+                       for i in range(0, len(sharehashes), HASH_SIZE + 2)]
+            results = dict([struct.unpack(">H32s", data)
+                            for data in results])
+            return results
+        d.addCallback(_build_share_hash_chain)
+        return d
+
+
+    def get_encprivkey(self, queue=False):
+        """
+        I return the encrypted private key.
+        """
+        d = self._maybe_fetch_offsets_and_header()
+
+        def _make_readvs(ignored):
+            privkey_offset = self._offsets['enc_privkey']
+            if self._version_number == 0:
+                privkey_length = self._offsets['EOF'] - privkey_offset
+            else:
+                privkey_length = self._offsets['block_hash_tree'] - privkey_offset
+            readvs = [(privkey_offset, privkey_length)]
+            return readvs
+        d.addCallback(_make_readvs)
+        d.addCallback(lambda readvs:
+            self._read(readvs, queue=queue))
+        def _process_results(results):
+            assert self.shnum in results
+            privkey = results[self.shnum][0]
+            return privkey
+        d.addCallback(_process_results)
+        return d
+
+
+    def get_signature(self, queue=False):
+        """
+        I return the signature of my share.
+        """
+        d = self._maybe_fetch_offsets_and_header()
+
+        def _make_readvs(ignored):
+            signature_offset = self._offsets['signature']
+            if self._version_number == 1:
+                signature_length = self._offsets['verification_key'] - signature_offset
+            else:
+                signature_length = self._offsets['share_hash_chain'] - signature_offset
+            readvs = [(signature_offset, signature_length)]
+            return readvs
+        d.addCallback(_make_readvs)
+        d.addCallback(lambda readvs:
+            self._read(readvs, queue=queue))
+        def _process_results(results):
+            assert self.shnum in results
+            signature = results[self.shnum][0]
+            return signature
+        d.addCallback(_process_results)
+        return d
+
+
+    def get_verification_key(self, queue=False):
+        """
+        I return the verification key.
+        """
+        d = self._maybe_fetch_offsets_and_header()
+
+        def _make_readvs(ignored):
+            if self._version_number == 1:
+                vk_offset = self._offsets['verification_key']
+                vk_length = self._offsets['EOF'] - vk_offset
+            else:
+                vk_offset = struct.calcsize(">BQ32s16sBBQQLLLLQQ")
+                vk_length = self._offsets['signature'] - vk_offset
+            readvs = [(vk_offset, vk_length)]
+            return readvs
+        d.addCallback(_make_readvs)
+        d.addCallback(lambda readvs:
+            self._read(readvs, queue=queue))
+        def _process_results(results):
+            assert self.shnum in results
+            verification_key = results[self.shnum][0]
+            return verification_key
+        d.addCallback(_process_results)
+        return d
+
+
+    def get_encoding_parameters(self):
+        """
+        I return (k, n, segsize, datalen)
+        """
+        d = self._maybe_fetch_offsets_and_header()
+        d.addCallback(lambda ignored:
+            (self._required_shares,
+             self._total_shares,
+             self._segment_size,
+             self._data_length))
+        return d
+
+
+    def get_seqnum(self):
+        """
+        I return the sequence number for this share.
+        """
+        d = self._maybe_fetch_offsets_and_header()
+        d.addCallback(lambda ignored:
+            self._sequence_number)
+        return d
+
+
+    def get_root_hash(self):
+        """
+        I return the root of the block hash tree
+        """
+        d = self._maybe_fetch_offsets_and_header()
+        d.addCallback(lambda ignored: self._root_hash)
+        return d
+
+
+    def get_checkstring(self):
+        """
+        I return the packed representation of the following:
+
+            - version number
+            - sequence number
+            - root hash
+            - salt hash
+
+        which my users use as a checkstring to detect other writers.
+        """
+        d = self._maybe_fetch_offsets_and_header()
+        def _build_checkstring(ignored):
+            if self._salt:
+                checkstring = strut.pack(PREFIX,
+                                         self._version_number,
+                                         self._sequence_number,
+                                         self._root_hash,
+                                         self._salt)
+            else:
+                checkstring = struct.pack(MDMFCHECKSTRING,
+                                          self._version_number,
+                                          self._sequence_number,
+                                          self._root_hash)
+
+            return checkstring
+        d.addCallback(_build_checkstring)
+        return d
+
+
+    def get_prefix(self, force_remote):
+        d = self._maybe_fetch_offsets_and_header(force_remote)
+        d.addCallback(lambda ignored:
+            self._build_prefix())
+        return d
+
+
+    def _build_prefix(self):
+        # The prefix is another name for the part of the remote share
+        # that gets signed. It consists of everything up to and
+        # including the datalength, packed by struct.
+        if self._version_number == SDMF_VERSION:
+            return struct.pack(SIGNED_PREFIX,
+                           self._version_number,
+                           self._sequence_number,
+                           self._root_hash,
+                           self._salt,
+                           self._required_shares,
+                           self._total_shares,
+                           self._segment_size,
+                           self._data_length)
+
+        else:
+            return struct.pack(MDMFSIGNABLEHEADER,
+                           self._version_number,
+                           self._sequence_number,
+                           self._root_hash,
+                           self._required_shares,
+                           self._total_shares,
+                           self._segment_size,
+                           self._data_length)
+
+
+    def _get_offsets_tuple(self):
+        # The offsets tuple is another component of the version
+        # information tuple. It is basically our offsets dictionary,
+        # itemized and in a tuple.
+        return self._offsets.copy()
+
+
+    def get_verinfo(self):
+        """
+        I return my verinfo tuple. This is used by the ServermapUpdater
+        to keep track of versions of mutable files.
+
+        The verinfo tuple for MDMF files contains:
+            - seqnum
+            - root hash
+            - a blank (nothing)
+            - segsize
+            - datalen
+            - k
+            - n
+            - prefix (the thing that you sign)
+            - a tuple of offsets
+
+        We include the nonce in MDMF to simplify processing of version
+        information tuples.
+
+        The verinfo tuple for SDMF files is the same, but contains a
+        16-byte IV instead of a hash of salts.
+        """
+        d = self._maybe_fetch_offsets_and_header()
+        def _build_verinfo(ignored):
+            if self._version_number == SDMF_VERSION:
+                salt_to_use = self._salt
+            else:
+                salt_to_use = None
+            return (self._sequence_number,
+                    self._root_hash,
+                    salt_to_use,
+                    self._segment_size,
+                    self._data_length,
+                    self._required_shares,
+                    self._total_shares,
+                    self._build_prefix(),
+                    self._get_offsets_tuple())
+        d.addCallback(_build_verinfo)
+        return d
+
+
+    def flush(self):
+        """
+        I flush my queue of read vectors.
+        """
+        d = self._read(self._readvs)
+        def _then(results):
+            self._readvs = []
+            if isinstance(results, failure.Failure):
+                self._queue_errbacks.notify(results)
+            else:
+                self._queue_observers.notify(results)
+            self._queue_observers = observer.ObserverList()
+            self._queue_errbacks = observer.ObserverList()
+        d.addBoth(_then)
+
+
+    def _read(self, readvs, force_remote=False, queue=False):
+        unsatisfiable = filter(lambda x: x[0] + x[1] > len(self._data), readvs)
+        # TODO: It's entirely possible to tweak this so that it just
+        # fulfills the requests that it can, and not demand that all
+        # requests are satisfiable before running it.
+        if not unsatisfiable and not force_remote:
+            results = [self._data[offset:offset+length]
+                       for (offset, length) in readvs]
+            results = {self.shnum: results}
+            return defer.succeed(results)
+        else:
+            if queue:
+                start = len(self._readvs)
+                self._readvs += readvs
+                end = len(self._readvs)
+                def _get_results(results, start, end):
+                    if not self.shnum in results:
+                        return {self._shnum: [""]}
+                    return {self.shnum: results[self.shnum][start:end]}
+                d = defer.Deferred()
+                d.addCallback(_get_results, start, end)
+                self._queue_observers.subscribe(d.callback)
+                self._queue_errbacks.subscribe(d.errback)
+                return d
+            return self._rref.callRemote("slot_readv",
+                                         self._storage_index,
+                                         [self.shnum],
+                                         readvs)
+
+
+    def is_sdmf(self):
+        """I tell my caller whether or not my remote file is SDMF or MDMF
+        """
+        d = self._maybe_fetch_offsets_and_header()
+        d.addCallback(lambda ignored:
+            self._version_number == 0)
+        return d
+
+
+class LayoutInvalid(Exception):
+    """
+    This isn't a valid MDMF mutable file
+    """
hunk ./src/allmydata/test/test_storage.py 2
 
-import time, os.path, stat, re, simplejson, struct
+import time, os.path, stat, re, simplejson, struct, shutil
 
 from twisted.trial import unittest
 
hunk ./src/allmydata/test/test_storage.py 22
 from allmydata.storage.expirer import LeaseCheckingCrawler
 from allmydata.immutable.layout import WriteBucketProxy, WriteBucketProxy_v2, \
      ReadBucketProxy
-from allmydata.interfaces import BadWriteEnablerError
-from allmydata.test.common import LoggingServiceParent
+from allmydata.mutable.layout import MDMFSlotWriteProxy, MDMFSlotReadProxy, \
+                                     LayoutInvalid, MDMFSIGNABLEHEADER, \
+                                     SIGNED_PREFIX, MDMFHEADER, MDMFOFFSETS
+from allmydata.interfaces import BadWriteEnablerError, MDMF_VERSION, \
+                                 SDMF_VERSION
+from allmydata.test.common import LoggingServiceParent, ShouldFailMixin
 from allmydata.test.common_web import WebRenderingMixin
 from allmydata.web.storage import StorageStatus, remove_prefix
 
hunk ./src/allmydata/test/test_storage.py 105
 
 class RemoteBucket:
 
+    def __init__(self):
+        self.read_count = 0
+        self.write_count = 0
+
     def callRemote(self, methname, *args, **kwargs):
         def _call():
             meth = getattr(self.target, "remote_" + methname)
hunk ./src/allmydata/test/test_storage.py 113
             return meth(*args, **kwargs)
+
+        if methname == "slot_readv":
+            self.read_count += 1
+        if methname == "slot_writev":
+            self.write_count += 1
+
         return defer.maybeDeferred(_call)
 
hunk ./src/allmydata/test/test_storage.py 121
+
 class BucketProxy(unittest.TestCase):
     def make_bucket(self, name, size):
         basedir = os.path.join("storage", "BucketProxy", name)
hunk ./src/allmydata/test/test_storage.py 1298
         self.failUnless(os.path.exists(prefixdir), prefixdir)
         self.failIf(os.path.exists(bucketdir), bucketdir)
 
+
+class MDMFProxies(unittest.TestCase, ShouldFailMixin):
+    def setUp(self):
+        self.sparent = LoggingServiceParent()
+        self._lease_secret = itertools.count()
+        self.ss = self.create("MDMFProxies storage test server")
+        self.rref = RemoteBucket()
+        self.rref.target = self.ss
+        self.secrets = (self.write_enabler("we_secret"),
+                        self.renew_secret("renew_secret"),
+                        self.cancel_secret("cancel_secret"))
+        self.segment = "aaaaaa"
+        self.block = "aa"
+        self.salt = "a" * 16
+        self.block_hash = "a" * 32
+        self.block_hash_tree = [self.block_hash for i in xrange(6)]
+        self.share_hash = self.block_hash
+        self.share_hash_chain = dict([(i, self.share_hash) for i in xrange(6)])
+        self.signature = "foobarbaz"
+        self.verification_key = "vvvvvv"
+        self.encprivkey = "private"
+        self.root_hash = self.block_hash
+        self.salt_hash = self.root_hash
+        self.salt_hash_tree = [self.salt_hash for i in xrange(6)]
+        self.block_hash_tree_s = self.serialize_blockhashes(self.block_hash_tree)
+        self.share_hash_chain_s = self.serialize_sharehashes(self.share_hash_chain)
+        # blockhashes and salt hashes are serialized in the same way,
+        # only we lop off the first element and store that in the
+        # header.
+        self.salt_hash_tree_s = self.serialize_blockhashes(self.salt_hash_tree[1:])
+
+
+    def tearDown(self):
+        self.sparent.stopService()
+        shutil.rmtree(self.workdir("MDMFProxies storage test server"))
+
+
+    def write_enabler(self, we_tag):
+        return hashutil.tagged_hash("we_blah", we_tag)
+
+
+    def renew_secret(self, tag):
+        return hashutil.tagged_hash("renew_blah", str(tag))
+
+
+    def cancel_secret(self, tag):
+        return hashutil.tagged_hash("cancel_blah", str(tag))
+
+
+    def workdir(self, name):
+        basedir = os.path.join("storage", "MutableServer", name)
+        return basedir
+
+
+    def create(self, name):
+        workdir = self.workdir(name)
+        ss = StorageServer(workdir, "\x00" * 20)
+        ss.setServiceParent(self.sparent)
+        return ss
+
+
+    def build_test_mdmf_share(self, tail_segment=False, empty=False):
+        # Start with the checkstring
+        data = struct.pack(">BQ32s",
+                           1,
+                           0,
+                           self.root_hash)
+        self.checkstring = data
+        # Next, the encoding parameters
+        if tail_segment:
+            data += struct.pack(">BBQQ",
+                                3,
+                                10,
+                                6,
+                                33)
+        elif empty:
+            data += struct.pack(">BBQQ",
+                                3,
+                                10,
+                                0,
+                                0)
+        else:
+            data += struct.pack(">BBQQ",
+                                3,
+                                10,
+                                6,
+                                36)
+        # Now we'll build the offsets.
+        sharedata = ""
+        if not tail_segment and not empty:
+            for i in xrange(6):
+                sharedata += self.salt + self.block
+        elif tail_segment:
+            for i in xrange(5):
+                sharedata += self.salt + self.block
+            sharedata += self.salt + "a"
+
+        # The encrypted private key comes after the shares + salts
+        offset_size = struct.calcsize(MDMFOFFSETS)
+        encrypted_private_key_offset = len(data) + offset_size + len(sharedata)
+        # The blockhashes come after the private key
+        blockhashes_offset = encrypted_private_key_offset + len(self.encprivkey)
+        # The sharehashes come after the salt hashes
+        sharehashes_offset = blockhashes_offset + len(self.block_hash_tree_s)
+        # The signature comes after the share hash chain
+        signature_offset = sharehashes_offset + len(self.share_hash_chain_s)
+        # The verification key comes after the signature
+        verification_offset = signature_offset + len(self.signature)
+        # The EOF comes after the verification key
+        eof_offset = verification_offset + len(self.verification_key)
+        data += struct.pack(MDMFOFFSETS,
+                            encrypted_private_key_offset,
+                            blockhashes_offset,
+                            sharehashes_offset,
+                            signature_offset,
+                            verification_offset,
+                            eof_offset)
+        self.offsets = {}
+        self.offsets['enc_privkey'] = encrypted_private_key_offset
+        self.offsets['block_hash_tree'] = blockhashes_offset
+        self.offsets['share_hash_chain'] = sharehashes_offset
+        self.offsets['signature'] = signature_offset
+        self.offsets['verification_key'] = verification_offset
+        self.offsets['EOF'] = eof_offset
+        # Next, we'll add in the salts and share data,
+        data += sharedata
+        # the private key,
+        data += self.encprivkey
+        # the block hash tree,
+        data += self.block_hash_tree_s
+        # the share hash chain,
+        data += self.share_hash_chain_s
+        # the signature,
+        data += self.signature
+        # and the verification key
+        data += self.verification_key
+        return data
+
+
+    def write_test_share_to_server(self,
+                                   storage_index,
+                                   tail_segment=False,
+                                   empty=False):
+        """
+        I write some data for the read tests to read to self.ss
+
+        If tail_segment=True, then I will write a share that has a
+        smaller tail segment than other segments.
+        """
+        write = self.ss.remote_slot_testv_and_readv_and_writev
+        data = self.build_test_mdmf_share(tail_segment, empty)
+        # Finally, we write the whole thing to the storage server in one
+        # pass.
+        testvs = [(0, 1, "eq", "")]
+        tws = {}
+        tws[0] = (testvs, [(0, data)], None)
+        readv = [(0, 1)]
+        results = write(storage_index, self.secrets, tws, readv)
+        self.failUnless(results[0])
+
+
+    def build_test_sdmf_share(self, empty=False):
+        if empty:
+            sharedata = ""
+        else:
+            sharedata = self.segment * 6
+        blocksize = len(sharedata) / 3
+        block = sharedata[:blocksize]
+        prefix = struct.pack(">BQ32s16s BBQQ",
+                             0, # version,
+                             0,
+                             self.root_hash,
+                             self.salt,
+                             3,
+                             10,
+                             len(sharedata),
+                             len(sharedata),
+                            )
+        post_offset = struct.calcsize(">BQ32s16sBBQQLLLLQQ")
+        signature_offset = post_offset + len(self.verification_key)
+        sharehashes_offset = signature_offset + len(self.signature)
+        blockhashes_offset = sharehashes_offset + len(self.share_hash_chain_s)
+        sharedata_offset = blockhashes_offset + len(self.block_hash_tree_s)
+        encprivkey_offset = sharedata_offset + len(block)
+        eof_offset = encprivkey_offset + len(self.encprivkey)
+        offsets = struct.pack(">LLLLQQ",
+                              signature_offset,
+                              sharehashes_offset,
+                              blockhashes_offset,
+                              sharedata_offset,
+                              encprivkey_offset,
+                              eof_offset)
+        final_share = "".join([prefix,
+                           offsets,
+                           self.verification_key,
+                           self.signature,
+                           self.share_hash_chain_s,
+                           self.block_hash_tree_s,
+                           block,
+                           self.encprivkey])
+        self.offsets = {}
+        self.offsets['signature'] = signature_offset
+        self.offsets['share_hash_chain'] = sharehashes_offset
+        self.offsets['block_hash_tree'] = blockhashes_offset
+        self.offsets['share_data'] = sharedata_offset
+        self.offsets['enc_privkey'] = encprivkey_offset
+        self.offsets['EOF'] = eof_offset
+        return final_share
+
+
+    def write_sdmf_share_to_server(self,
+                                   storage_index,
+                                   empty=False):
+        # Some tests need SDMF shares to verify that we can still 
+        # read them. This method writes one, which resembles but is not
+        assert self.rref
+        write = self.ss.remote_slot_testv_and_readv_and_writev
+        share = self.build_test_sdmf_share(empty)
+        testvs = [(0, 1, "eq", "")]
+        tws = {}
+        tws[0] = (testvs, [(0, share)], None)
+        readv = []
+        results = write(storage_index, self.secrets, tws, readv)
+        self.failUnless(results[0])
+
+
+    def test_read(self):
+        self.write_test_share_to_server("si1")
+        mr = MDMFSlotReadProxy(self.rref, "si1", 0)
+        # Check that every method equals what we expect it to.
+        d = defer.succeed(None)
+        def _check_block_and_salt((block, salt)):
+            self.failUnlessEqual(block, self.block)
+            self.failUnlessEqual(salt, self.salt)
+
+        for i in xrange(6):
+            d.addCallback(lambda ignored, i=i:
+                mr.get_block_and_salt(i))
+            d.addCallback(_check_block_and_salt)
+
+        d.addCallback(lambda ignored:
+            mr.get_encprivkey())
+        d.addCallback(lambda encprivkey:
+            self.failUnlessEqual(self.encprivkey, encprivkey))
+
+        d.addCallback(lambda ignored:
+            mr.get_blockhashes())
+        d.addCallback(lambda blockhashes:
+            self.failUnlessEqual(self.block_hash_tree, blockhashes))
+
+        d.addCallback(lambda ignored:
+            mr.get_sharehashes())
+        d.addCallback(lambda sharehashes:
+            self.failUnlessEqual(self.share_hash_chain, sharehashes))
+
+        d.addCallback(lambda ignored:
+            mr.get_signature())
+        d.addCallback(lambda signature:
+            self.failUnlessEqual(signature, self.signature))
+
+        d.addCallback(lambda ignored:
+            mr.get_verification_key())
+        d.addCallback(lambda verification_key:
+            self.failUnlessEqual(verification_key, self.verification_key))
+
+        d.addCallback(lambda ignored:
+            mr.get_seqnum())
+        d.addCallback(lambda seqnum:
+            self.failUnlessEqual(seqnum, 0))
+
+        d.addCallback(lambda ignored:
+            mr.get_root_hash())
+        d.addCallback(lambda root_hash:
+            self.failUnlessEqual(self.root_hash, root_hash))
+
+        d.addCallback(lambda ignored:
+            mr.get_seqnum())
+        d.addCallback(lambda seqnum:
+            self.failUnlessEqual(0, seqnum))
+
+        d.addCallback(lambda ignored:
+            mr.get_encoding_parameters())
+        def _check_encoding_parameters((k, n, segsize, datalen)):
+            self.failUnlessEqual(k, 3)
+            self.failUnlessEqual(n, 10)
+            self.failUnlessEqual(segsize, 6)
+            self.failUnlessEqual(datalen, 36)
+        d.addCallback(_check_encoding_parameters)
+
+        d.addCallback(lambda ignored:
+            mr.get_checkstring())
+        d.addCallback(lambda checkstring:
+            self.failUnlessEqual(checkstring, checkstring))
+        return d
+
+
+    def test_read_with_different_tail_segment_size(self):
+        self.write_test_share_to_server("si1", tail_segment=True)
+        mr = MDMFSlotReadProxy(self.rref, "si1", 0)
+        d = mr.get_block_and_salt(5)
+        def _check_tail_segment(results):
+            block, salt = results
+            self.failUnlessEqual(len(block), 1)
+            self.failUnlessEqual(block, "a")
+        d.addCallback(_check_tail_segment)
+        return d
+
+
+    def test_get_block_with_invalid_segnum(self):
+        self.write_test_share_to_server("si1")
+        mr = MDMFSlotReadProxy(self.rref, "si1", 0)
+        d = defer.succeed(None)
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "test invalid segnum",
+                            None,
+                            mr.get_block_and_salt, 7))
+        return d
+
+
+    def test_get_encoding_parameters_first(self):
+        self.write_test_share_to_server("si1")
+        mr = MDMFSlotReadProxy(self.rref, "si1", 0)
+        d = mr.get_encoding_parameters()
+        def _check_encoding_parameters((k, n, segment_size, datalen)):
+            self.failUnlessEqual(k, 3)
+            self.failUnlessEqual(n, 10)
+            self.failUnlessEqual(segment_size, 6)
+            self.failUnlessEqual(datalen, 36)
+        d.addCallback(_check_encoding_parameters)
+        return d
+
+
+    def test_get_seqnum_first(self):
+        self.write_test_share_to_server("si1")
+        mr = MDMFSlotReadProxy(self.rref, "si1", 0)
+        d = mr.get_seqnum()
+        d.addCallback(lambda seqnum:
+            self.failUnlessEqual(seqnum, 0))
+        return d
+
+
+    def test_get_root_hash_first(self):
+        self.write_test_share_to_server("si1")
+        mr = MDMFSlotReadProxy(self.rref, "si1", 0)
+        d = mr.get_root_hash()
+        d.addCallback(lambda root_hash:
+            self.failUnlessEqual(root_hash, self.root_hash))
+        return d
+
+
+    def test_get_checkstring_first(self):
+        self.write_test_share_to_server("si1")
+        mr = MDMFSlotReadProxy(self.rref, "si1", 0)
+        d = mr.get_checkstring()
+        d.addCallback(lambda checkstring:
+            self.failUnlessEqual(checkstring, self.checkstring))
+        return d
+
+
+    def test_write_read_vectors(self):
+        # When writing for us, the storage server will return to us a
+        # read vector, along with its result. If a write fails because
+        # the test vectors failed, this read vector can help us to
+        # diagnose the problem. This test ensures that the read vector
+        # is working appropriately.
+        mw = self._make_new_mw("si1", 0)
+        d = defer.succeed(None)
+
+        # Write one share. This should return a checkstring of nothing,
+        # since there is no data there.
+        d.addCallback(lambda ignored:
+            mw.put_block(self.block, 0, self.salt))
+        def _check_first_write(results):
+            result, readvs = results
+            self.failUnless(result)
+            self.failIf(readvs)
+        d.addCallback(_check_first_write)
+        # Now, there should be a different checkstring returned when
+        # we write other shares
+        d.addCallback(lambda ignored:
+            mw.put_block(self.block, 1, self.salt))
+        def _check_next_write(results):
+            result, readvs = results
+            self.failUnless(result)
+            self.expected_checkstring = mw.get_checkstring()
+            self.failUnlessIn(0, readvs)
+            self.failUnlessEqual(readvs[0][0], self.expected_checkstring)
+        d.addCallback(_check_next_write)
+        # Add the other four shares
+        for i in xrange(2, 6):
+            d.addCallback(lambda ignored, i=i:
+                mw.put_block(self.block, i, self.salt))
+            d.addCallback(_check_next_write)
+        # Add the encrypted private key
+        d.addCallback(lambda ignored:
+            mw.put_encprivkey(self.encprivkey))
+        d.addCallback(_check_next_write)
+        # Add the block hash tree and share hash tree
+        d.addCallback(lambda ignored:
+            mw.put_blockhashes(self.block_hash_tree))
+        d.addCallback(_check_next_write)
+        d.addCallback(lambda ignored:
+            mw.put_sharehashes(self.share_hash_chain))
+        d.addCallback(_check_next_write)
+        # Add the root hash and the salt hash. This should change the
+        # checkstring, but not in a way that we'll be able to see right
+        # now, since the read vectors are applied before the write
+        # vectors.
+        d.addCallback(lambda ignored:
+            mw.put_root_hash(self.root_hash))
+        def _check_old_testv_after_new_one_is_written(results):
+            result, readvs = results
+            self.failUnless(result)
+            self.failUnlessIn(0, readvs)
+            self.failUnlessEqual(self.expected_checkstring,
+                                 readvs[0][0])
+            new_checkstring = mw.get_checkstring()
+            self.failIfEqual(new_checkstring,
+                             readvs[0][0])
+        d.addCallback(_check_old_testv_after_new_one_is_written)
+        # Now add the signature. This should succeed, meaning that the
+        # data gets written and the read vector matches what the writer
+        # thinks should be there.
+        d.addCallback(lambda ignored:
+            mw.put_signature(self.signature))
+        d.addCallback(_check_next_write)
+        # The checkstring remains the same for the rest of the process.
+        return d
+
+
+    def test_blockhashes_after_share_hash_chain(self):
+        mw = self._make_new_mw("si1", 0)
+        d = defer.succeed(None)
+        # Put everything up to and including the share hash chain
+        for i in xrange(6):
+            d.addCallback(lambda ignored, i=i:
+                mw.put_block(self.block, i, self.salt))
+        d.addCallback(lambda ignored:
+            mw.put_encprivkey(self.encprivkey))
+        d.addCallback(lambda ignored:
+            mw.put_blockhashes(self.block_hash_tree))
+        d.addCallback(lambda ignored:
+            mw.put_sharehashes(self.share_hash_chain))
+
+        # Now try to put the block hash tree again.
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "test repeat salthashes",
+                            None,
+                            mw.put_blockhashes, self.block_hash_tree))
+        return d
+
+
+    def test_encprivkey_after_blockhashes(self):
+        mw = self._make_new_mw("si1", 0)
+        d = defer.succeed(None)
+        # Put everything up to and including the block hash tree
+        for i in xrange(6):
+            d.addCallback(lambda ignored, i=i:
+                mw.put_block(self.block, i, self.salt))
+        d.addCallback(lambda ignored:
+            mw.put_encprivkey(self.encprivkey))
+        d.addCallback(lambda ignored:
+            mw.put_blockhashes(self.block_hash_tree))
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "out of order private key",
+                            None,
+                            mw.put_encprivkey, self.encprivkey))
+        return d
+
+
+    def test_share_hash_chain_after_signature(self):
+        mw = self._make_new_mw("si1", 0)
+        d = defer.succeed(None)
+        # Put everything up to and including the signature
+        for i in xrange(6):
+            d.addCallback(lambda ignored, i=i:
+                mw.put_block(self.block, i, self.salt))
+        d.addCallback(lambda ignored:
+            mw.put_encprivkey(self.encprivkey))
+        d.addCallback(lambda ignored:
+            mw.put_blockhashes(self.block_hash_tree))
+        d.addCallback(lambda ignored:
+            mw.put_sharehashes(self.share_hash_chain))
+        d.addCallback(lambda ignored:
+            mw.put_root_hash(self.root_hash))
+        d.addCallback(lambda ignored:
+            mw.put_signature(self.signature))
+        # Now try to put the share hash chain again. This should fail
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "out of order share hash chain",
+                            None,
+                            mw.put_sharehashes, self.share_hash_chain))
+        return d
+
+
+    def test_signature_after_verification_key(self):
+        mw = self._make_new_mw("si1", 0)
+        d = defer.succeed(None)
+        # Put everything up to and including the verification key.
+        for i in xrange(6):
+            d.addCallback(lambda ignored, i=i:
+                mw.put_block(self.block, i, self.salt))
+        d.addCallback(lambda ignored:
+            mw.put_encprivkey(self.encprivkey))
+        d.addCallback(lambda ignored:
+            mw.put_blockhashes(self.block_hash_tree))
+        d.addCallback(lambda ignored:
+            mw.put_sharehashes(self.share_hash_chain))
+        d.addCallback(lambda ignored:
+            mw.put_root_hash(self.root_hash))
+        d.addCallback(lambda ignored:
+            mw.put_signature(self.signature))
+        d.addCallback(lambda ignored:
+            mw.put_verification_key(self.verification_key))
+        # Now try to put the signature again. This should fail
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "signature after verification",
+                            None,
+                            mw.put_signature, self.signature))
+        return d
+
+
+    def test_uncoordinated_write(self):
+        # Make two mutable writers, both pointing to the same storage
+        # server, both at the same storage index, and try writing to the
+        # same share.
+        mw1 = self._make_new_mw("si1", 0)
+        mw2 = self._make_new_mw("si1", 0)
+        d = defer.succeed(None)
+        def _check_success(results):
+            result, readvs = results
+            self.failUnless(result)
+
+        def _check_failure(results):
+            result, readvs = results
+            self.failIf(result)
+
+        d.addCallback(lambda ignored:
+            mw1.put_block(self.block, 0, self.salt))
+        d.addCallback(_check_success)
+        d.addCallback(lambda ignored:
+            mw2.put_block(self.block, 0, self.salt))
+        d.addCallback(_check_failure)
+        return d
+
+
+    def test_invalid_salt_size(self):
+        # Salts need to be 16 bytes in size. Writes that attempt to
+        # write more or less than this should be rejected.
+        mw = self._make_new_mw("si1", 0)
+        invalid_salt = "a" * 17 # 17 bytes
+        another_invalid_salt = "b" * 15 # 15 bytes
+        d = defer.succeed(None)
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "salt too big",
+                            None,
+                            mw.put_block, self.block, 0, invalid_salt))
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "salt too small",
+                            None,
+                            mw.put_block, self.block, 0,
+                            another_invalid_salt))
+        return d
+
+
+    def test_write_test_vectors(self):
+        # If we give the write proxy a bogus test vector at 
+        # any point during the process, it should fail to write.
+        mw = self._make_new_mw("si1", 0)
+        mw.set_checkstring("this is a lie")
+        # The initial write should be expecting to find the improbable
+        # checkstring above in place; finding nothing, it should fail.
+        d = defer.succeed(None)
+        d.addCallback(lambda ignored:
+            mw.put_block(self.block, 0, self.salt))
+        def _check_failure(results):
+            result, readv = results
+            self.failIf(result)
+        d.addCallback(_check_failure)
+        # Now set the checkstring to the empty string, which
+        # indicates that no share is there.
+        d.addCallback(lambda ignored:
+            mw.set_checkstring(""))
+        d.addCallback(lambda ignored:
+            mw.put_block(self.block, 0, self.salt))
+        def _check_success(results):
+            result, readv = results
+            self.failUnless(result)
+        d.addCallback(_check_success)
+        # Now set the checkstring to something wrong
+        d.addCallback(lambda ignored:
+            mw.set_checkstring("something wrong"))
+        # This should fail to do anything
+        d.addCallback(lambda ignored:
+            mw.put_block(self.block, 1, self.salt))
+        d.addCallback(_check_failure)
+        # Now set it back to what it should be.
+        d.addCallback(lambda ignored:
+            mw.set_checkstring(mw.get_checkstring()))
+        for i in xrange(1, 6):
+            d.addCallback(lambda ignored, i=i:
+                mw.put_block(self.block, i, self.salt))
+            d.addCallback(_check_success)
+        d.addCallback(lambda ignored:
+            mw.put_encprivkey(self.encprivkey))
+        d.addCallback(_check_success)
+        d.addCallback(lambda ignored:
+            mw.put_blockhashes(self.block_hash_tree))
+        d.addCallback(_check_success)
+        d.addCallback(lambda ignored:
+            mw.put_sharehashes(self.share_hash_chain))
+        d.addCallback(_check_success)
+        def _keep_old_checkstring(ignored):
+            self.old_checkstring = mw.get_checkstring()
+            mw.set_checkstring("foobarbaz")
+        d.addCallback(_keep_old_checkstring)
+        d.addCallback(lambda ignored:
+            mw.put_root_hash(self.root_hash))
+        d.addCallback(_check_failure)
+        d.addCallback(lambda ignored:
+            self.failUnlessEqual(self.old_checkstring, mw.get_checkstring()))
+        def _restore_old_checkstring(ignored):
+            mw.set_checkstring(self.old_checkstring)
+        d.addCallback(_restore_old_checkstring)
+        d.addCallback(lambda ignored:
+            mw.put_root_hash(self.root_hash))
+        d.addCallback(_check_success)
+        # The checkstring should have been set appropriately for us on
+        # the last write; if we try to change it to something else,
+        # that change should cause the verification key step to fail.
+        d.addCallback(lambda ignored:
+            mw.set_checkstring("something else"))
+        d.addCallback(lambda ignored:
+            mw.put_signature(self.signature))
+        d.addCallback(_check_failure)
+        d.addCallback(lambda ignored:
+            mw.set_checkstring(mw.get_checkstring()))
+        d.addCallback(lambda ignored:
+            mw.put_signature(self.signature))
+        d.addCallback(_check_success)
+        d.addCallback(lambda ignored:
+            mw.put_verification_key(self.verification_key))
+        d.addCallback(_check_success)
+        return d
+
+
+    def test_offset_only_set_on_success(self):
+        # The write proxy should be smart enough to detect when a write
+        # has failed, and to temper its definition of progress based on
+        # that.
+        mw = self._make_new_mw("si1", 0)
+        d = defer.succeed(None)
+        for i in xrange(1, 6):
+            d.addCallback(lambda ignored, i=i:
+                mw.put_block(self.block, i, self.salt))
+        def _break_checkstring(ignored):
+            self._old_checkstring = mw.get_checkstring()
+            mw.set_checkstring("foobarbaz")
+
+        def _fix_checkstring(ignored):
+            mw.set_checkstring(self._old_checkstring)
+
+        d.addCallback(_break_checkstring)
+
+        # Setting the encrypted private key shouldn't work now, which is
+        # to be expected and is tested elsewhere. We also want to make
+        # sure that we can't add the block hash tree after a failed
+        # write of this sort.
+        d.addCallback(lambda ignored:
+            mw.put_encprivkey(self.encprivkey))
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "test out-of-order blockhashes",
+                            None,
+                            mw.put_blockhashes, self.block_hash_tree))
+        d.addCallback(_fix_checkstring)
+        d.addCallback(lambda ignored:
+            mw.put_encprivkey(self.encprivkey))
+        d.addCallback(_break_checkstring)
+        d.addCallback(lambda ignored:
+            mw.put_blockhashes(self.block_hash_tree))
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "test out-of-order sharehashes",
+                            None,
+                            mw.put_sharehashes, self.share_hash_chain))
+        d.addCallback(_fix_checkstring)
+        d.addCallback(lambda ignored:
+            mw.put_blockhashes(self.block_hash_tree))
+        d.addCallback(_break_checkstring)
+        d.addCallback(lambda ignored:
+            mw.put_sharehashes(self.share_hash_chain))
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "out-of-order root hash",
+                            None,
+                            mw.put_root_hash, self.root_hash))
+        d.addCallback(_fix_checkstring)
+        d.addCallback(lambda ignored:
+            mw.put_sharehashes(self.share_hash_chain))
+        d.addCallback(_break_checkstring)
+        d.addCallback(lambda ignored:
+            mw.put_root_hash(self.root_hash))
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "out-of-order signature",
+                            None,
+                            mw.put_signature, self.signature))
+        d.addCallback(_fix_checkstring)
+        d.addCallback(lambda ignored:
+            mw.put_root_hash(self.root_hash))
+        d.addCallback(_break_checkstring)
+        d.addCallback(lambda ignored:
+            mw.put_signature(self.signature))
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "out-of-order verification key",
+                            None,
+                            mw.put_verification_key,
+                            self.verification_key))
+        d.addCallback(_fix_checkstring)
+        d.addCallback(lambda ignored:
+            mw.put_signature(self.signature))
+        d.addCallback(_break_checkstring)
+        d.addCallback(lambda ignored:
+            mw.put_verification_key(self.verification_key))
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "out-of-order finish",
+                            None,
+                            mw.finish_publishing))
+        return d
+
+
+    def serialize_blockhashes(self, blockhashes):
+        return "".join(blockhashes)
+
+
+    def serialize_sharehashes(self, sharehashes):
+        ret = "".join([struct.pack(">H32s", i, sharehashes[i])
+                        for i in sorted(sharehashes.keys())])
+        return ret
+
+
+    def test_write(self):
+        # This translates to a file with 6 6-byte segments, and with 2-byte
+        # blocks.
+        mw = self._make_new_mw("si1", 0)
+        mw2 = self._make_new_mw("si1", 1)
+        # Test writing some blocks.
+        read = self.ss.remote_slot_readv
+        expected_sharedata_offset = struct.calcsize(MDMFHEADER)
+        written_block_size = 2 + len(self.salt)
+        written_block = self.block + self.salt
+        def _check_block_write(i, share):
+            self.failUnlessEqual(read("si1", [share], [(expected_sharedata_offset + (i * written_block_size), written_block_size)]),
+                                {share: [written_block]})
+        d = defer.succeed(None)
+        for i in xrange(6):
+            d.addCallback(lambda ignored, i=i:
+                mw.put_block(self.block, i, self.salt))
+            d.addCallback(lambda ignored, i=i:
+                _check_block_write(i, 0))
+        # Now try the same thing, but with share 1 instead of share 0.
+        for i in xrange(6):
+            d.addCallback(lambda ignored, i=i:
+                mw2.put_block(self.block, i, self.salt))
+            d.addCallback(lambda ignored, i=i:
+                _check_block_write(i, 1))
+
+        # Next, we make a fake encrypted private key, and put it onto the 
+        # storage server.
+        d.addCallback(lambda ignored:
+            mw.put_encprivkey(self.encprivkey))
+        expected_private_key_offset = expected_sharedata_offset + \
+                                      len(written_block) * 6
+        self.failUnlessEqual(len(self.encprivkey), 7)
+        d.addCallback(lambda ignored:
+            self.failUnlessEqual(read("si1", [0], [(expected_private_key_offset, 7)]),
+                                 {0: [self.encprivkey]}))
+
+        # Next, we put a fake block hash tree.
+        d.addCallback(lambda ignored:
+            mw.put_blockhashes(self.block_hash_tree))
+        expected_block_hash_offset = expected_private_key_offset + len(self.encprivkey)
+        self.failUnlessEqual(len(self.block_hash_tree_s), 32 * 6)
+        d.addCallback(lambda ignored:
+            self.failUnlessEqual(read("si1", [0], [(expected_block_hash_offset, 32 * 6)]),
+                                 {0: [self.block_hash_tree_s]}))
+
+        # Next, put a fake share hash chain
+        d.addCallback(lambda ignored:
+            mw.put_sharehashes(self.share_hash_chain))
+        expected_share_hash_offset = expected_block_hash_offset + len(self.block_hash_tree_s)
+        d.addCallback(lambda ignored:
+            self.failUnlessEqual(read("si1", [0],[(expected_share_hash_offset, (32 + 2) * 6)]),
+                                 {0: [self.share_hash_chain_s]}))
+
+        # Next, we put what is supposed to be the root hash of
+        # our share hash tree but isn't        
+        d.addCallback(lambda ignored:
+            mw.put_root_hash(self.root_hash))
+        # The root hash gets inserted at byte 9 (its position is in the header,
+        # and is fixed). 
+        def _check(ignored):
+            self.failUnlessEqual(read("si1", [0], [(9, 32)]),
+                                 {0: [self.root_hash]})
+        d.addCallback(_check)
+
+        # Next, we put a signature of the header block.
+        d.addCallback(lambda ignored:
+            mw.put_signature(self.signature))
+        expected_signature_offset = expected_share_hash_offset + len(self.share_hash_chain_s)
+        self.failUnlessEqual(len(self.signature), 9)
+        d.addCallback(lambda ignored:
+            self.failUnlessEqual(read("si1", [0], [(expected_signature_offset, 9)]),
+                                 {0: [self.signature]}))
+
+        # Next, we put the verification key
+        d.addCallback(lambda ignored:
+            mw.put_verification_key(self.verification_key))
+        expected_verification_key_offset = expected_signature_offset + len(self.signature)
+        self.failUnlessEqual(len(self.verification_key), 6)
+        d.addCallback(lambda ignored:
+            self.failUnlessEqual(read("si1", [0], [(expected_verification_key_offset, 6)]),
+                                 {0: [self.verification_key]}))
+
+        def _check_signable(ignored):
+            # Make sure that the signable is what we think it should be.
+            signable = mw.get_signable()
+            verno, seq, roothash, k, n, segsize, datalen = \
+                                            struct.unpack(">BQ32sBBQQ",
+                                                          signable)
+            self.failUnlessEqual(verno, 1)
+            self.failUnlessEqual(seq, 0)
+            self.failUnlessEqual(roothash, self.root_hash)
+            self.failUnlessEqual(k, 3)
+            self.failUnlessEqual(n, 10)
+            self.failUnlessEqual(segsize, 6)
+            self.failUnlessEqual(datalen, 36)
+        d.addCallback(_check_signable)
+        # Next, we cause the offset table to be published.
+        d.addCallback(lambda ignored:
+            mw.finish_publishing())
+        expected_eof_offset = expected_verification_key_offset + len(self.verification_key)
+
+        def _check_offsets(ignored):
+            # Check the version number to make sure that it is correct.
+            expected_version_number = struct.pack(">B", 1)
+            self.failUnlessEqual(read("si1", [0], [(0, 1)]),
+                                 {0: [expected_version_number]})
+            # Check the sequence number to make sure that it is correct
+            expected_sequence_number = struct.pack(">Q", 0)
+            self.failUnlessEqual(read("si1", [0], [(1, 8)]),
+                                 {0: [expected_sequence_number]})
+            # Check that the encoding parameters (k, N, segement size, data
+            # length) are what they should be. These are  3, 10, 6, 36
+            expected_k = struct.pack(">B", 3)
+            self.failUnlessEqual(read("si1", [0], [(41, 1)]),
+                                 {0: [expected_k]})
+            expected_n = struct.pack(">B", 10)
+            self.failUnlessEqual(read("si1", [0], [(42, 1)]),
+                                 {0: [expected_n]})
+            expected_segment_size = struct.pack(">Q", 6)
+            self.failUnlessEqual(read("si1", [0], [(43, 8)]),
+                                 {0: [expected_segment_size]})
+            expected_data_length = struct.pack(">Q", 36)
+            self.failUnlessEqual(read("si1", [0], [(51, 8)]),
+                                 {0: [expected_data_length]})
+            expected_offset = struct.pack(">Q", expected_private_key_offset)
+            self.failUnlessEqual(read("si1", [0], [(59, 8)]),
+                                 {0: [expected_offset]})
+            expected_offset = struct.pack(">Q", expected_block_hash_offset)
+            self.failUnlessEqual(read("si1", [0], [(67, 8)]),
+                                 {0: [expected_offset]})
+            expected_offset = struct.pack(">Q", expected_share_hash_offset)
+            self.failUnlessEqual(read("si1", [0], [(75, 8)]),
+                                 {0: [expected_offset]})
+            expected_offset = struct.pack(">Q", expected_signature_offset)
+            self.failUnlessEqual(read("si1", [0], [(83, 8)]),
+                                 {0: [expected_offset]})
+            expected_offset = struct.pack(">Q", expected_verification_key_offset)
+            self.failUnlessEqual(read("si1", [0], [(91, 8)]),
+                                 {0: [expected_offset]})
+            expected_offset = struct.pack(">Q", expected_eof_offset)
+            self.failUnlessEqual(read("si1", [0], [(99, 8)]),
+                                 {0: [expected_offset]})
+        d.addCallback(_check_offsets)
+        return d
+
+    def _make_new_mw(self, si, share, datalength=36):
+        # This is a file of size 36 bytes. Since it has a segment
+        # size of 6, we know that it has 6 byte segments, which will
+        # be split into blocks of 2 bytes because our FEC k
+        # parameter is 3.
+        mw = MDMFSlotWriteProxy(share, self.rref, si, self.secrets, 0, 3, 10,
+                                6, datalength)
+        return mw
+
+
+    def test_write_rejected_with_too_many_blocks(self):
+        mw = self._make_new_mw("si0", 0)
+
+        # Try writing too many blocks. We should not be able to write
+        # more than 6
+        # blocks into each share.
+        d = defer.succeed(None)
+        for i in xrange(6):
+            d.addCallback(lambda ignored, i=i:
+                mw.put_block(self.block, i, self.salt))
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "too many blocks",
+                            None,
+                            mw.put_block, self.block, 7, self.salt))
+        return d
+
+
+    def test_write_rejected_with_invalid_salt(self):
+        # Try writing an invalid salt. Salts are 16 bytes -- any more or
+        # less should cause an error.
+        mw = self._make_new_mw("si1", 0)
+        bad_salt = "a" * 17 # 17 bytes
+        d = defer.succeed(None)
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "test_invalid_salt",
+                            None, mw.put_block, self.block, 7, bad_salt))
+        return d
+
+
+    def test_write_rejected_with_invalid_root_hash(self):
+        # Try writing an invalid root hash. This should be SHA256d, and
+        # 32 bytes long as a result.
+        mw = self._make_new_mw("si2", 0)
+        # 17 bytes != 32 bytes
+        invalid_root_hash = "a" * 17
+        d = defer.succeed(None)
+        # Before this test can work, we need to put some blocks + salts,
+        # a block hash tree, and a share hash tree. Otherwise, we'll see
+        # failures that match what we are looking for, but are caused by
+        # the constraints imposed on operation ordering.
+        for i in xrange(6):
+            d.addCallback(lambda ignored, i=i:
+                mw.put_block(self.block, i, self.salt))
+        d.addCallback(lambda ignored:
+            mw.put_encprivkey(self.encprivkey))
+        d.addCallback(lambda ignored:
+            mw.put_blockhashes(self.block_hash_tree))
+        d.addCallback(lambda ignored:
+            mw.put_sharehashes(self.share_hash_chain))
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "invalid root hash",
+                            None, mw.put_root_hash, invalid_root_hash))
+        return d
+
+
+    def test_write_rejected_with_invalid_blocksize(self):
+        # The blocksize implied by the writer that we get from
+        # _make_new_mw is 2bytes -- any more or any less than this
+        # should be cause for failure, unless it is the tail segment, in
+        # which case it may not be failure.
+        invalid_block = "a"
+        mw = self._make_new_mw("si3", 0, 33) # implies a tail segment with
+                                             # one byte blocks
+        # 1 bytes != 2 bytes
+        d = defer.succeed(None)
+        d.addCallback(lambda ignored, invalid_block=invalid_block:
+            self.shouldFail(LayoutInvalid, "test blocksize too small",
+                            None, mw.put_block, invalid_block, 0,
+                            self.salt))
+        invalid_block = invalid_block * 3
+        # 3 bytes != 2 bytes
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "test blocksize too large",
+                            None,
+                            mw.put_block, invalid_block, 0, self.salt))
+        for i in xrange(5):
+            d.addCallback(lambda ignored, i=i:
+                mw.put_block(self.block, i, self.salt))
+        # Try to put an invalid tail segment
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "test invalid tail segment",
+                            None,
+                            mw.put_block, self.block, 5, self.salt))
+        valid_block = "a"
+        d.addCallback(lambda ignored:
+            mw.put_block(valid_block, 5, self.salt))
+        return d
+
+
+    def test_write_enforces_order_constraints(self):
+        # We require that the MDMFSlotWriteProxy be interacted with in a
+        # specific way.
+        # That way is:
+        # 0: __init__
+        # 1: write blocks and salts
+        # 2: Write the encrypted private key
+        # 3: Write the block hashes
+        # 4: Write the share hashes
+        # 5: Write the root hash and salt hash
+        # 6: Write the signature and verification key
+        # 7: Write the file.
+        # 
+        # Some of these can be performed out-of-order, and some can't.
+        # The dependencies that I want to test here are:
+        #  - Private key before block hashes
+        #  - share hashes and block hashes before root hash
+        #  - root hash before signature
+        #  - signature before verification key
+        mw0 = self._make_new_mw("si0", 0)
+        # Write some shares
+        d = defer.succeed(None)
+        for i in xrange(6):
+            d.addCallback(lambda ignored, i=i:
+                mw0.put_block(self.block, i, self.salt))
+        # Try to write the block hashes before writing the encrypted
+        # private key
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "block hashes before key",
+                            None, mw0.put_blockhashes,
+                            self.block_hash_tree))
+
+        # Write the private key.
+        d.addCallback(lambda ignored:
+            mw0.put_encprivkey(self.encprivkey))
+
+
+        # Try to write the share hash chain without writing the block
+        # hash tree
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "share hash chain before "
+                                           "salt hash tree",
+                            None,
+                            mw0.put_sharehashes, self.share_hash_chain))
+
+        # Try to write the root hash and without writing either the
+        # block hashes or the or the share hashes
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "root hash before share hashes",
+                            None,
+                            mw0.put_root_hash, self.root_hash))
+
+        # Now write the block hashes and try again
+        d.addCallback(lambda ignored:
+            mw0.put_blockhashes(self.block_hash_tree))
+
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "root hash before share hashes",
+                            None, mw0.put_root_hash, self.root_hash))
+
+        # We haven't yet put the root hash on the share, so we shouldn't
+        # be able to sign it.
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "signature before root hash",
+                            None, mw0.put_signature, self.signature))
+
+        d.addCallback(lambda ignored:
+            self.failUnlessRaises(LayoutInvalid, mw0.get_signable))
+
+        # ..and, since that fails, we also shouldn't be able to put the
+        # verification key.
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "key before signature",
+                            None, mw0.put_verification_key,
+                            self.verification_key))
+
+        # Now write the share hashes.
+        d.addCallback(lambda ignored:
+            mw0.put_sharehashes(self.share_hash_chain))
+        # We should be able to write the root hash now too
+        d.addCallback(lambda ignored:
+            mw0.put_root_hash(self.root_hash))
+
+        # We should still be unable to put the verification key
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "key before signature",
+                            None, mw0.put_verification_key,
+                            self.verification_key))
+
+        d.addCallback(lambda ignored:
+            mw0.put_signature(self.signature))
+
+        # We shouldn't be able to write the offsets to the remote server
+        # until the offset table is finished; IOW, until we have written
+        # the verification key.
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "offsets before verification key",
+                            None,
+                            mw0.finish_publishing))
+
+        d.addCallback(lambda ignored:
+            mw0.put_verification_key(self.verification_key))
+        return d
+
+
+    def test_end_to_end(self):
+        mw = self._make_new_mw("si1", 0)
+        # Write a share using the mutable writer, and make sure that the
+        # reader knows how to read everything back to us.
+        d = defer.succeed(None)
+        for i in xrange(6):
+            d.addCallback(lambda ignored, i=i:
+                mw.put_block(self.block, i, self.salt))
+        d.addCallback(lambda ignored:
+            mw.put_encprivkey(self.encprivkey))
+        d.addCallback(lambda ignored:
+            mw.put_blockhashes(self.block_hash_tree))
+        d.addCallback(lambda ignored:
+            mw.put_sharehashes(self.share_hash_chain))
+        d.addCallback(lambda ignored:
+            mw.put_root_hash(self.root_hash))
+        d.addCallback(lambda ignored:
+            mw.put_signature(self.signature))
+        d.addCallback(lambda ignored:
+            mw.put_verification_key(self.verification_key))
+        d.addCallback(lambda ignored:
+            mw.finish_publishing())
+
+        mr = MDMFSlotReadProxy(self.rref, "si1", 0)
+        def _check_block_and_salt((block, salt)):
+            self.failUnlessEqual(block, self.block)
+            self.failUnlessEqual(salt, self.salt)
+
+        for i in xrange(6):
+            d.addCallback(lambda ignored, i=i:
+                mr.get_block_and_salt(i))
+            d.addCallback(_check_block_and_salt)
+
+        d.addCallback(lambda ignored:
+            mr.get_encprivkey())
+        d.addCallback(lambda encprivkey:
+            self.failUnlessEqual(self.encprivkey, encprivkey))
+
+        d.addCallback(lambda ignored:
+            mr.get_blockhashes())
+        d.addCallback(lambda blockhashes:
+            self.failUnlessEqual(self.block_hash_tree, blockhashes))
+
+        d.addCallback(lambda ignored:
+            mr.get_sharehashes())
+        d.addCallback(lambda sharehashes:
+            self.failUnlessEqual(self.share_hash_chain, sharehashes))
+
+        d.addCallback(lambda ignored:
+            mr.get_signature())
+        d.addCallback(lambda signature:
+            self.failUnlessEqual(signature, self.signature))
+
+        d.addCallback(lambda ignored:
+            mr.get_verification_key())
+        d.addCallback(lambda verification_key:
+            self.failUnlessEqual(verification_key, self.verification_key))
+
+        d.addCallback(lambda ignored:
+            mr.get_seqnum())
+        d.addCallback(lambda seqnum:
+            self.failUnlessEqual(seqnum, 0))
+
+        d.addCallback(lambda ignored:
+            mr.get_root_hash())
+        d.addCallback(lambda root_hash:
+            self.failUnlessEqual(self.root_hash, root_hash))
+
+        d.addCallback(lambda ignored:
+            mr.get_encoding_parameters())
+        def _check_encoding_parameters((k, n, segsize, datalen)):
+            self.failUnlessEqual(k, 3)
+            self.failUnlessEqual(n, 10)
+            self.failUnlessEqual(segsize, 6)
+            self.failUnlessEqual(datalen, 36)
+        d.addCallback(_check_encoding_parameters)
+
+        d.addCallback(lambda ignored:
+            mr.get_checkstring())
+        d.addCallback(lambda checkstring:
+            self.failUnlessEqual(checkstring, mw.get_checkstring()))
+        return d
+
+
+    def test_is_sdmf(self):
+        # The MDMFSlotReadProxy should also know how to read SDMF files,
+        # since it will encounter them on the grid. Callers use the
+        # is_sdmf method to test this.
+        self.write_sdmf_share_to_server("si1")
+        mr = MDMFSlotReadProxy(self.rref, "si1", 0)
+        d = mr.is_sdmf()
+        d.addCallback(lambda issdmf:
+            self.failUnless(issdmf))
+        return d
+
+
+    def test_reads_sdmf(self):
+        # The slot read proxy should, naturally, know how to tell us
+        # about data in the SDMF format
+        self.write_sdmf_share_to_server("si1")
+        mr = MDMFSlotReadProxy(self.rref, "si1", 0)
+        d = defer.succeed(None)
+        d.addCallback(lambda ignored:
+            mr.is_sdmf())
+        d.addCallback(lambda issdmf:
+            self.failUnless(issdmf))
+
+        # What do we need to read?
+        #  - The sharedata
+        #  - The salt
+        d.addCallback(lambda ignored:
+            mr.get_block_and_salt(0))
+        def _check_block_and_salt(results):
+            block, salt = results
+            # Our original file is 36 bytes long. Then each share is 12
+            # bytes in size. The share is composed entirely of the
+            # letter a. self.block contains 2 as, so 6 * self.block is
+            # what we are looking for.
+            self.failUnlessEqual(block, self.block * 6)
+            self.failUnlessEqual(salt, self.salt)
+        d.addCallback(_check_block_and_salt)
+
+        #  - The blockhashes
+        d.addCallback(lambda ignored:
+            mr.get_blockhashes())
+        d.addCallback(lambda blockhashes:
+            self.failUnlessEqual(self.block_hash_tree,
+                                 blockhashes,
+                                 blockhashes))
+        #  - The sharehashes
+        d.addCallback(lambda ignored:
+            mr.get_sharehashes())
+        d.addCallback(lambda sharehashes:
+            self.failUnlessEqual(self.share_hash_chain,
+                                 sharehashes))
+        #  - The keys
+        d.addCallback(lambda ignored:
+            mr.get_encprivkey())
+        d.addCallback(lambda encprivkey:
+            self.failUnlessEqual(encprivkey, self.encprivkey, encprivkey))
+        d.addCallback(lambda ignored:
+            mr.get_verification_key())
+        d.addCallback(lambda verification_key:
+            self.failUnlessEqual(verification_key,
+                                 self.verification_key,
+                                 verification_key))
+        #  - The signature
+        d.addCallback(lambda ignored:
+            mr.get_signature())
+        d.addCallback(lambda signature:
+            self.failUnlessEqual(signature, self.signature, signature))
+
+        #  - The sequence number
+        d.addCallback(lambda ignored:
+            mr.get_seqnum())
+        d.addCallback(lambda seqnum:
+            self.failUnlessEqual(seqnum, 0, seqnum))
+
+        #  - The root hash
+        d.addCallback(lambda ignored:
+            mr.get_root_hash())
+        d.addCallback(lambda root_hash:
+            self.failUnlessEqual(root_hash, self.root_hash, root_hash))
+        return d
+
+
+    def test_only_reads_one_segment_sdmf(self):
+        # SDMF shares have only one segment, so it doesn't make sense to
+        # read more segments than that. The reader should know this and
+        # complain if we try to do that.
+        self.write_sdmf_share_to_server("si1")
+        mr = MDMFSlotReadProxy(self.rref, "si1", 0)
+        d = defer.succeed(None)
+        d.addCallback(lambda ignored:
+            mr.is_sdmf())
+        d.addCallback(lambda issdmf:
+            self.failUnless(issdmf))
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "test bad segment",
+                            None,
+                            mr.get_block_and_salt, 1))
+        return d
+
+
+    def test_read_with_prefetched_mdmf_data(self):
+        # The MDMFSlotReadProxy will prefill certain fields if you pass
+        # it data that you have already fetched. This is useful for
+        # cases like the Servermap, which prefetches ~2kb of data while
+        # finding out which shares are on the remote peer so that it
+        # doesn't waste round trips.
+        mdmf_data = self.build_test_mdmf_share()
+        self.write_test_share_to_server("si1")
+        def _make_mr(ignored, length):
+            mr = MDMFSlotReadProxy(self.rref, "si1", 0, mdmf_data[:length])
+            return mr
+
+        d = defer.succeed(None)
+        # This should be enough to fill in both the encoding parameters
+        # and the table of offsets, which will complete the version
+        # information tuple.
+        d.addCallback(_make_mr, 107)
+        d.addCallback(lambda mr:
+            mr.get_verinfo())
+        def _check_verinfo(verinfo):
+            self.failUnless(verinfo)
+            self.failUnlessEqual(len(verinfo), 9)
+            (seqnum,
+             root_hash,
+             salt_hash,
+             segsize,
+             datalen,
+             k,
+             n,
+             prefix,
+             offsets) = verinfo
+            self.failUnlessEqual(seqnum, 0)
+            self.failUnlessEqual(root_hash, self.root_hash)
+            self.failUnlessEqual(segsize, 6)
+            self.failUnlessEqual(datalen, 36)
+            self.failUnlessEqual(k, 3)
+            self.failUnlessEqual(n, 10)
+            expected_prefix = struct.pack(MDMFSIGNABLEHEADER,
+                                          1,
+                                          seqnum,
+                                          root_hash,
+                                          k,
+                                          n,
+                                          segsize,
+                                          datalen)
+            self.failUnlessEqual(expected_prefix, prefix)
+            self.failUnlessEqual(self.rref.read_count, 0)
+        d.addCallback(_check_verinfo)
+        # This is not enough data to read a block and a share, so the
+        # wrapper should attempt to read this from the remote server.
+        d.addCallback(_make_mr, 107)
+        d.addCallback(lambda mr:
+            mr.get_block_and_salt(0))
+        def _check_block_and_salt((block, salt)):
+            self.failUnlessEqual(block, self.block)
+            self.failUnlessEqual(salt, self.salt)
+            self.failUnlessEqual(self.rref.read_count, 1)
+        # This should be enough data to read one block.
+        d.addCallback(_make_mr, 249)
+        d.addCallback(lambda mr:
+            mr.get_block_and_salt(0))
+        d.addCallback(_check_block_and_salt)
+        return d
+
+
+    def test_read_with_prefetched_sdmf_data(self):
+        sdmf_data = self.build_test_sdmf_share()
+        self.write_sdmf_share_to_server("si1")
+        def _make_mr(ignored, length):
+            mr = MDMFSlotReadProxy(self.rref, "si1", 0, sdmf_data[:length])
+            return mr
+
+        d = defer.succeed(None)
+        # This should be enough to get us the encoding parameters,
+        # offset table, and everything else we need to build a verinfo
+        # string.
+        d.addCallback(_make_mr, 107)
+        d.addCallback(lambda mr:
+            mr.get_verinfo())
+        def _check_verinfo(verinfo):
+            self.failUnless(verinfo)
+            self.failUnlessEqual(len(verinfo), 9)
+            (seqnum,
+             root_hash,
+             salt,
+             segsize,
+             datalen,
+             k,
+             n,
+             prefix,
+             offsets) = verinfo
+            self.failUnlessEqual(seqnum, 0)
+            self.failUnlessEqual(root_hash, self.root_hash)
+            self.failUnlessEqual(salt, self.salt)
+            self.failUnlessEqual(segsize, 36)
+            self.failUnlessEqual(datalen, 36)
+            self.failUnlessEqual(k, 3)
+            self.failUnlessEqual(n, 10)
+            expected_prefix = struct.pack(SIGNED_PREFIX,
+                                          0,
+                                          seqnum,
+                                          root_hash,
+                                          salt,
+                                          k,
+                                          n,
+                                          segsize,
+                                          datalen)
+            self.failUnlessEqual(expected_prefix, prefix)
+            self.failUnlessEqual(self.rref.read_count, 0)
+        d.addCallback(_check_verinfo)
+        # This shouldn't be enough to read any share data.
+        d.addCallback(_make_mr, 107)
+        d.addCallback(lambda mr:
+            mr.get_block_and_salt(0))
+        def _check_block_and_salt((block, salt)):
+            self.failUnlessEqual(block, self.block * 6)
+            self.failUnlessEqual(salt, self.salt)
+            # TODO: Fix the read routine so that it reads only the data
+            #       that it has cached if it can't read all of it.
+            self.failUnlessEqual(self.rref.read_count, 2)
+
+        # This should be enough to read share data.
+        d.addCallback(_make_mr, self.offsets['share_data'])
+        d.addCallback(lambda mr:
+            mr.get_block_and_salt(0))
+        d.addCallback(_check_block_and_salt)
+        return d
+
+
+    def test_read_with_empty_mdmf_file(self):
+        # Some tests upload a file with no contents to test things
+        # unrelated to the actual handling of the content of the file.
+        # The reader should behave intelligently in these cases.
+        self.write_test_share_to_server("si1", empty=True)
+        mr = MDMFSlotReadProxy(self.rref, "si1", 0)
+        # We should be able to get the encoding parameters, and they
+        # should be correct.
+        d = defer.succeed(None)
+        d.addCallback(lambda ignored:
+            mr.get_encoding_parameters())
+        def _check_encoding_parameters(params):
+            self.failUnlessEqual(len(params), 4)
+            k, n, segsize, datalen = params
+            self.failUnlessEqual(k, 3)
+            self.failUnlessEqual(n, 10)
+            self.failUnlessEqual(segsize, 0)
+            self.failUnlessEqual(datalen, 0)
+        d.addCallback(_check_encoding_parameters)
+
+        # We should not be able to fetch a block, since there are no
+        # blocks to fetch
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "get block on empty file",
+                            None,
+                            mr.get_block_and_salt, 0))
+        return d
+
+
+    def test_read_with_empty_sdmf_file(self):
+        self.write_sdmf_share_to_server("si1", empty=True)
+        mr = MDMFSlotReadProxy(self.rref, "si1", 0)
+        # We should be able to get the encoding parameters, and they
+        # should be correct
+        d = defer.succeed(None)
+        d.addCallback(lambda ignored:
+            mr.get_encoding_parameters())
+        def _check_encoding_parameters(params):
+            self.failUnlessEqual(len(params), 4)
+            k, n, segsize, datalen = params
+            self.failUnlessEqual(k, 3)
+            self.failUnlessEqual(n, 10)
+            self.failUnlessEqual(segsize, 0)
+            self.failUnlessEqual(datalen, 0)
+        d.addCallback(_check_encoding_parameters)
+
+        # It does not make sense to get a block in this format, so we
+        # should not be able to.
+        d.addCallback(lambda ignored:
+            self.shouldFail(LayoutInvalid, "get block on an empty file",
+                            None,
+                            mr.get_block_and_salt, 0))
+        return d
+
+
+    def test_verinfo_with_sdmf_file(self):
+        self.write_sdmf_share_to_server("si1")
+        mr = MDMFSlotReadProxy(self.rref, "si1", 0)
+        # We should be able to get the version information.
+        d = defer.succeed(None)
+        d.addCallback(lambda ignored:
+            mr.get_verinfo())
+        def _check_verinfo(verinfo):
+            self.failUnless(verinfo)
+            self.failUnlessEqual(len(verinfo), 9)
+            (seqnum,
+             root_hash,
+             salt,
+             segsize,
+             datalen,
+             k,
+             n,
+             prefix,
+             offsets) = verinfo
+            self.failUnlessEqual(seqnum, 0)
+            self.failUnlessEqual(root_hash, self.root_hash)
+            self.failUnlessEqual(salt, self.salt)
+            self.failUnlessEqual(segsize, 36)
+            self.failUnlessEqual(datalen, 36)
+            self.failUnlessEqual(k, 3)
+            self.failUnlessEqual(n, 10)
+            expected_prefix = struct.pack(">BQ32s16s BBQQ",
+                                          0,
+                                          seqnum,
+                                          root_hash,
+                                          salt,
+                                          k,
+                                          n,
+                                          segsize,
+                                          datalen)
+            self.failUnlessEqual(prefix, expected_prefix)
+            self.failUnlessEqual(offsets, self.offsets)
+        d.addCallback(_check_verinfo)
+        return d
+
+
+    def test_verinfo_with_mdmf_file(self):
+        self.write_test_share_to_server("si1")
+        mr = MDMFSlotReadProxy(self.rref, "si1", 0)
+        d = defer.succeed(None)
+        d.addCallback(lambda ignored:
+            mr.get_verinfo())
+        def _check_verinfo(verinfo):
+            self.failUnless(verinfo)
+            self.failUnlessEqual(len(verinfo), 9)
+            (seqnum,
+             root_hash,
+             IV,
+             segsize,
+             datalen,
+             k,
+             n,
+             prefix,
+             offsets) = verinfo
+            self.failUnlessEqual(seqnum, 0)
+            self.failUnlessEqual(root_hash, self.root_hash)
+            self.failIf(IV)
+            self.failUnlessEqual(segsize, 6)
+            self.failUnlessEqual(datalen, 36)
+            self.failUnlessEqual(k, 3)
+            self.failUnlessEqual(n, 10)
+            expected_prefix = struct.pack(">BQ32s BBQQ",
+                                          1,
+                                          seqnum,
+                                          root_hash,
+                                          k,
+                                          n,
+                                          segsize,
+                                          datalen)
+            self.failUnlessEqual(prefix, expected_prefix)
+            self.failUnlessEqual(offsets, self.offsets)
+        d.addCallback(_check_verinfo)
+        return d
+
+
+    def test_reader_queue(self):
+        self.write_test_share_to_server('si1')
+        mr = MDMFSlotReadProxy(self.rref, "si1", 0)
+        d1 = mr.get_block_and_salt(0, queue=True)
+        d2 = mr.get_blockhashes(queue=True)
+        d3 = mr.get_sharehashes(queue=True)
+        d4 = mr.get_signature(queue=True)
+        d5 = mr.get_verification_key(queue=True)
+        dl = defer.DeferredList([d1, d2, d3, d4, d5])
+        mr.flush()
+        def _print(results):
+            self.failUnlessEqual(len(results), 5)
+            # We have one read for version information and offsets, and
+            # one for everything else.
+            self.failUnlessEqual(self.rref.read_count, 2)
+            block, salt = results[0][1] # results[0] is a boolean that says
+                                           # whether or not the operation
+                                           # worked.
+            self.failUnlessEqual(self.block, block)
+            self.failUnlessEqual(self.salt, salt)
+
+            blockhashes = results[1][1]
+            self.failUnlessEqual(self.block_hash_tree, blockhashes)
+
+            sharehashes = results[2][1]
+            self.failUnlessEqual(self.share_hash_chain, sharehashes)
+
+            signature = results[3][1]
+            self.failUnlessEqual(self.signature, signature)
+
+            verification_key = results[4][1]
+            self.failUnlessEqual(self.verification_key, verification_key)
+        dl.addCallback(_print)
+        return dl
+
+
 class Stats(unittest.TestCase):
 
     def setUp(self):
}
[Alter the ServermapUpdater to find MDMF files
Kevan Carstensen <kevan@isnotajoke.com>**20100626234118
 Ignore-this: 25f6278209c2983ba8f307cfe0fde0
 
 The servermapupdater should find MDMF files on a grid in the same way
 that it finds SDMF files. This patch makes it do that.
] {
hunk ./src/allmydata/mutable/servermap.py 7
 from itertools import count
 from twisted.internet import defer
 from twisted.python import failure
-from foolscap.api import DeadReferenceError, RemoteException, eventually
+from foolscap.api import DeadReferenceError, RemoteException, eventually, \
+                         fireEventually
 from allmydata.util import base32, hashutil, idlib, log
 from allmydata.storage.server import si_b2a
 from allmydata.interfaces import IServermapUpdaterStatus
hunk ./src/allmydata/mutable/servermap.py 17
 from allmydata.mutable.common import MODE_CHECK, MODE_ANYTHING, MODE_WRITE, MODE_READ, \
      DictOfSets, CorruptShareError, NeedMoreDataError
 from allmydata.mutable.layout import unpack_prefix_and_signature, unpack_header, unpack_share, \
-     SIGNED_PREFIX_LENGTH
+     SIGNED_PREFIX_LENGTH, MDMFSlotReadProxy
 
 class UpdateStatus:
     implements(IServermapUpdaterStatus)
hunk ./src/allmydata/mutable/servermap.py 254
         """Return a set of versionids, one for each version that is currently
         recoverable."""
         versionmap = self.make_versionmap()
-
         recoverable_versions = set()
         for (verinfo, shares) in versionmap.items():
             (seqnum, root_hash, IV, segsize, datalength, k, N, prefix,
hunk ./src/allmydata/mutable/servermap.py 366
         self._servers_responded = set()
 
         # how much data should we read?
+        # SDMF:
         #  * if we only need the checkstring, then [0:75]
         #  * if we need to validate the checkstring sig, then [543ish:799ish]
         #  * if we need the verification key, then [107:436ish]
hunk ./src/allmydata/mutable/servermap.py 374
         #  * if we need the encrypted private key, we want [-1216ish:]
         #   * but we can't read from negative offsets
         #   * the offset table tells us the 'ish', also the positive offset
-        # A future version of the SMDF slot format should consider using
-        # fixed-size slots so we can retrieve less data. For now, we'll just
-        # read 2000 bytes, which also happens to read enough actual data to
-        # pre-fetch a 9-entry dirnode.
+        # MDMF:
+        #  * Checkstring? [0:72]
+        #  * If we want to validate the checkstring, then [0:72], [143:?] --
+        #    the offset table will tell us for sure.
+        #  * If we need the verification key, we have to consult the offset
+        #    table as well.
+        # At this point, we don't know which we are. Our filenode can
+        # tell us, but it might be lying -- in some cases, we're
+        # responsible for telling it which kind of file it is.
         self._read_size = 4000
         if mode == MODE_CHECK:
             # we use unpack_prefix_and_signature, so we need 1k
hunk ./src/allmydata/mutable/servermap.py 432
         self._queries_completed = 0
 
         sb = self._storage_broker
+        # All of the peers, permuted by the storage index, as usual.
         full_peerlist = sb.get_servers_for_index(self._storage_index)
         self.full_peerlist = full_peerlist # for use later, immutable
         self.extra_peers = full_peerlist[:] # peers are removed as we use them
hunk ./src/allmydata/mutable/servermap.py 439
         self._good_peers = set() # peers who had some shares
         self._empty_peers = set() # peers who don't have any shares
         self._bad_peers = set() # peers to whom our queries failed
+        self._readers = {} # peerid -> dict(sharewriters), filled in
+                           # after responses come in.
 
         k = self._node.get_required_shares()
hunk ./src/allmydata/mutable/servermap.py 443
+        # For what cases can these conditions work?
         if k is None:
             # make a guess
             k = 3
hunk ./src/allmydata/mutable/servermap.py 456
         self.num_peers_to_query = k + self.EPSILON
 
         if self.mode == MODE_CHECK:
+            # We want to query all of the peers.
             initial_peers_to_query = dict(full_peerlist)
             must_query = set(initial_peers_to_query.keys())
             self.extra_peers = []
hunk ./src/allmydata/mutable/servermap.py 464
             # we're planning to replace all the shares, so we want a good
             # chance of finding them all. We will keep searching until we've
             # seen epsilon that don't have a share.
+            # We don't query all of the peers because that could take a while.
             self.num_peers_to_query = N + self.EPSILON
             initial_peers_to_query, must_query = self._build_initial_querylist()
             self.required_num_empty_peers = self.EPSILON
hunk ./src/allmydata/mutable/servermap.py 474
             # might also avoid the round trip required to read the encrypted
             # private key.
 
-        else:
+        else: # MODE_READ, MODE_ANYTHING
+            # 2k peers is good enough.
             initial_peers_to_query, must_query = self._build_initial_querylist()
 
         # this is a set of peers that we are required to get responses from:
hunk ./src/allmydata/mutable/servermap.py 490
         # before we can consider ourselves finished, and self.extra_peers
         # contains the overflow (peers that we should tap if we don't get
         # enough responses)
+        # I guess that self._must_query is a subset of
+        # initial_peers_to_query?
+        assert set(must_query).issubset(set(initial_peers_to_query))
 
         self._send_initial_requests(initial_peers_to_query)
         self._status.timings["initial_queries"] = time.time() - self._started
hunk ./src/allmydata/mutable/servermap.py 549
         # errors that aren't handled by _query_failed (and errors caused by
         # _query_failed) get logged, but we still want to check for doneness.
         d.addErrback(log.err)
-        d.addBoth(self._check_for_done)
         d.addErrback(self._fatal_error)
hunk ./src/allmydata/mutable/servermap.py 550
+        d.addCallback(self._check_for_done)
         return d
 
     def _do_read(self, ss, peerid, storage_index, shnums, readv):
hunk ./src/allmydata/mutable/servermap.py 569
         d = ss.callRemote("slot_readv", storage_index, shnums, readv)
         return d
 
+
+    def _got_corrupt_share(self, e, shnum, peerid, data, lp):
+        """
+        I am called when a remote server returns a corrupt share in
+        response to one of our queries. By corrupt, I mean a share
+        without a valid signature. I then record the failure, notify the
+        server of the corruption, and record the share as bad.
+        """
+        f = failure.Failure(e)
+        self.log(format="bad share: %(f_value)s", f_value=str(f),
+                 failure=f, parent=lp, level=log.WEIRD, umid="h5llHg")
+        # Notify the server that its share is corrupt.
+        self.notify_server_corruption(peerid, shnum, str(e))
+        # By flagging this as a bad peer, we won't count any of
+        # the other shares on that peer as valid, though if we
+        # happen to find a valid version string amongst those
+        # shares, we'll keep track of it so that we don't need
+        # to validate the signature on those again.
+        self._bad_peers.add(peerid)
+        self._last_failure = f
+        # XXX: Use the reader for this?
+        checkstring = data[:SIGNED_PREFIX_LENGTH]
+        self._servermap.mark_bad_share(peerid, shnum, checkstring)
+        self._servermap.problems.append(f)
+
+
+    def _cache_good_sharedata(self, verinfo, shnum, now, data):
+        """
+        If one of my queries returns successfully (which means that we
+        were able to and successfully did validate the signature), I
+        cache the data that we initially fetched from the storage
+        server. This will help reduce the number of roundtrips that need
+        to occur when the file is downloaded, or when the file is
+        updated.
+        """
+        self._node._add_to_cache(verinfo, shnum, 0, data, now)
+
+
     def _got_results(self, datavs, peerid, readsize, stuff, started):
         lp = self.log(format="got result from [%(peerid)s], %(numshares)d shares",
                       peerid=idlib.shortnodeid_b2a(peerid),
hunk ./src/allmydata/mutable/servermap.py 630
         else:
             self._empty_peers.add(peerid)
 
-        last_verinfo = None
-        last_shnum = None
+        ss, storage_index = stuff
+        ds = []
+
         for shnum,datav in datavs.items():
             data = datav[0]
hunk ./src/allmydata/mutable/servermap.py 635
-            try:
-                verinfo = self._got_results_one_share(shnum, data, peerid, lp)
-                last_verinfo = verinfo
-                last_shnum = shnum
-                self._node._add_to_cache(verinfo, shnum, 0, data, now)
-            except CorruptShareError, e:
-                # log it and give the other shares a chance to be processed
-                f = failure.Failure()
-                self.log(format="bad share: %(f_value)s", f_value=str(f.value),
-                         failure=f, parent=lp, level=log.WEIRD, umid="h5llHg")
-                self.notify_server_corruption(peerid, shnum, str(e))
-                self._bad_peers.add(peerid)
-                self._last_failure = f
-                checkstring = data[:SIGNED_PREFIX_LENGTH]
-                self._servermap.mark_bad_share(peerid, shnum, checkstring)
-                self._servermap.problems.append(f)
-                pass
-
-        self._status.timings["cumulative_verify"] += (time.time() - now)
+            reader = MDMFSlotReadProxy(ss,
+                                       storage_index,
+                                       shnum,
+                                       data)
+            self._readers.setdefault(peerid, dict())[shnum] = reader
+            # our goal, with each response, is to validate the version
+            # information and share data as best we can at this point --
+            # we do this by validating the signature. To do this, we
+            # need to do the following:
+            #   - If we don't already have the public key, fetch the
+            #     public key. We use this to validate the signature.
+            if not self._node.get_pubkey():
+                # fetch and set the public key.
+                d = reader.get_verification_key()
+                d.addCallback(lambda results, shnum=shnum, peerid=peerid:
+                    self._try_to_set_pubkey(results, peerid, shnum, lp))
+                # XXX: Make self._pubkey_query_failed?
+                d.addErrback(lambda error, shnum=shnum, peerid=peerid:
+                    self._got_corrupt_share(error, shnum, peerid, data, lp))
+            else:
+                # we already have the public key.
+                d = defer.succeed(None)
+            # Neither of these two branches return anything of
+            # consequence, so the first entry in our deferredlist will
+            # be None.
 
hunk ./src/allmydata/mutable/servermap.py 661
-        if self._need_privkey and last_verinfo:
-            # send them a request for the privkey. We send one request per
-            # server.
-            lp2 = self.log("sending privkey request",
-                           parent=lp, level=log.NOISY)
-            (seqnum, root_hash, IV, segsize, datalength, k, N, prefix,
-             offsets_tuple) = last_verinfo
-            o = dict(offsets_tuple)
+            # - Next, we need the version information. We almost
+            #   certainly got this by reading the first thousand or so
+            #   bytes of the share on the storage server, so we
+            #   shouldn't need to fetch anything at this step.
+            d2 = reader.get_verinfo()
+            d2.addErrback(lambda error, shnum=shnum, peerid=peerid:
+                self._got_corrupt_share(error, shnum, peerid, data, lp))
+            # - Next, we need the signature. For an SDMF share, it is
+            #   likely that we fetched this when doing our initial fetch
+            #   to get the version information. In MDMF, this lives at
+            #   the end of the share, so unless the file is quite small,
+            #   we'll need to do a remote fetch to get it.
+            d3 = reader.get_signature()
+            d3.addErrback(lambda error, shnum=shnum, peerid=peerid:
+                self._got_corrupt_share(error, shnum, peerid, data, lp))
+            #  Once we have all three of these responses, we can move on
+            #  to validating the signature
 
hunk ./src/allmydata/mutable/servermap.py 679
-            self._queries_outstanding.add(peerid)
-            readv = [ (o['enc_privkey'], (o['EOF'] - o['enc_privkey'])) ]
-            ss = self._servermap.connections[peerid]
-            privkey_started = time.time()
-            d = self._do_read(ss, peerid, self._storage_index,
-                              [last_shnum], readv)
-            d.addCallback(self._got_privkey_results, peerid, last_shnum,
-                          privkey_started, lp2)
-            d.addErrback(self._privkey_query_failed, peerid, last_shnum, lp2)
-            d.addErrback(log.err)
-            d.addCallback(self._check_for_done)
-            d.addErrback(self._fatal_error)
+            # Does the node already have a privkey? If not, we'll try to
+            # fetch it here.
+            if self._need_privkey:
+                d4 = reader.get_encprivkey()
+                d4.addCallback(lambda results, shnum=shnum, peerid=peerid:
+                    self._try_to_validate_privkey(results, peerid, shnum, lp))
+                d4.addErrback(lambda error, shnum=shnum, peerid=peerid:
+                    self._privkey_query_failed(error, shnum, data, lp))
+            else:
+                d4 = defer.succeed(None)
 
hunk ./src/allmydata/mutable/servermap.py 690
+            dl = defer.DeferredList([d, d2, d3, d4])
+            dl.addCallback(lambda results, shnum=shnum, peerid=peerid:
+                self._got_signature_one_share(results, shnum, peerid, lp))
+            dl.addErrback(lambda error, shnum=shnum, data=data:
+               self._got_corrupt_share(error, shnum, peerid, data, lp))
+            dl.addCallback(lambda verinfo, shnum=shnum, peerid=peerid, data=data:
+                self._cache_good_sharedata(verinfo, shnum, now, data))
+            ds.append(dl)
+        # dl is a deferred list that will fire when all of the shares
+        # that we found on this peer are done processing. When dl fires,
+        # we know that processing is done, so we can decrement the
+        # semaphore-like thing that we incremented earlier.
+        dl = defer.DeferredList(ds, fireOnOneErrback=True)
+        # Are we done? Done means that there are no more queries to
+        # send, that there are no outstanding queries, and that we
+        # haven't received any queries that are still processing. If we
+        # are done, self._check_for_done will cause the done deferred
+        # that we returned to our caller to fire, which tells them that
+        # they have a complete servermap, and that we won't be touching
+        # the servermap anymore.
+        dl.addCallback(self._check_for_done)
+        dl.addErrback(self._fatal_error)
         # all done!
         self.log("_got_results done", parent=lp, level=log.NOISY)
hunk ./src/allmydata/mutable/servermap.py 714
+        return dl
+
+
+    def _try_to_set_pubkey(self, pubkey_s, peerid, shnum, lp):
+        if self._node.get_pubkey():
+            return # don't go through this again if we don't have to
+        fingerprint = hashutil.ssk_pubkey_fingerprint_hash(pubkey_s)
+        assert len(fingerprint) == 32
+        if fingerprint != self._node.get_fingerprint():
+            raise CorruptShareError(peerid, shnum,
+                                "pubkey doesn't match fingerprint")
+        self._node._populate_pubkey(self._deserialize_pubkey(pubkey_s))
+        assert self._node.get_pubkey()
+
 
     def notify_server_corruption(self, peerid, shnum, reason):
         ss = self._servermap.connections[peerid]
hunk ./src/allmydata/mutable/servermap.py 734
         ss.callRemoteOnly("advise_corrupt_share",
                           "mutable", self._storage_index, shnum, reason)
 
-    def _got_results_one_share(self, shnum, data, peerid, lp):
+
+    def _got_signature_one_share(self, results, shnum, peerid, lp):
+        # It is our job to give versioninfo to our caller. We need to
+        # raise CorruptShareError if the share is corrupt for any
+        # reason, something that our caller will handle.
         self.log(format="_got_results: got shnum #%(shnum)d from peerid %(peerid)s",
                  shnum=shnum,
                  peerid=idlib.shortnodeid_b2a(peerid),
hunk ./src/allmydata/mutable/servermap.py 744
                  level=log.NOISY,
                  parent=lp)
-
-        # this might raise NeedMoreDataError, if the pubkey and signature
-        # live at some weird offset. That shouldn't happen, so I'm going to
-        # treat it as a bad share.
-        (seqnum, root_hash, IV, k, N, segsize, datalength,
-         pubkey_s, signature, prefix) = unpack_prefix_and_signature(data)
-
-        if not self._node.get_pubkey():
-            fingerprint = hashutil.ssk_pubkey_fingerprint_hash(pubkey_s)
-            assert len(fingerprint) == 32
-            if fingerprint != self._node.get_fingerprint():
-                raise CorruptShareError(peerid, shnum,
-                                        "pubkey doesn't match fingerprint")
-            self._node._populate_pubkey(self._deserialize_pubkey(pubkey_s))
-
-        if self._need_privkey:
-            self._try_to_extract_privkey(data, peerid, shnum, lp)
-
-        (ig_version, ig_seqnum, ig_root_hash, ig_IV, ig_k, ig_N,
-         ig_segsize, ig_datalen, offsets) = unpack_header(data)
+        _, verinfo, signature, __ = results
+        (seqnum,
+         root_hash,
+         saltish,
+         segsize,
+         datalen,
+         k,
+         n,
+         prefix,
+         offsets) = verinfo[1]
         offsets_tuple = tuple( [(key,value) for key,value in offsets.items()] )
 
hunk ./src/allmydata/mutable/servermap.py 756
-        verinfo = (seqnum, root_hash, IV, segsize, datalength, k, N, prefix,
+        # XXX: This should be done for us in the method, so
+        # presumably you can go in there and fix it.
+        verinfo = (seqnum,
+                   root_hash,
+                   saltish,
+                   segsize,
+                   datalen,
+                   k,
+                   n,
+                   prefix,
                    offsets_tuple)
hunk ./src/allmydata/mutable/servermap.py 767
+        # This tuple uniquely identifies a share on the grid; we use it
+        # to keep track of the ones that we've already seen.
 
         if verinfo not in self._valid_versions:
hunk ./src/allmydata/mutable/servermap.py 771
-            # it's a new pair. Verify the signature.
-            valid = self._node.get_pubkey().verify(prefix, signature)
+            # This is a new version tuple, and we need to validate it
+            # against the public key before keeping track of it.
+            assert self._node.get_pubkey()
+            valid = self._node.get_pubkey().verify(prefix, signature[1])
             if not valid:
hunk ./src/allmydata/mutable/servermap.py 776
-                raise CorruptShareError(peerid, shnum, "signature is invalid")
+                raise CorruptShareError(peerid, shnum,
+                                        "signature is invalid")
 
hunk ./src/allmydata/mutable/servermap.py 779
-            # ok, it's a valid verinfo. Add it to the list of validated
-            # versions.
-            self.log(" found valid version %d-%s from %s-sh%d: %d-%d/%d/%d"
-                     % (seqnum, base32.b2a(root_hash)[:4],
-                        idlib.shortnodeid_b2a(peerid), shnum,
-                        k, N, segsize, datalength),
-                     parent=lp)
-            self._valid_versions.add(verinfo)
-        # We now know that this is a valid candidate verinfo.
+        # ok, it's a valid verinfo. Add it to the list of validated
+        # versions.
+        self.log(" found valid version %d-%s from %s-sh%d: %d-%d/%d/%d"
+                 % (seqnum, base32.b2a(root_hash)[:4],
+                    idlib.shortnodeid_b2a(peerid), shnum,
+                    k, n, segsize, datalen),
+                    parent=lp)
+        self._valid_versions.add(verinfo)
+        # We now know that this is a valid candidate verinfo. Whether or
+        # not this instance of it is valid is a matter for the next
+        # statement; at this point, we just know that if we see this
+        # version info again, that its signature checks out and that
+        # we're okay to skip the signature-checking step.
 
hunk ./src/allmydata/mutable/servermap.py 793
+        # (peerid, shnum) are bound in the method invocation.
         if (peerid, shnum) in self._servermap.bad_shares:
             # we've been told that the rest of the data in this share is
             # unusable, so don't add it to the servermap.
hunk ./src/allmydata/mutable/servermap.py 808
         self.versionmap.add(verinfo, (shnum, peerid, timestamp))
         return verinfo
 
+
     def _deserialize_pubkey(self, pubkey_s):
         verifier = rsa.create_verifying_key_from_string(pubkey_s)
         return verifier
hunk ./src/allmydata/mutable/servermap.py 813
 
-    def _try_to_extract_privkey(self, data, peerid, shnum, lp):
-        try:
-            r = unpack_share(data)
-        except NeedMoreDataError, e:
-            # this share won't help us. oh well.
-            offset = e.encprivkey_offset
-            length = e.encprivkey_length
-            self.log("shnum %d on peerid %s: share was too short (%dB) "
-                     "to get the encprivkey; [%d:%d] ought to hold it" %
-                     (shnum, idlib.shortnodeid_b2a(peerid), len(data),
-                      offset, offset+length),
-                     parent=lp)
-            # NOTE: if uncoordinated writes are taking place, someone might
-            # change the share (and most probably move the encprivkey) before
-            # we get a chance to do one of these reads and fetch it. This
-            # will cause us to see a NotEnoughSharesError(unable to fetch
-            # privkey) instead of an UncoordinatedWriteError . This is a
-            # nuisance, but it will go away when we move to DSA-based mutable
-            # files (since the privkey will be small enough to fit in the
-            # write cap).
-
-            return
-
-        (seqnum, root_hash, IV, k, N, segsize, datalen,
-         pubkey, signature, share_hash_chain, block_hash_tree,
-         share_data, enc_privkey) = r
-
-        return self._try_to_validate_privkey(enc_privkey, peerid, shnum, lp)
 
     def _try_to_validate_privkey(self, enc_privkey, peerid, shnum, lp):
hunk ./src/allmydata/mutable/servermap.py 815
-
+        """
+        Given a writekey from a remote server, I validate it against the
+        writekey stored in my node. If it is valid, then I set the
+        privkey and encprivkey properties of the node.
+        """
         alleged_privkey_s = self._node._decrypt_privkey(enc_privkey)
         alleged_writekey = hashutil.ssk_writekey_hash(alleged_privkey_s)
         if alleged_writekey != self._node.get_writekey():
hunk ./src/allmydata/mutable/servermap.py 892
         self._queries_completed += 1
         self._last_failure = f
 
-    def _got_privkey_results(self, datavs, peerid, shnum, started, lp):
-        now = time.time()
-        elapsed = now - started
-        self._status.add_per_server_time(peerid, "privkey", started, elapsed)
-        self._queries_outstanding.discard(peerid)
-        if not self._need_privkey:
-            return
-        if shnum not in datavs:
-            self.log("privkey wasn't there when we asked it",
-                     level=log.WEIRD, umid="VA9uDQ")
-            return
-        datav = datavs[shnum]
-        enc_privkey = datav[0]
-        self._try_to_validate_privkey(enc_privkey, peerid, shnum, lp)
 
     def _privkey_query_failed(self, f, peerid, shnum, lp):
         self._queries_outstanding.discard(peerid)
hunk ./src/allmydata/mutable/servermap.py 906
         self._servermap.problems.append(f)
         self._last_failure = f
 
+
     def _check_for_done(self, res):
         # exit paths:
         #  return self._send_more_queries(outstanding) : send some more queries
hunk ./src/allmydata/mutable/servermap.py 912
         #  return self._done() : all done
         #  return : keep waiting, no new queries
-
         lp = self.log(format=("_check_for_done, mode is '%(mode)s', "
                               "%(outstanding)d queries outstanding, "
                               "%(extra)d extra peers available, "
hunk ./src/allmydata/mutable/servermap.py 1117
         self._servermap.last_update_time = self._started
         # the servermap will not be touched after this
         self.log("servermap: %s" % self._servermap.summarize_versions())
+
         eventually(self._done_deferred.callback, self._servermap)
 
     def _fatal_error(self, f):
hunk ./src/allmydata/test/test_mutable.py 637
         d.addCallback(_created)
         return d
 
-    def publish_multiple(self):
+    def publish_mdmf(self):
+        # like publish_one, except that the result is guaranteed to be
+        # an MDMF file.
+        # self.CONTENTS should have more than one segment.
+        self.CONTENTS = "This is an MDMF file" * 100000
+        self._storage = FakeStorage()
+        self._nodemaker = make_nodemaker(self._storage)
+        self._storage_broker = self._nodemaker.storage_broker
+        d = self._nodemaker.create_mutable_file(self.CONTENTS, version=1)
+        def _created(node):
+            self._fn = node
+            self._fn2 = self._nodemaker.create_from_cap(node.get_uri())
+        d.addCallback(_created)
+        return d
+
+
+    def publish_sdmf(self):
+        # like publish_one, except that the result is guaranteed to be
+        # an SDMF file
+        self.CONTENTS = "This is an SDMF file" * 1000
+        self._storage = FakeStorage()
+        self._nodemaker = make_nodemaker(self._storage)
+        self._storage_broker = self._nodemaker.storage_broker
+        d = self._nodemaker.create_mutable_file(self.CONTENTS, version=0)
+        def _created(node):
+            self._fn = node
+            self._fn2 = self._nodemaker.create_from_cap(node.get_uri())
+        d.addCallback(_created)
+        return d
+
+
+    def publish_multiple(self, version=0):
         self.CONTENTS = ["Contents 0",
                          "Contents 1",
                          "Contents 2",
hunk ./src/allmydata/test/test_mutable.py 677
         self._copied_shares = {}
         self._storage = FakeStorage()
         self._nodemaker = make_nodemaker(self._storage)
-        d = self._nodemaker.create_mutable_file(self.CONTENTS[0]) # seqnum=1
+        d = self._nodemaker.create_mutable_file(self.CONTENTS[0], version=version) # seqnum=1
         def _created(node):
             self._fn = node
             # now create multiple versions of the same file, and accumulate
hunk ./src/allmydata/test/test_mutable.py 906
         return d
 
 
+    def test_servermapupdater_finds_mdmf_files(self):
+        # setUp already published an MDMF file for us. We just need to
+        # make sure that when we run the ServermapUpdater, the file is
+        # reported to have one recoverable version.
+        d = defer.succeed(None)
+        d.addCallback(lambda ignored:
+            self.publish_mdmf())
+        d.addCallback(lambda ignored:
+            self.make_servermap(mode=MODE_CHECK))
+        # Calling make_servermap also updates the servermap in the mode
+        # that we specify, so we just need to see what it says.
+        def _check_servermap(sm):
+            self.failUnlessEqual(len(sm.recoverable_versions()), 1)
+        d.addCallback(_check_servermap)
+        return d
+
+
+    def test_servermapupdater_finds_sdmf_files(self):
+        d = defer.succeed(None)
+        d.addCallback(lambda ignored:
+            self.publish_sdmf())
+        d.addCallback(lambda ignored:
+            self.make_servermap(mode=MODE_CHECK))
+        d.addCallback(lambda servermap:
+            self.failUnlessEqual(len(servermap.recoverable_versions()), 1))
+        return d
+
 
 class Roundtrip(unittest.TestCase, testutil.ShouldFailMixin, PublishMixin):
     def setUp(self):
hunk ./src/allmydata/test/test_mutable.py 1050
         return d
     test_no_servers_download.timeout = 15
 
+
     def _test_corrupt_all(self, offset, substring,
                           should_succeed=False, corrupt_early=True,
                           failure_checker=None):
}
[Make a segmented mutable uploader
Kevan Carstensen <kevan@isnotajoke.com>**20100626234204
 Ignore-this: d199af8ab0bc64d8ed2bc19c5437bfba
 
 The mutable file uploader should be able to publish files with one
 segment and files with multiple segments. This patch makes it do that.
 This is still incomplete, and rather ugly -- I need to flesh out error
 handling, I need to write tests, and I need to remove some of the uglier
 kludges in the process before I can call this done.
] {
hunk ./src/allmydata/mutable/publish.py 8
 from zope.interface import implements
 from twisted.internet import defer
 from twisted.python import failure
-from allmydata.interfaces import IPublishStatus
+from allmydata.interfaces import IPublishStatus, SDMF_VERSION, MDMF_VERSION
 from allmydata.util import base32, hashutil, mathutil, idlib, log
 from allmydata import hashtree, codec
 from allmydata.storage.server import si_b2a
hunk ./src/allmydata/mutable/publish.py 19
      UncoordinatedWriteError, NotEnoughServersError
 from allmydata.mutable.servermap import ServerMap
 from allmydata.mutable.layout import pack_prefix, pack_share, unpack_header, pack_checkstring, \
-     unpack_checkstring, SIGNED_PREFIX
+     unpack_checkstring, SIGNED_PREFIX, MDMFSlotWriteProxy
+
+KiB = 1024
+DEFAULT_MAX_SEGMENT_SIZE = 128 * KiB
 
 class PublishStatus:
     implements(IPublishStatus)
hunk ./src/allmydata/mutable/publish.py 112
         self._status.set_helper(False)
         self._status.set_progress(0.0)
         self._status.set_active(True)
+        # We use this to control how the file is written.
+        version = self._node.get_version()
+        assert version in (SDMF_VERSION, MDMF_VERSION)
+        self._version = version
 
     def get_status(self):
         return self._status
hunk ./src/allmydata/mutable/publish.py 134
         simultaneous write.
         """
 
-        # 1: generate shares (SDMF: files are small, so we can do it in RAM)
-        # 2: perform peer selection, get candidate servers
-        #  2a: send queries to n+epsilon servers, to determine current shares
-        #  2b: based upon responses, create target map
-        # 3: send slot_testv_and_readv_and_writev messages
-        # 4: as responses return, update share-dispatch table
-        # 4a: may need to run recovery algorithm
-        # 5: when enough responses are back, we're done
+        # 0. Setup encoding parameters, encoder, and other such things.
+        # 1. Encrypt, encode, and publish segments.
 
         self.log("starting publish, datalen is %s" % len(newdata))
         self._status.set_size(len(newdata))
hunk ./src/allmydata/mutable/publish.py 187
         self.bad_peers = set() # peerids who have errbacked/refused requests
 
         self.newdata = newdata
-        self.salt = os.urandom(16)
 
hunk ./src/allmydata/mutable/publish.py 188
+        # This will set self.segment_size, self.num_segments, and
+        # self.fec.
         self.setup_encoding_parameters()
 
         # if we experience any surprises (writes which were rejected because
hunk ./src/allmydata/mutable/publish.py 238
             self.bad_share_checkstrings[key] = old_checkstring
             self.connections[peerid] = self._servermap.connections[peerid]
 
-        # create the shares. We'll discard these as they are delivered. SDMF:
-        # we're allowed to hold everything in memory.
+        # Now, the process dovetails -- if this is an SDMF file, we need
+        # to write an SDMF file. Otherwise, we need to write an MDMF
+        # file.
+        if self._version == MDMF_VERSION:
+            return self._publish_mdmf()
+        else:
+            return self._publish_sdmf()
+        #return self.done_deferred
+
+    def _publish_mdmf(self):
+        # Next, we find homes for all of the shares that we don't have
+        # homes for yet.
+        # TODO: Make this part do peer selection.
+        self.update_goal()
+        self.writers = {}
+        # For each (peerid, shnum) in self.goal, we make an
+        # MDMFSlotWriteProxy for that peer. We'll use this to write
+        # shares to the peer.
+        for key in self.goal:
+            peerid, shnum = key
+            write_enabler = self._node.get_write_enabler(peerid)
+            renew_secret = self._node.get_renewal_secret(peerid)
+            cancel_secret = self._node.get_cancel_secret(peerid)
+            secrets = (write_enabler, renew_secret, cancel_secret)
+
+            self.writers[shnum] =  MDMFSlotWriteProxy(shnum,
+                                                      self.connections[peerid],
+                                                      self._storage_index,
+                                                      secrets,
+                                                      self._new_seqnum,
+                                                      self.required_shares,
+                                                      self.total_shares,
+                                                      self.segment_size,
+                                                      len(self.newdata))
+            if (peerid, shnum) in self._servermap.servermap:
+                old_versionid, old_timestamp = self._servermap.servermap[key]
+                (old_seqnum, old_root_hash, old_salt, old_segsize,
+                 old_datalength, old_k, old_N, old_prefix,
+                 old_offsets_tuple) = old_versionid
+                self.writers[shnum].set_checkstring(old_seqnum, old_root_hash)
+
+        # Now, we start pushing shares.
+        self._status.timings["setup"] = time.time() - self._started
+        def _start_pushing(res):
+            self._started_pushing = time.time()
+            return res
+
+        # First, we encrypt, encode, and publish the shares that we need
+        # to encrypt, encode, and publish.
+
+        # This will eventually hold the block hash chain for each share
+        # that we publish. We define it this way so that empty publishes
+        # will still have something to write to the remote slot.
+        self.blockhashes = dict([(i, []) for i in xrange(self.total_shares)])
+        self.sharehash_leaves = None # eventually [sharehashes]
+        self.sharehashes = {} # shnum -> [sharehash leaves necessary to 
+                              # validate the share]
 
hunk ./src/allmydata/mutable/publish.py 296
+        d = defer.succeed(None)
+        self.log("Starting push")
+        for i in xrange(self.num_segments - 1):
+            d.addCallback(lambda ignored, i=i:
+                self.push_segment(i))
+            d.addCallback(self._turn_barrier)
+        # We have at least one segment, so we will have a tail segment
+        if self.num_segments > 0:
+            d.addCallback(lambda ignored:
+                self.push_tail_segment())
+
+        d.addCallback(lambda ignored:
+            self.push_encprivkey())
+        d.addCallback(lambda ignored:
+            self.push_blockhashes())
+        d.addCallback(lambda ignored:
+            self.push_sharehashes())
+        d.addCallback(lambda ignored:
+            self.push_toplevel_hashes_and_signature())
+        d.addCallback(lambda ignored:
+            self.finish_publishing())
+        return d
+
+
+    def _publish_sdmf(self):
         self._status.timings["setup"] = time.time() - self._started
hunk ./src/allmydata/mutable/publish.py 322
+        self.salt = os.urandom(16)
+
         d = self._encrypt_and_encode()
         d.addCallback(self._generate_shares)
         def _start_pushing(res):
hunk ./src/allmydata/mutable/publish.py 335
 
         return self.done_deferred
 
+
     def setup_encoding_parameters(self):
hunk ./src/allmydata/mutable/publish.py 337
-        segment_size = len(self.newdata)
+        if self._version == MDMF_VERSION:
+            segment_size = DEFAULT_MAX_SEGMENT_SIZE # 128 KiB by default
+        else:
+            segment_size = len(self.newdata) # SDMF is only one segment
         # this must be a multiple of self.required_shares
         segment_size = mathutil.next_multiple(segment_size,
                                               self.required_shares)
hunk ./src/allmydata/mutable/publish.py 350
                                                   segment_size)
         else:
             self.num_segments = 0
-        assert self.num_segments in [0, 1,] # SDMF restrictions
+        if self._version == SDMF_VERSION:
+            assert self.num_segments in (0, 1) # SDMF
+            return
+        # calculate the tail segment size.
+        self.tail_segment_size = len(self.newdata) % segment_size
+
+        if self.tail_segment_size == 0:
+            # The tail segment is the same size as the other segments.
+            self.tail_segment_size = segment_size
+
+        # We'll make an encoder ahead-of-time for the normal-sized
+        # segments (defined as any segment of segment_size size.
+        # (the part of the code that puts the tail segment will make its
+        #  own encoder for that part)
+        fec = codec.CRSEncoder()
+        fec.set_params(self.segment_size,
+                       self.required_shares, self.total_shares)
+        self.piece_size = fec.get_block_size()
+        self.fec = fec
+
+
+    def push_segment(self, segnum):
+        started = time.time()
+        segsize = self.segment_size
+        self.log("Pushing segment %d of %d" % (segnum + 1, self.num_segments))
+        data = self.newdata[segsize * segnum:segsize*(segnum + 1)]
+        assert len(data) == segsize
+
+        salt = os.urandom(16)
+
+        key = hashutil.ssk_readkey_data_hash(salt, self.readkey)
+        enc = AES(key)
+        crypttext = enc.process(data)
+        assert len(crypttext) == len(data)
+
+        now = time.time()
+        self._status.timings["encrypt"] = now - started
+        started = now
+
+        # now apply FEC
+
+        self._status.set_status("Encoding")
+        crypttext_pieces = [None] * self.required_shares
+        piece_size = self.piece_size
+        for i in range(len(crypttext_pieces)):
+            offset = i * piece_size
+            piece = crypttext[offset:offset+piece_size]
+            piece = piece + "\x00"*(piece_size - len(piece)) # padding
+            crypttext_pieces[i] = piece
+            assert len(piece) == piece_size
+        d = self.fec.encode(crypttext_pieces)
+        def _done_encoding(res):
+            elapsed = time.time() - started
+            self._status.timings["encode"] = elapsed
+            return res
+        d.addCallback(_done_encoding)
+
+        def _push_shares_and_salt(results):
+            shares, shareids = results
+            dl = []
+            for i in xrange(len(shares)):
+                sharedata = shares[i]
+                shareid = shareids[i]
+                block_hash = hashutil.block_hash(salt + sharedata)
+                self.blockhashes[shareid].append(block_hash)
+
+                # find the writer for this share
+                d = self.writers[shareid].put_block(sharedata, segnum, salt)
+                dl.append(d)
+            # TODO: Naturally, we need to check on the results of these.
+            return defer.DeferredList(dl)
+        d.addCallback(_push_shares_and_salt)
+        return d
+
+
+    def push_tail_segment(self):
+        # This is essentially the same as push_segment, except that we
+        # don't use the cached encoder that we use elsewhere.
+        self.log("Pushing tail segment")
+        started = time.time()
+        segsize = self.segment_size
+        data = self.newdata[segsize * (self.num_segments-1):]
+        assert len(data) == self.tail_segment_size
+        salt = os.urandom(16)
+
+        key = hashutil.ssk_readkey_data_hash(salt, self.readkey)
+        enc = AES(key)
+        crypttext = enc.process(data)
+        assert len(crypttext) == len(data)
+
+        now = time.time()
+        self._status.timings['encrypt'] = now - started
+        started = now
+
+        self._status.set_status("Encoding")
+        tail_fec = codec.CRSEncoder()
+        tail_fec.set_params(self.tail_segment_size,
+                            self.required_shares,
+                            self.total_shares)
+
+        crypttext_pieces = [None] * self.required_shares
+        piece_size = tail_fec.get_block_size()
+        for i in range(len(crypttext_pieces)):
+            offset = i * piece_size
+            piece = crypttext[offset:offset+piece_size]
+            piece = piece + "\x00"*(piece_size - len(piece)) # padding
+            crypttext_pieces[i] = piece
+            assert len(piece) == piece_size
+        d = tail_fec.encode(crypttext_pieces)
+        def _push_shares_and_salt(results):
+            shares, shareids = results
+            dl = []
+            for i in xrange(len(shares)):
+                sharedata = shares[i]
+                shareid = shareids[i]
+                block_hash = hashutil.block_hash(salt + sharedata)
+                self.blockhashes[shareid].append(block_hash)
+                # find the writer for this share
+                d = self.writers[shareid].put_block(sharedata,
+                                                    self.num_segments - 1,
+                                                    salt)
+                dl.append(d)
+            # TODO: Naturally, we need to check on the results of these.
+            return defer.DeferredList(dl)
+        d.addCallback(_push_shares_and_salt)
+        return d
+
+
+    def push_encprivkey(self):
+        started = time.time()
+        encprivkey = self._encprivkey
+        dl = []
+        def _spy_on_writer(results):
+            print results
+            return results
+        for shnum, writer in self.writers.iteritems():
+            d = writer.put_encprivkey(encprivkey)
+            dl.append(d)
+        d = defer.DeferredList(dl)
+        return d
+
+
+    def push_blockhashes(self):
+        started = time.time()
+        dl = []
+        def _spy_on_results(results):
+            print results
+            return results
+        self.sharehash_leaves = [None] * len(self.blockhashes)
+        for shnum, blockhashes in self.blockhashes.iteritems():
+            t = hashtree.HashTree(blockhashes)
+            self.blockhashes[shnum] = list(t)
+            # set the leaf for future use.
+            self.sharehash_leaves[shnum] = t[0]
+            d = self.writers[shnum].put_blockhashes(self.blockhashes[shnum])
+            dl.append(d)
+        d = defer.DeferredList(dl)
+        return d
+
+
+    def push_sharehashes(self):
+        share_hash_tree = hashtree.HashTree(self.sharehash_leaves)
+        share_hash_chain = {}
+        ds = []
+        def _spy_on_results(results):
+            print results
+            return results
+        for shnum in xrange(len(self.sharehash_leaves)):
+            needed_indices = share_hash_tree.needed_hashes(shnum)
+            self.sharehashes[shnum] = dict( [ (i, share_hash_tree[i])
+                                             for i in needed_indices] )
+            d = self.writers[shnum].put_sharehashes(self.sharehashes[shnum])
+            ds.append(d)
+        self.root_hash = share_hash_tree[0]
+        d = defer.DeferredList(ds)
+        return d
+
+
+    def push_toplevel_hashes_and_signature(self):
+        # We need to to three things here:
+        #   - Push the root hash and salt hash
+        #   - Get the checkstring of the resulting layout; sign that.
+        #   - Push the signature
+        ds = []
+        def _spy_on_results(results):
+            print results
+            return results
+        for shnum in xrange(self.total_shares):
+            d = self.writers[shnum].put_root_hash(self.root_hash)
+            ds.append(d)
+        d = defer.DeferredList(ds)
+        def _make_and_place_signature(ignored):
+            signable = self.writers[0].get_signable()
+            self.signature = self._privkey.sign(signable)
+
+            ds = []
+            for (shnum, writer) in self.writers.iteritems():
+                d = writer.put_signature(self.signature)
+                ds.append(d)
+            return defer.DeferredList(ds)
+        d.addCallback(_make_and_place_signature)
+        return d
+
+
+    def finish_publishing(self):
+        # We're almost done -- we just need to put the verification key
+        # and the offsets
+        ds = []
+        verification_key = self._pubkey.serialize()
+
+        def _spy_on_results(results):
+            print results
+            return results
+        for (shnum, writer) in self.writers.iteritems():
+            d = writer.put_verification_key(verification_key)
+            d.addCallback(lambda ignored, writer=writer:
+                writer.finish_publishing())
+            ds.append(d)
+        return defer.DeferredList(ds)
+
+
+    def _turn_barrier(self, res):
+        # putting this method in a Deferred chain imposes a guaranteed
+        # reactor turn between the pre- and post- portions of that chain.
+        # This can be useful to limit memory consumption: since Deferreds do
+        # not do tail recursion, code which uses defer.succeed(result) for
+        # consistency will cause objects to live for longer than you might
+        # normally expect.
+        return fireEventually(res)
+
 
     def _fatal_error(self, f):
         self.log("error during loop", failure=f, level=log.UNUSUAL)
hunk ./src/allmydata/mutable/publish.py 716
             self.log_goal(self.goal, "after update: ")
 
 
-
     def _encrypt_and_encode(self):
         # this returns a Deferred that fires with a list of (sharedata,
         # sharenum) tuples. TODO: cache the ciphertext, only produce the
hunk ./src/allmydata/mutable/publish.py 757
         d.addCallback(_done_encoding)
         return d
 
+
     def _generate_shares(self, shares_and_shareids):
         # this sets self.shares and self.root_hash
         self.log("_generate_shares")
hunk ./src/allmydata/mutable/publish.py 1145
             self._status.set_progress(1.0)
         eventually(self.done_deferred.callback, res)
 
-
hunk ./src/allmydata/test/test_mutable.py 248
         d.addCallback(_created)
         return d
 
+
+    def test_create_mdmf(self):
+        d = self.nodemaker.create_mutable_file(version=MDMF_VERSION)
+        def _created(n):
+            self.failUnless(isinstance(n, MutableFileNode))
+            self.failUnlessEqual(n.get_storage_index(), n._storage_index)
+            sb = self.nodemaker.storage_broker
+            peer0 = sorted(sb.get_all_serverids())[0]
+            shnums = self._storage._peers[peer0].keys()
+            self.failUnlessEqual(len(shnums), 1)
+        d.addCallback(_created)
+        return d
+
+
     def test_serialize(self):
         n = MutableFileNode(None, None, {"k": 3, "n": 10}, None)
         calls = []
hunk ./src/allmydata/test/test_mutable.py 334
         d.addCallback(_created)
         return d
 
+
+    def test_create_mdmf_with_initial_contents(self):
+        initial_contents = "foobarbaz" * 131072 # 900KiB
+        d = self.nodemaker.create_mutable_file(initial_contents,
+                                               version=MDMF_VERSION)
+        def _created(n):
+            d = n.download_best_version()
+            d.addCallback(lambda data:
+                self.failUnlessEqual(data, initial_contents))
+            d.addCallback(lambda ignored:
+                n.overwrite(initial_contents + "foobarbaz"))
+            d.addCallback(lambda ignored:
+                n.download_best_version())
+            d.addCallback(lambda data:
+                self.failUnlessEqual(data, initial_contents +
+                                           "foobarbaz"))
+            return d
+        d.addCallback(_created)
+        return d
+
+
     def test_create_with_initial_contents_function(self):
         data = "initial contents"
         def _make_contents(n):
hunk ./src/allmydata/test/test_mutable.py 370
         d.addCallback(lambda data2: self.failUnlessEqual(data2, data))
         return d
 
+
+    def test_create_mdmf_with_initial_contents_function(self):
+        data = "initial contents" * 100000
+        def _make_contents(n):
+            self.failUnless(isinstance(n, MutableFileNode))
+            key = n.get_writekey()
+            self.failUnless(isinstance(key, str), key)
+            self.failUnlessEqual(len(key), 16)
+            return data
+        d = self.nodemaker.create_mutable_file(_make_contents,
+                                               version=MDMF_VERSION)
+        d.addCallback(lambda n:
+            n.download_best_version())
+        d.addCallback(lambda data2:
+            self.failUnlessEqual(data2, data))
+        return d
+
+
     def test_create_with_too_large_contents(self):
         BIG = "a" * (self.OLD_MAX_SEGMENT_SIZE + 1)
         d = self.nodemaker.create_mutable_file(BIG)
}
[Write a segmented mutable downloader
Kevan Carstensen <kevan@isnotajoke.com>**20100626234314
 Ignore-this: d2bef531cde1b5c38f2eb28afdd4b17c
 
 The segmented mutable downloader can deal with MDMF files (files with
 one or more segments in MDMF format) and SDMF files (files with one
 segment in SDMF format). It is backwards compatible with the old
 file format.
 
 This patch also contains tests for the segmented mutable downloader.
] {
hunk ./src/allmydata/mutable/retrieve.py 8
 from twisted.internet import defer
 from twisted.python import failure
 from foolscap.api import DeadReferenceError, eventually, fireEventually
-from allmydata.interfaces import IRetrieveStatus, NotEnoughSharesError
-from allmydata.util import hashutil, idlib, log
+from allmydata.interfaces import IRetrieveStatus, NotEnoughSharesError, \
+                                 MDMF_VERSION, SDMF_VERSION
+from allmydata.util import hashutil, idlib, log, mathutil
 from allmydata import hashtree, codec
 from allmydata.storage.server import si_b2a
 from pycryptopp.cipher.aes import AES
hunk ./src/allmydata/mutable/retrieve.py 17
 from pycryptopp.publickey import rsa
 
 from allmydata.mutable.common import DictOfSets, CorruptShareError, UncoordinatedWriteError
-from allmydata.mutable.layout import SIGNED_PREFIX, unpack_share_data
+from allmydata.mutable.layout import SIGNED_PREFIX, unpack_share_data, \
+                                     MDMFSlotReadProxy
 
 class RetrieveStatus:
     implements(IRetrieveStatus)
hunk ./src/allmydata/mutable/retrieve.py 104
         self.verinfo = verinfo
         # during repair, we may be called upon to grab the private key, since
         # it wasn't picked up during a verify=False checker run, and we'll
-        # need it for repair to generate the a new version.
+        # need it for repair to generate a new version.
         self._need_privkey = fetch_privkey
         if self._node.get_privkey():
             self._need_privkey = False
hunk ./src/allmydata/mutable/retrieve.py 109
 
+        if self._need_privkey:
+            # TODO: Evaluate the need for this. We'll use it if we want
+            # to limit how many queries are on the wire for the privkey
+            # at once.
+            self._privkey_query_markers = [] # one Marker for each time we've
+                                             # tried to get the privkey.
+
         self._status = RetrieveStatus()
         self._status.set_storage_index(self._storage_index)
         self._status.set_helper(False)
hunk ./src/allmydata/mutable/retrieve.py 125
          offsets_tuple) = self.verinfo
         self._status.set_size(datalength)
         self._status.set_encoding(k, N)
+        self.readers = {}
 
     def get_status(self):
         return self._status
hunk ./src/allmydata/mutable/retrieve.py 149
         self.remaining_sharemap = DictOfSets()
         for (shnum, peerid, timestamp) in shares:
             self.remaining_sharemap.add(shnum, peerid)
+            # If the servermap update fetched anything, it fetched at least 1
+            # KiB, so we ask for that much.
+            # TODO: Change the cache methods to allow us to fetch all of the
+            # data that they have, then change this method to do that.
+            any_cache, timestamp = self._node._read_from_cache(self.verinfo,
+                                                               shnum,
+                                                               0,
+                                                               1000)
+            ss = self.servermap.connections[peerid]
+            reader = MDMFSlotReadProxy(ss,
+                                       self._storage_index,
+                                       shnum,
+                                       any_cache)
+            reader.peerid = peerid
+            self.readers[shnum] = reader
+
 
         self.shares = {} # maps shnum to validated blocks
hunk ./src/allmydata/mutable/retrieve.py 167
+        self._active_readers = [] # list of active readers for this dl.
+        self._validated_readers = set() # set of readers that we have
+                                        # validated the prefix of
+        self._block_hash_trees = {} # shnum => hashtree
+        # TODO: Make this into a file-backed consumer or something to
+        # conserve memory.
+        self._plaintext = ""
 
         # how many shares do we need?
hunk ./src/allmydata/mutable/retrieve.py 176
-        (seqnum, root_hash, IV, segsize, datalength, k, N, prefix,
+        (seqnum,
+         root_hash,
+         IV,
+         segsize,
+         datalength,
+         k,
+         N,
+         prefix,
          offsets_tuple) = self.verinfo
hunk ./src/allmydata/mutable/retrieve.py 185
-        assert len(self.remaining_sharemap) >= k
-        # we start with the lowest shnums we have available, since FEC is
-        # faster if we're using "primary shares"
-        self.active_shnums = set(sorted(self.remaining_sharemap.keys())[:k])
-        for shnum in self.active_shnums:
-            # we use an arbitrary peer who has the share. If shares are
-            # doubled up (more than one share per peer), we could make this
-            # run faster by spreading the load among multiple peers. But the
-            # algorithm to do that is more complicated than I want to write
-            # right now, and a well-provisioned grid shouldn't have multiple
-            # shares per peer.
-            peerid = list(self.remaining_sharemap[shnum])[0]
-            self.get_data(shnum, peerid)
 
hunk ./src/allmydata/mutable/retrieve.py 186
-        # control flow beyond this point: state machine. Receiving responses
-        # from queries is the input. We might send out more queries, or we
-        # might produce a result.
 
hunk ./src/allmydata/mutable/retrieve.py 187
+        # We need one share hash tree for the entire file; its leaves
+        # are the roots of the block hash trees for the shares that
+        # comprise it, and its root is in the verinfo.
+        self.share_hash_tree = hashtree.IncompleteHashTree(N)
+        self.share_hash_tree.set_hashes({0: root_hash})
+
+        # This will set up both the segment decoder and the tail segment
+        # decoder, as well as a variety of other instance variables that
+        # the download process will use.
+        self._setup_encoding_parameters()
+        assert len(self.remaining_sharemap) >= k
+
+        self.log("starting download")
+        self._add_active_peers()
+        # The download process beyond this is a state machine.
+        # _add_active_peers will select the peers that we want to use
+        # for the download, and then attempt to start downloading. After
+        # each segment, it will check for doneness, reacting to broken
+        # peers and corrupt shares as necessary. If it runs out of good
+        # peers before downloading all of the segments, _done_deferred
+        # will errback.  Otherwise, it will eventually callback with the
+        # contents of the mutable file.
         return self._done_deferred
 
hunk ./src/allmydata/mutable/retrieve.py 211
-    def get_data(self, shnum, peerid):
-        self.log(format="sending sh#%(shnum)d request to [%(peerid)s]",
-                 shnum=shnum,
-                 peerid=idlib.shortnodeid_b2a(peerid),
-                 level=log.NOISY)
-        ss = self.servermap.connections[peerid]
-        started = time.time()
-        (seqnum, root_hash, IV, segsize, datalength, k, N, prefix,
+
+    def _setup_encoding_parameters(self):
+        """
+        I set up the encoding parameters, including k, n, the number
+        of segments associated with this file, and the segment decoder.
+        """
+        (seqnum,
+         root_hash,
+         IV,
+         segsize,
+         datalength,
+         k,
+         n,
+         known_prefix,
          offsets_tuple) = self.verinfo
hunk ./src/allmydata/mutable/retrieve.py 226
-        offsets = dict(offsets_tuple)
+        self._required_shares = k
+        self._total_shares = n
+        self._segment_size = segsize
+        self._data_length = datalength
+
+        if not IV:
+            self._version = MDMF_VERSION
+        else:
+            self._version = SDMF_VERSION
+
+        if datalength and segsize:
+            self._num_segments = mathutil.div_ceil(datalength, segsize)
+            self._tail_data_size = datalength % segsize
+        else:
+            self._num_segments = 0
+            self._tail_data_size = 0
 
hunk ./src/allmydata/mutable/retrieve.py 243
-        # we read the checkstring, to make sure that the data we grab is from
-        # the right version.
-        readv = [ (0, struct.calcsize(SIGNED_PREFIX)) ]
+        self._segment_decoder = codec.CRSDecoder()
+        self._segment_decoder.set_params(segsize, k, n)
+        self._current_segment = 0
 
hunk ./src/allmydata/mutable/retrieve.py 247
-        # We also read the data, and the hashes necessary to validate them
-        # (share_hash_chain, block_hash_tree, share_data). We don't read the
-        # signature or the pubkey, since that was handled during the
-        # servermap phase, and we'll be comparing the share hash chain
-        # against the roothash that was validated back then.
+        if  not self._tail_data_size:
+            self._tail_data_size = segsize
 
hunk ./src/allmydata/mutable/retrieve.py 250
-        readv.append( (offsets['share_hash_chain'],
-                       offsets['enc_privkey'] - offsets['share_hash_chain'] ) )
+        self._tail_segment_size = mathutil.next_multiple(self._tail_data_size,
+                                                         self._required_shares)
+        if self._tail_segment_size == self._segment_size:
+            self._tail_decoder = self._segment_decoder
+        else:
+            self._tail_decoder = codec.CRSDecoder()
+            self._tail_decoder.set_params(self._tail_segment_size,
+                                          self._required_shares,
+                                          self._total_shares)
 
hunk ./src/allmydata/mutable/retrieve.py 260
-        # if we need the private key (for repair), we also fetch that
-        if self._need_privkey:
-            readv.append( (offsets['enc_privkey'],
-                           offsets['EOF'] - offsets['enc_privkey']) )
+        self.log("got encoding parameters: "
+                 "k: %d "
+                 "n: %d "
+                 "%d segments of %d bytes each (%d byte tail segment)" % \
+                 (k, n, self._num_segments, self._segment_size,
+                  self._tail_segment_size))
 
hunk ./src/allmydata/mutable/retrieve.py 267
-        m = Marker()
-        self._outstanding_queries[m] = (peerid, shnum, started)
+        for i in xrange(self._total_shares):
+            # So we don't have to do this later.
+            self._block_hash_trees[i] = hashtree.IncompleteHashTree(self._num_segments)
 
hunk ./src/allmydata/mutable/retrieve.py 271
-        # ask the cache first
-        got_from_cache = False
-        datavs = []
-        for (offset, length) in readv:
-            (data, timestamp) = self._node._read_from_cache(self.verinfo, shnum,
-                                                            offset, length)
-            if data is not None:
-                datavs.append(data)
-        if len(datavs) == len(readv):
-            self.log("got data from cache")
-            got_from_cache = True
-            d = fireEventually({shnum: datavs})
-            # datavs is a dict mapping shnum to a pair of strings
-        else:
-            d = self._do_read(ss, peerid, self._storage_index, [shnum], readv)
-        self.remaining_sharemap.discard(shnum, peerid)
+        # If we have more than one segment, we are an SDMF file, which
+        # means that we need to validate the salts as we receive them.
+        self._salt_hash_tree = hashtree.IncompleteHashTree(self._num_segments)
+        self._salt_hash_tree[0] = IV # from the prefix.
 
hunk ./src/allmydata/mutable/retrieve.py 276
-        d.addCallback(self._got_results, m, peerid, started, got_from_cache)
-        d.addErrback(self._query_failed, m, peerid)
-        # errors that aren't handled by _query_failed (and errors caused by
-        # _query_failed) get logged, but we still want to check for doneness.
-        def _oops(f):
-            self.log(format="problem in _query_failed for sh#%(shnum)d to %(peerid)s",
-                     shnum=shnum,
-                     peerid=idlib.shortnodeid_b2a(peerid),
-                     failure=f,
-                     level=log.WEIRD, umid="W0xnQA")
-        d.addErrback(_oops)
-        d.addBoth(self._check_for_done)
-        # any error during _check_for_done means the download fails. If the
-        # download is successful, _check_for_done will fire _done by itself.
-        d.addErrback(self._done)
-        d.addErrback(log.err)
-        return d # purely for testing convenience
 
hunk ./src/allmydata/mutable/retrieve.py 277
-    def _do_read(self, ss, peerid, storage_index, shnums, readv):
-        # isolate the callRemote to a separate method, so tests can subclass
-        # Publish and override it
-        d = ss.callRemote("slot_readv", storage_index, shnums, readv)
-        return d
+    def _add_active_peers(self):
+        """
+        I populate self._active_readers with enough active readers to
+        retrieve the contents of this mutable file. I am called before
+        downloading starts, and (eventually) after each validation
+        error, connection error, or other problem in the download.
+        """
+        # TODO: It would be cool to investigate other heuristics for
+        # reader selection. For instance, the cost (in time the user
+        # spends waiting for their file) of selecting a really slow peer
+        # that happens to have a primary share is probably more than
+        # selecting a really fast peer that doesn't have a primary
+        # share. Maybe the servermap could be extended to provide this
+        # information; it could keep track of latency information while
+        # it gathers more important data, and then this routine could
+        # use that to select active readers.
+        #
+        # (these and other questions would be easier to answer with a
+        #  robust, configurable tahoe-lafs simulator, which modeled node
+        #  failures, differences in node speed, and other characteristics
+        #  that we expect storage servers to have.  You could have
+        #  presets for really stable grids (like allmydata.com),
+        #  friendnets, make it easy to configure your own settings, and
+        #  then simulate the effect of big changes on these use cases
+        #  instead of just reasoning about what the effect might be. Out
+        #  of scope for MDMF, though.)
 
hunk ./src/allmydata/mutable/retrieve.py 304
-    def remove_peer(self, peerid):
-        for shnum in list(self.remaining_sharemap.keys()):
-            self.remaining_sharemap.discard(shnum, peerid)
+        # We need at least self._required_shares readers to download a
+        # segment.
+        needed = self._required_shares - len(self._active_readers)
+        # XXX: Why don't format= log messages work here?
+        self.log("adding %d peers to the active peers list" % needed)
 
hunk ./src/allmydata/mutable/retrieve.py 310
-    def _got_results(self, datavs, marker, peerid, started, got_from_cache):
-        now = time.time()
-        elapsed = now - started
-        if not got_from_cache:
-            self._status.add_fetch_timing(peerid, elapsed)
-        self.log(format="got results (%(shares)d shares) from [%(peerid)s]",
-                 shares=len(datavs),
-                 peerid=idlib.shortnodeid_b2a(peerid),
-                 level=log.NOISY)
-        self._outstanding_queries.pop(marker, None)
-        if not self._running:
-            return
+        # We favor lower numbered shares, since FEC is faster with
+        # primary shares than with other shares, and lower-numbered
+        # shares are more likely to be primary than higher numbered
+        # shares.
+        active_shnums = set(sorted(self.remaining_sharemap.keys()))
+        # We shouldn't consider adding shares that we already have; this
+        # will cause problems later.
+        active_shnums -= set([reader.shnum for reader in self._active_readers])
+        active_shnums = list(active_shnums)[:needed]
+        if len(active_shnums) < needed:
+            # We don't have enough readers to retrieve the file; fail.
+            return self._failed()
 
hunk ./src/allmydata/mutable/retrieve.py 323
-        # note that we only ask for a single share per query, so we only
-        # expect a single share back. On the other hand, we use the extra
-        # shares if we get them.. seems better than an assert().
+        for shnum in active_shnums:
+            self._active_readers.append(self.readers[shnum])
+            self.log("added reader for share %d" % shnum)
+        assert len(self._active_readers) == self._required_shares
+        # Conceptually, this is part of the _add_active_peers step. It
+        # validates the prefixes of newly added readers to make sure
+        # that they match what we are expecting for self.verinfo. If
+        # validation is successful, _validate_active_prefixes will call
+        # _download_current_segment for us. If validation is
+        # unsuccessful, then _validate_prefixes will remove the peer and
+        # call _add_active_peers again, where we will attempt to rectify
+        # the problem by choosing another peer.
+        return self._validate_active_prefixes()
 
hunk ./src/allmydata/mutable/retrieve.py 337
-        for shnum,datav in datavs.items():
-            (prefix, hash_and_data) = datav[:2]
-            try:
-                self._got_results_one_share(shnum, peerid,
-                                            prefix, hash_and_data)
-            except CorruptShareError, e:
-                # log it and give the other shares a chance to be processed
-                f = failure.Failure()
-                self.log(format="bad share: %(f_value)s",
-                         f_value=str(f.value), failure=f,
-                         level=log.WEIRD, umid="7fzWZw")
-                self.notify_server_corruption(peerid, shnum, str(e))
-                self.remove_peer(peerid)
-                self.servermap.mark_bad_share(peerid, shnum, prefix)
-                self._bad_shares.add( (peerid, shnum) )
-                self._status.problems[peerid] = f
-                self._last_failure = f
-                pass
-            if self._need_privkey and len(datav) > 2:
-                lp = None
-                self._try_to_validate_privkey(datav[2], peerid, shnum, lp)
-        # all done!
 
hunk ./src/allmydata/mutable/retrieve.py 338
-    def notify_server_corruption(self, peerid, shnum, reason):
-        ss = self.servermap.connections[peerid]
-        ss.callRemoteOnly("advise_corrupt_share",
-                          "mutable", self._storage_index, shnum, reason)
+    def _validate_active_prefixes(self):
+        """
+        I check to make sure that the prefixes on the peers that I am
+        currently reading from match the prefix that we want to see, as
+        said in self.verinfo.
 
hunk ./src/allmydata/mutable/retrieve.py 344
-    def _got_results_one_share(self, shnum, peerid,
-                               got_prefix, got_hash_and_data):
-        self.log("_got_results: got shnum #%d from peerid %s"
-                 % (shnum, idlib.shortnodeid_b2a(peerid)))
-        (seqnum, root_hash, IV, segsize, datalength, k, N, prefix,
+        If I find that all of the active peers have acceptable prefixes,
+        I pass control to _download_current_segment, which will use
+        those peers to do cool things. If I find that some of the active
+        peers have unacceptable prefixes, I will remove them from active
+        peers (and from further consideration) and call
+        _add_active_peers to attempt to rectify the situation. I keep
+        track of which peers I have already validated so that I don't
+        need to do so again.
+        """
+        assert self._active_readers, "No more active readers"
+
+        ds = []
+        new_readers = set(self._active_readers) - self._validated_readers
+        self.log('validating %d newly-added active readers' % len(new_readers))
+
+        for reader in new_readers:
+            # We force a remote read here -- otherwise, we are relying
+            # on cached data that we already verified as valid, and we
+            # won't detect an uncoordinated write that has occurred
+            # since the last servermap update.
+            d = reader.get_prefix(force_remote=True)
+            d.addCallback(self._try_to_validate_prefix, reader)
+            ds.append(d)
+        dl = defer.DeferredList(ds, consumeErrors=True)
+        def _check_results(results):
+            # Each result in results will be of the form (success, msg).
+            # We don't care about msg, but success will tell us whether
+            # or not the checkstring validated. If it didn't, we need to
+            # remove the offending (peer,share) from our active readers,
+            # and ensure that active readers is again populated.
+            bad_readers = []
+            for i, result in enumerate(results):
+                if not result[0]:
+                    reader = self._active_readers[i]
+                    f = result[1]
+                    assert isinstance(f, failure.Failure)
+
+                    self.log("The reader %s failed to "
+                             "properly validate: %s" % \
+                             (reader, str(f.value)))
+                    bad_readers.append((reader, f))
+                else:
+                    reader = self._active_readers[i]
+                    self.log("the reader %s checks out, so we'll use it" % \
+                             reader)
+                    self._validated_readers.add(reader)
+                    # Each time we validate a reader, we check to see if
+                    # we need the private key. If we do, we politely ask
+                    # for it and then continue computing. If we find
+                    # that we haven't gotten it at the end of
+                    # segment decoding, then we'll take more drastic
+                    # measures.
+                    if self._need_privkey:
+                        d = reader.get_encprivkey()
+                        d.addCallback(self._try_to_validate_privkey, reader)
+            if bad_readers:
+                # We do them all at once, or else we screw up list indexing.
+                for (reader, f) in bad_readers:
+                    self._mark_bad_share(reader, f)
+                return self._add_active_peers()
+            else:
+                return self._download_current_segment()
+            # The next step will assert that it has enough active
+            # readers to fetch shares; we just need to remove it.
+        dl.addCallback(_check_results)
+        return dl
+
+
+    def _try_to_validate_prefix(self, prefix, reader):
+        """
+        I check that the prefix returned by a candidate server for
+        retrieval matches the prefix that the servermap knows about
+        (and, hence, the prefix that was validated earlier). If it does,
+        I return True, which means that I approve of the use of the
+        candidate server for segment retrieval. If it doesn't, I return
+        False, which means that another server must be chosen.
+        """
+        (seqnum,
+         root_hash,
+         IV,
+         segsize,
+         datalength,
+         k,
+         N,
+         known_prefix,
          offsets_tuple) = self.verinfo
hunk ./src/allmydata/mutable/retrieve.py 430
-        assert len(got_prefix) == len(prefix), (len(got_prefix), len(prefix))
-        if got_prefix != prefix:
-            msg = "someone wrote to the data since we read the servermap: prefix changed"
-            raise UncoordinatedWriteError(msg)
-        (share_hash_chain, block_hash_tree,
-         share_data) = unpack_share_data(self.verinfo, got_hash_and_data)
+        if known_prefix != prefix:
+            self.log("prefix from share %d doesn't match" % reader.shnum)
+            raise UncoordinatedWriteError("Mismatched prefix -- this could "
+                                          "indicate an uncoordinated write")
+        # Otherwise, we're okay -- no issues.
 
hunk ./src/allmydata/mutable/retrieve.py 436
-        assert isinstance(share_data, str)
-        # build the block hash tree. SDMF has only one leaf.
-        leaves = [hashutil.block_hash(share_data)]
-        t = hashtree.HashTree(leaves)
-        if list(t) != block_hash_tree:
-            raise CorruptShareError(peerid, shnum, "block hash tree failure")
-        share_hash_leaf = t[0]
-        t2 = hashtree.IncompleteHashTree(N)
-        # root_hash was checked by the signature
-        t2.set_hashes({0: root_hash})
-        try:
-            t2.set_hashes(hashes=share_hash_chain,
-                          leaves={shnum: share_hash_leaf})
-        except (hashtree.BadHashError, hashtree.NotEnoughHashesError,
-                IndexError), e:
-            msg = "corrupt hashes: %s" % (e,)
-            raise CorruptShareError(peerid, shnum, msg)
-        self.log(" data valid! len=%d" % len(share_data))
-        # each query comes down to this: placing validated share data into
-        # self.shares
-        self.shares[shnum] = share_data
 
hunk ./src/allmydata/mutable/retrieve.py 437
-    def _try_to_validate_privkey(self, enc_privkey, peerid, shnum, lp):
+    def _remove_reader(self, reader):
+        """
+        At various points, we will wish to remove a peer from
+        consideration and/or use. These include, but are not necessarily
+        limited to:
 
hunk ./src/allmydata/mutable/retrieve.py 443
-        alleged_privkey_s = self._node._decrypt_privkey(enc_privkey)
-        alleged_writekey = hashutil.ssk_writekey_hash(alleged_privkey_s)
-        if alleged_writekey != self._node.get_writekey():
-            self.log("invalid privkey from %s shnum %d" %
-                     (idlib.nodeid_b2a(peerid)[:8], shnum),
-                     parent=lp, level=log.WEIRD, umid="YIw4tA")
-            return
+            - A connection error.
+            - A mismatched prefix (that is, a prefix that does not match
+              our conception of the version information string).
+            - A failing block hash, salt hash, or share hash, which can
+              indicate disk failure/bit flips, or network trouble.
 
hunk ./src/allmydata/mutable/retrieve.py 449
-        # it's good
-        self.log("got valid privkey from shnum %d on peerid %s" %
-                 (shnum, idlib.shortnodeid_b2a(peerid)),
-                 parent=lp)
-        privkey = rsa.create_signing_key_from_string(alleged_privkey_s)
-        self._node._populate_encprivkey(enc_privkey)
-        self._node._populate_privkey(privkey)
-        self._need_privkey = False
+        This method will do that. I will make sure that the
+        (shnum,reader) combination represented by my reader argument is
+        not used for anything else during this download. I will not
+        advise the reader of any corruption, something that my callers
+        may wish to do on their own.
+        """
+        # TODO: When you're done writing this, see if this is ever
+        # actually used for something that _mark_bad_share isn't. I have
+        # a feeling that they will be used for very similar things, and
+        # that having them both here is just going to be an epic amount
+        # of code duplication.
+        #
+        # (well, okay, not epic, but meaningful)
+        self.log("removing reader %s" % reader)
+        # Remove the reader from _active_readers
+        self._active_readers.remove(reader)
+        # TODO: self.readers.remove(reader)?
+        for shnum in list(self.remaining_sharemap.keys()):
+            self.remaining_sharemap.discard(shnum, reader.peerid)
 
hunk ./src/allmydata/mutable/retrieve.py 469
-    def _query_failed(self, f, marker, peerid):
-        self.log(format="query to [%(peerid)s] failed",
-                 peerid=idlib.shortnodeid_b2a(peerid),
-                 level=log.NOISY)
-        self._status.problems[peerid] = f
-        self._outstanding_queries.pop(marker, None)
-        if not self._running:
-            return
-        self._last_failure = f
-        self.remove_peer(peerid)
-        level = log.WEIRD
-        if f.check(DeadReferenceError):
-            level = log.UNUSUAL
-        self.log(format="error during query: %(f_value)s",
-                 f_value=str(f.value), failure=f, level=level, umid="gOJB5g")
 
hunk ./src/allmydata/mutable/retrieve.py 470
-    def _check_for_done(self, res):
-        # exit paths:
-        #  return : keep waiting, no new queries
-        #  return self._send_more_queries(outstanding) : send some more queries
-        #  fire self._done(plaintext) : download successful
-        #  raise exception : download fails
+    def _mark_bad_share(self, reader, f):
+        """
+        I mark the (peerid, shnum) encapsulated by my reader argument as
+        a bad share, which means that it will not be used anywhere else.
 
hunk ./src/allmydata/mutable/retrieve.py 475
-        self.log(format="_check_for_done: running=%(running)s, decoding=%(decoding)s",
-                 running=self._running, decoding=self._decoding,
-                 level=log.NOISY)
-        if not self._running:
-            return
-        if self._decoding:
-            return
-        (seqnum, root_hash, IV, segsize, datalength, k, N, prefix,
-         offsets_tuple) = self.verinfo
+        There are several reasons to want to mark something as a bad
+        share. These include:
 
hunk ./src/allmydata/mutable/retrieve.py 478
-        if len(self.shares) < k:
-            # we don't have enough shares yet
-            return self._maybe_send_more_queries(k)
-        if self._need_privkey:
-            # we got k shares, but none of them had a valid privkey. TODO:
-            # look further. Adding code to do this is a bit complicated, and
-            # I want to avoid that complication, and this should be pretty
-            # rare (k shares with bitflips in the enc_privkey but not in the
-            # data blocks). If we actually do get here, the subsequent repair
-            # will fail for lack of a privkey.
-            self.log("got k shares but still need_privkey, bummer",
-                     level=log.WEIRD, umid="MdRHPA")
+            - A connection error to the peer.
+            - A mismatched prefix (that is, a prefix that does not match
+              our local conception of the version information string).
+            - A failing block hash, salt hash, share hash, or other
+              integrity check.
 
hunk ./src/allmydata/mutable/retrieve.py 484
-        # we have enough to finish. All the shares have had their hashes
-        # checked, so if something fails at this point, we don't know how
-        # to fix it, so the download will fail.
+        This method will ensure that readers that we wish to mark bad
+        (for these reasons or other reasons) are not used for the rest
+        of the download. Additionally, it will attempt to tell the
+        remote peer (with no guarantee of success) that its share is
+        corrupt.
+        """
+        self.log("marking share %d on server %s as bad" % \
+                 (reader.shnum, reader))
+        self._remove_reader(reader)
+        self._bad_shares.add((reader.peerid, reader.shnum))
+        self._status.problems[reader.peerid] = f
+        self._last_failure = f
+        self.notify_server_corruption(reader.peerid, reader.shnum,
+                                      str(f.value))
 
hunk ./src/allmydata/mutable/retrieve.py 499
-        self._decoding = True # avoid reentrancy
-        self._status.set_status("decoding")
-        now = time.time()
-        elapsed = now - self._started
-        self._status.timings["fetch"] = elapsed
 
hunk ./src/allmydata/mutable/retrieve.py 500
-        d = defer.maybeDeferred(self._decode)
-        d.addCallback(self._decrypt, IV, self._node.get_readkey())
-        d.addBoth(self._done)
-        return d # purely for test convenience
+    def _download_current_segment(self):
+        """
+        I download, validate, decode, decrypt, and assemble the segment
+        that this Retrieve is currently responsible for downloading.
+        """
+        assert len(self._active_readers) >= self._required_shares
+        if self._current_segment < self._num_segments:
+            d = self._process_segment(self._current_segment)
+        else:
+            d = defer.succeed(None)
+        d.addCallback(self._check_for_done)
+        return d
 
hunk ./src/allmydata/mutable/retrieve.py 513
-    def _maybe_send_more_queries(self, k):
-        # we don't have enough shares yet. Should we send out more queries?
-        # There are some number of queries outstanding, each for a single
-        # share. If we can generate 'needed_shares' additional queries, we do
-        # so. If we can't, then we know this file is a goner, and we raise
-        # NotEnoughSharesError.
-        self.log(format=("_maybe_send_more_queries, have=%(have)d, k=%(k)d, "
-                         "outstanding=%(outstanding)d"),
-                 have=len(self.shares), k=k,
-                 outstanding=len(self._outstanding_queries),
-                 level=log.NOISY)
 
hunk ./src/allmydata/mutable/retrieve.py 514
-        remaining_shares = k - len(self.shares)
-        needed = remaining_shares - len(self._outstanding_queries)
-        if not needed:
-            # we have enough queries in flight already
+    def _process_segment(self, segnum):
+        """
+        I download, validate, decode, and decrypt one segment of the
+        file that this Retrieve is retrieving. This means coordinating
+        the process of getting k blocks of that file, validating them,
+        assembling them into one segment with the decoder, and then
+        decrypting them.
+        """
+        self.log("processing segment %d" % segnum)
 
hunk ./src/allmydata/mutable/retrieve.py 524
-            # TODO: but if they've been in flight for a long time, and we
-            # have reason to believe that new queries might respond faster
-            # (i.e. we've seen other queries come back faster, then consider
-            # sending out new queries. This could help with peers which have
-            # silently gone away since the servermap was updated, for which
-            # we're still waiting for the 15-minute TCP disconnect to happen.
-            self.log("enough queries are in flight, no more are needed",
-                     level=log.NOISY)
-            return
+        # TODO: The old code uses a marker. Should this code do that
+        # too? What did the Marker do?
+        assert len(self._active_readers) >= self._required_shares
+
+        # We need to ask each of our active readers for its block and
+        # salt. We will then validate those. If validation is
+        # successful, we will assemble the results into plaintext.
+        ds = []
+        for reader in self._active_readers:
+            d = reader.get_block_and_salt(segnum, queue=True)
+            d2 = self._get_needed_hashes(reader, segnum)
+            dl = defer.DeferredList([d, d2], consumeErrors=True)
+            dl.addCallback(self._validate_block, segnum, reader)
+            dl.addErrback(self._validation_or_decoding_failed, [reader])
+            ds.append(dl)
+            reader.flush()
+        dl = defer.DeferredList(ds)
+        dl.addCallback(self._maybe_decode_and_decrypt_segment, segnum)
+        return dl
 
hunk ./src/allmydata/mutable/retrieve.py 544
-        outstanding_shnums = set([shnum
-                                  for (peerid, shnum, started)
-                                  in self._outstanding_queries.values()])
-        # prefer low-numbered shares, they are more likely to be primary
-        available_shnums = sorted(self.remaining_sharemap.keys())
-        for shnum in available_shnums:
-            if shnum in outstanding_shnums:
-                # skip ones that are already in transit
-                continue
-            if shnum not in self.remaining_sharemap:
-                # no servers for that shnum. note that DictOfSets removes
-                # empty sets from the dict for us.
-                continue
-            peerid = list(self.remaining_sharemap[shnum])[0]
-            # get_data will remove that peerid from the sharemap, and add the
-            # query to self._outstanding_queries
-            self._status.set_status("Retrieving More Shares")
-            self.get_data(shnum, peerid)
-            needed -= 1
-            if not needed:
+
+    def _maybe_decode_and_decrypt_segment(self, blocks_and_salts, segnum):
+        """
+        I take the results of fetching and validating the blocks from a
+        callback chain in another method. If the results are such that
+        they tell me that validation and fetching succeeded without
+        incident, I will proceed with decoding and decryption.
+        Otherwise, I will do nothing.
+        """
+        self.log("trying to decode and decrypt segment %d" % segnum)
+        failures = False
+        for block_and_salt in blocks_and_salts:
+            if not block_and_salt[0] or block_and_salt[1] == None:
+                self.log("some validation operations failed; not proceeding")
+                failures = True
                 break
hunk ./src/allmydata/mutable/retrieve.py 560
+        if not failures:
+            self.log("everything looks ok, building segment %d" % segnum)
+            d = self._decode_blocks(blocks_and_salts, segnum)
+            d.addCallback(self._decrypt_segment)
+            d.addErrback(self._validation_or_decoding_failed,
+                         self._active_readers)
+            d.addCallback(self._set_segment)
+            return d
+        else:
+            return defer.succeed(None)
+
+
+    def _set_segment(self, segment):
+        """
+        Given a plaintext segment, I register that segment with the
+        target that is handling the file download.
+        """
+        self.log("got plaintext for segment %d" % self._current_segment)
+        self._plaintext += segment
+        self._current_segment += 1
 
hunk ./src/allmydata/mutable/retrieve.py 581
-        # at this point, we have as many outstanding queries as we can. If
-        # needed!=0 then we might not have enough to recover the file.
-        if needed:
-            format = ("ran out of peers: "
-                      "have %(have)d shares (k=%(k)d), "
-                      "%(outstanding)d queries in flight, "
-                      "need %(need)d more, "
-                      "found %(bad)d bad shares")
-            args = {"have": len(self.shares),
-                    "k": k,
-                    "outstanding": len(self._outstanding_queries),
-                    "need": needed,
-                    "bad": len(self._bad_shares),
-                    }
-            self.log(format=format,
-                     level=log.WEIRD, umid="ezTfjw", **args)
-            err = NotEnoughSharesError("%s, last failure: %s" %
-                                      (format % args, self._last_failure))
-            if self._bad_shares:
-                self.log("We found some bad shares this pass. You should "
-                         "update the servermap and try again to check "
-                         "more peers",
-                         level=log.WEIRD, umid="EFkOlA")
-                err.servermap = self.servermap
-            raise err
 
hunk ./src/allmydata/mutable/retrieve.py 582
+    def _validation_or_decoding_failed(self, f, readers):
+        """
+        I am called when a block or a salt fails to correctly validate, or when
+        the decryption or decoding operation fails for some reason.  I react to
+        this failure by notifying the remote server of corruption, and then
+        removing the remote peer from further activity.
+        """
+        assert isinstance(readers, list)
+        bad_shnums = [reader.shnum for reader in readers]
+
+        self.log("validation or decoding failed on share(s) %s, peer(s) %s "
+                 ", segment %d: %s" % \
+                 (bad_shnums, readers, self._current_segment, str(f)))
+        for reader in readers:
+            self._mark_bad_share(reader, f)
         return
 
hunk ./src/allmydata/mutable/retrieve.py 599
-    def _decode(self):
-        started = time.time()
-        (seqnum, root_hash, IV, segsize, datalength, k, N, prefix,
-         offsets_tuple) = self.verinfo
 
hunk ./src/allmydata/mutable/retrieve.py 600
-        # shares_dict is a dict mapping shnum to share data, but the codec
-        # wants two lists.
-        shareids = []; shares = []
-        for shareid, share in self.shares.items():
+    def _validate_block(self, results, segnum, reader):
+        """
+        I validate a block from one share on a remote server.
+        """
+        # Grab the part of the block hash tree that is necessary to
+        # validate this block, then generate the block hash root.
+        self.log("validating share %d for segment %d" % (reader.shnum,
+                                                             segnum))
+        # Did we fail to fetch either of the things that we were
+        # supposed to? Fail if so.
+        if not results[0][0] and results[1][0]:
+            # handled by the errback handler.
+
+            # These all get batched into one query, so the resulting
+            # failure should be the same for all of them, so we can just
+            # use the first one.
+            assert isinstance(results[0][1], failure.Failure)
+
+            f = results[0][1]
+            raise CorruptShareError(reader.peerid,
+                                    reader.shnum,
+                                    "Connection error: %s" % str(f))
+
+        block_and_salt, block_and_sharehashes = results
+        block, salt = block_and_salt[1]
+        blockhashes, sharehashes = block_and_sharehashes[1]
+
+        blockhashes = dict(enumerate(blockhashes[1]))
+        self.log("the reader gave me the following blockhashes: %s" % \
+                 blockhashes.keys())
+        self.log("the reader gave me the following sharehashes: %s" % \
+                 sharehashes[1].keys())
+        bht = self._block_hash_trees[reader.shnum]
+
+        if bht.needed_hashes(segnum, include_leaf=True):
+            try:
+                bht.set_hashes(blockhashes)
+            except (hashtree.BadHashError, hashtree.NotEnoughHashesError, \
+                    IndexError), e:
+                raise CorruptShareError(reader.peerid,
+                                        reader.shnum,
+                                        "block hash tree failure: %s" % e)
+
+        if self._version == MDMF_VERSION:
+            blockhash = hashutil.block_hash(salt + block)
+        else:
+            blockhash = hashutil.block_hash(block)
+        # If this works without an error, then validation is
+        # successful.
+        try:
+           bht.set_hashes(leaves={segnum: blockhash})
+        except (hashtree.BadHashError, hashtree.NotEnoughHashesError, \
+                IndexError), e:
+            raise CorruptShareError(reader.peerid,
+                                    reader.shnum,
+                                    "block hash tree failure: %s" % e)
+
+        # Reaching this point means that we know that this segment
+        # is correct. Now we need to check to see whether the share
+        # hash chain is also correct. 
+        # SDMF wrote share hash chains that didn't contain the
+        # leaves, which would be produced from the block hash tree.
+        # So we need to validate the block hash tree first. If
+        # successful, then bht[0] will contain the root for the
+        # shnum, which will be a leaf in the share hash tree, which
+        # will allow us to validate the rest of the tree.
+        if self.share_hash_tree.needed_hashes(reader.shnum,
+                                               include_leaf=True):
+            try:
+                self.share_hash_tree.set_hashes(hashes=sharehashes[1],
+                                            leaves={reader.shnum: bht[0]})
+            except (hashtree.BadHashError, hashtree.NotEnoughHashesError, \
+                    IndexError), e:
+                raise CorruptShareError(reader.peerid,
+                                        reader.shnum,
+                                        "corrupt hashes: %s" % e)
+
+        # TODO: Validate the salt, too.
+        self.log('share %d is valid for segment %d' % (reader.shnum,
+                                                       segnum))
+        return {reader.shnum: (block, salt)}
+
+
+    def _get_needed_hashes(self, reader, segnum):
+        """
+        I get the hashes needed to validate segnum from the reader, then return
+        to my caller when this is done.
+        """
+        bht = self._block_hash_trees[reader.shnum]
+        needed = bht.needed_hashes(segnum, include_leaf=True)
+        # The root of the block hash tree is also a leaf in the share
+        # hash tree. So we don't need to fetch it from the remote
+        # server. In the case of files with one segment, this means that
+        # we won't fetch any block hash tree from the remote server,
+        # since the hash of each share of the file is the entire block
+        # hash tree, and is a leaf in the share hash tree. This is fine,
+        # since any share corruption will be detected in the share hash
+        # tree.
+        #needed.discard(0)
+        self.log("getting blockhashes for segment %d, share %d: %s" % \
+                 (segnum, reader.shnum, str(needed)))
+        d1 = reader.get_blockhashes(needed, queue=True, force_remote=True)
+        if self.share_hash_tree.needed_hashes(reader.shnum):
+            need = self.share_hash_tree.needed_hashes(reader.shnum)
+            self.log("also need sharehashes for share %d: %s" % (reader.shnum,
+                                                                 str(need)))
+            d2 = reader.get_sharehashes(need, queue=True, force_remote=True)
+        else:
+            d2 = defer.succeed({}) # the logic in the next method
+                                   # expects a dict
+        dl = defer.DeferredList([d1, d2], consumeErrors=True)
+        return dl
+
+
+    def _decode_blocks(self, blocks_and_salts, segnum):
+        """
+        I take a list of k blocks and salts, and decode that into a
+        single encrypted segment.
+        """
+        d = {}
+        # We want to merge our dictionaries to the form 
+        # {shnum: blocks_and_salts}
+        #
+        # The dictionaries come from validate block that way, so we just
+        # need to merge them.
+        for block_and_salt in blocks_and_salts:
+            d.update(block_and_salt[1])
+
+        # All of these blocks should have the same salt; in SDMF, it is
+        # the file-wide IV, while in MDMF it is the per-segment salt. In
+        # either case, we just need to get one of them and use it.
+        #
+        # d.items()[0] is like (shnum, (block, salt))
+        # d.items()[0][1] is like (block, salt)
+        # d.items()[0][1][1] is the salt.
+        salt = d.items()[0][1][1]
+        # Next, extract just the blocks from the dict. We'll use the
+        # salt in the next step.
+        share_and_shareids = [(k, v[0]) for k, v in d.items()]
+        d2 = dict(share_and_shareids)
+        shareids = []
+        shares = []
+        for shareid, share in d2.items():
             shareids.append(shareid)
             shares.append(share)
 
hunk ./src/allmydata/mutable/retrieve.py 746
-        assert len(shareids) >= k, len(shareids)
+        assert len(shareids) >= self._required_shares, len(shareids)
         # zfec really doesn't want extra shares
hunk ./src/allmydata/mutable/retrieve.py 748
-        shareids = shareids[:k]
-        shares = shares[:k]
-
-        fec = codec.CRSDecoder()
-        fec.set_params(segsize, k, N)
-
-        self.log("params %s, we have %d shares" % ((segsize, k, N), len(shares)))
-        self.log("about to decode, shareids=%s" % (shareids,))
-        d = defer.maybeDeferred(fec.decode, shares, shareids)
-        def _done(buffers):
-            self._status.timings["decode"] = time.time() - started
-            self.log(" decode done, %d buffers" % len(buffers))
+        shareids = shareids[:self._required_shares]
+        shares = shares[:self._required_shares]
+        self.log("decoding segment %d" % segnum)
+        if segnum == self._num_segments - 1:
+            d = defer.maybeDeferred(self._tail_decoder.decode, shares, shareids)
+        else:
+            d = defer.maybeDeferred(self._segment_decoder.decode, shares, shareids)
+        def _process(buffers):
             segment = "".join(buffers)
hunk ./src/allmydata/mutable/retrieve.py 757
+            self.log(format="now decoding segment %(segnum)s of %(numsegs)s",
+                     segnum=segnum,
+                     numsegs=self._num_segments,
+                     level=log.NOISY)
             self.log(" joined length %d, datalength %d" %
hunk ./src/allmydata/mutable/retrieve.py 762
-                     (len(segment), datalength))
-            segment = segment[:datalength]
+                     (len(segment), self._data_length))
+            if segnum == self._num_segments - 1:
+                size_to_use = self._tail_data_size
+            else:
+                size_to_use = self._segment_size
+            segment = segment[:size_to_use]
             self.log(" segment len=%d" % len(segment))
hunk ./src/allmydata/mutable/retrieve.py 769
-            return segment
-        def _err(f):
-            self.log(" decode failed: %s" % f)
-            return f
-        d.addCallback(_done)
-        d.addErrback(_err)
+            return segment, salt
+        d.addCallback(_process)
         return d
 
hunk ./src/allmydata/mutable/retrieve.py 773
-    def _decrypt(self, crypttext, IV, readkey):
+
+    def _decrypt_segment(self, segment_and_salt):
+        """
+        I take a single segment and its salt, and decrypt it. I return
+        the plaintext of the segment that is in my argument.
+        """
+        segment, salt = segment_and_salt
         self._status.set_status("decrypting")
hunk ./src/allmydata/mutable/retrieve.py 781
+        self.log("decrypting segment %d" % self._current_segment)
         started = time.time()
hunk ./src/allmydata/mutable/retrieve.py 783
-        key = hashutil.ssk_readkey_data_hash(IV, readkey)
+        key = hashutil.ssk_readkey_data_hash(salt, self._node.get_readkey())
         decryptor = AES(key)
hunk ./src/allmydata/mutable/retrieve.py 785
-        plaintext = decryptor.process(crypttext)
+        plaintext = decryptor.process(segment)
         self._status.timings["decrypt"] = time.time() - started
         return plaintext
 
hunk ./src/allmydata/mutable/retrieve.py 789
-    def _done(self, res):
-        if not self._running:
+
+    def notify_server_corruption(self, peerid, shnum, reason):
+        ss = self.servermap.connections[peerid]
+        ss.callRemoteOnly("advise_corrupt_share",
+                          "mutable", self._storage_index, shnum, reason)
+
+
+    def _try_to_validate_privkey(self, enc_privkey, reader):
+
+        alleged_privkey_s = self._node._decrypt_privkey(enc_privkey)
+        alleged_writekey = hashutil.ssk_writekey_hash(alleged_privkey_s)
+        if alleged_writekey != self._node.get_writekey():
+            self.log("invalid privkey from %s shnum %d" %
+                     (reader, reader.shnum),
+                     level=log.WEIRD, umid="YIw4tA")
             return
hunk ./src/allmydata/mutable/retrieve.py 805
-        self._running = False
-        self._status.set_active(False)
-        self._status.timings["total"] = time.time() - self._started
-        # res is either the new contents, or a Failure
-        if isinstance(res, failure.Failure):
-            self.log("Retrieve done, with failure", failure=res,
-                     level=log.UNUSUAL)
-            self._status.set_status("Failed")
-        else:
-            self.log("Retrieve done, success!")
-            self._status.set_status("Finished")
-            self._status.set_progress(1.0)
-            # remember the encoding parameters, use them again next time
-            (seqnum, root_hash, IV, segsize, datalength, k, N, prefix,
-             offsets_tuple) = self.verinfo
-            self._node._populate_required_shares(k)
-            self._node._populate_total_shares(N)
-        eventually(self._done_deferred.callback, res)
 
hunk ./src/allmydata/mutable/retrieve.py 806
+        # it's good
+        self.log("got valid privkey from shnum %d on reader %s" %
+                 (reader.shnum, reader))
+        privkey = rsa.create_signing_key_from_string(alleged_privkey_s)
+        self._node._populate_encprivkey(enc_privkey)
+        self._node._populate_privkey(privkey)
+        self._need_privkey = False
+
+
+    def _check_for_done(self, res):
+        """
+        I check to see if this Retrieve object has successfully finished
+        its work.
+
+        I can exit in the following ways:
+            - If there are no more segments to download, then I exit by
+              causing self._done_deferred to fire with the plaintext
+              content requested by the caller.
+            - If there are still segments to be downloaded, and there
+              are enough active readers (readers which have not broken
+              and have not given us corrupt data) to continue
+              downloading, I send control back to
+              _download_current_segment.
+            - If there are still segments to be downloaded but there are
+              not enough active peers to download them, I ask
+              _add_active_peers to add more peers. If it is successful,
+              it will call _download_current_segment. If there are not
+              enough peers to retrieve the file, then that will cause
+              _done_deferred to errback.
+        """
+        self.log("checking for doneness")
+        if self._current_segment == self._num_segments:
+            # No more segments to download, we're done.
+            self.log("got plaintext, done")
+            return self._done()
+
+        if len(self._active_readers) >= self._required_shares:
+            # More segments to download, but we have enough good peers
+            # in self._active_readers that we can do that without issue,
+            # so go nab the next segment.
+            self.log("not done yet: on segment %d of %d" % \
+                     (self._current_segment + 1, self._num_segments))
+            return self._download_current_segment()
+
+        self.log("not done yet: on segment %d of %d, need to add peers" % \
+                 (self._current_segment + 1, self._num_segments))
+        return self._add_active_peers()
+
+
+    def _done(self):
+        """
+        I am called by _check_for_done when the download process has
+        finished successfully. After making some useful logging
+        statements, I return the decrypted contents to the owner of this
+        Retrieve object through self._done_deferred.
+        """
+        eventually(self._done_deferred.callback, self._plaintext)
+
+
+    def _failed(self):
+        """
+        I am called by _add_active_peers when there are not enough
+        active peers left to complete the download. After making some
+        useful logging statements, I return an exception to that effect
+        to the caller of this Retrieve object through
+        self._done_deferred.
+        """
+        format = ("ran out of peers: "
+                  "have %(have)d of %(total)d segments "
+                  "found %(bad)d bad shares "
+                  "encoding %(k)d-of-%(n)d")
+        args = {"have": self._current_segment,
+                "total": self._num_segments,
+                "k": self._required_shares,
+                "n": self._total_shares,
+                "bad": len(self._bad_shares)}
+        e = NotEnoughSharesError("%s, last failure: %s" % (format % args,
+                                                        str(self._last_failure)))
+        f = failure.Failure(e)
+        eventually(self._done_deferred.callback, f)
hunk ./src/allmydata/test/test_mutable.py 12
 from allmydata.util.hashutil import tagged_hash, ssk_writekey_hash, \
      ssk_pubkey_fingerprint_hash
 from allmydata.interfaces import IRepairResults, ICheckAndRepairResults, \
-     NotEnoughSharesError
+     NotEnoughSharesError, SDMF_VERSION, MDMF_VERSION
 from allmydata.monitor import Monitor
 from allmydata.test.common import ShouldFailMixin
 from allmydata.test.no_network import GridTestMixin
hunk ./src/allmydata/test/test_mutable.py 28
 from allmydata.mutable.retrieve import Retrieve
 from allmydata.mutable.publish import Publish
 from allmydata.mutable.servermap import ServerMap, ServermapUpdater
-from allmydata.mutable.layout import unpack_header, unpack_share
+from allmydata.mutable.layout import unpack_header, unpack_share, \
+                                     MDMFSlotReadProxy
 from allmydata.mutable.repairer import MustForceRepairError
 
 import allmydata.test.common_util as testutil
hunk ./src/allmydata/test/test_mutable.py 104
         d = fireEventually()
         d.addCallback(lambda res: _call())
         return d
+
     def callRemoteOnly(self, methname, *args, **kwargs):
         d = self.callRemote(methname, *args, **kwargs)
         d.addBoth(lambda ignore: None)
hunk ./src/allmydata/test/test_mutable.py 163
 def corrupt(res, s, offset, shnums_to_corrupt=None, offset_offset=0):
     # if shnums_to_corrupt is None, corrupt all shares. Otherwise it is a
     # list of shnums to corrupt.
+    ds = []
     for peerid in s._peers:
         shares = s._peers[peerid]
         for shnum in shares:
hunk ./src/allmydata/test/test_mutable.py 190
                 else:
                     offset1 = offset
                     offset2 = 0
-                if offset1 == "pubkey":
+                if offset1 == "pubkey" and IV:
                     real_offset = 107
hunk ./src/allmydata/test/test_mutable.py 192
+                elif offset1 == "share_data" and not IV:
+                    real_offset = 104
                 elif offset1 in o:
                     real_offset = o[offset1]
                 else:
hunk ./src/allmydata/test/test_mutable.py 327
         d.addCallback(_created)
         return d
 
+
+    def test_upload_and_download_mdmf(self):
+        d = self.nodemaker.create_mutable_file(version=MDMF_VERSION)
+        def _created(n):
+            d = defer.succeed(None)
+            d.addCallback(lambda ignored:
+                n.get_servermap(MODE_READ))
+            def _then(servermap):
+                dumped = servermap.dump(StringIO())
+                self.failUnlessIn("3-of-10", dumped.getvalue())
+            d.addCallback(_then)
+            # Now overwrite the contents with some new contents. We want 
+            # to make them big enough to force the file to be uploaded
+            # in more than one segment.
+            big_contents = "contents1" * 100000 # about 900 KiB
+            d.addCallback(lambda ignored:
+                n.overwrite(big_contents))
+            d.addCallback(lambda ignored:
+                n.download_best_version())
+            d.addCallback(lambda data:
+                self.failUnlessEqual(data, big_contents))
+            # Overwrite the contents again with some new contents. As
+            # before, they need to be big enough to force multiple
+            # segments, so that we make the downloader deal with
+            # multiple segments.
+            bigger_contents = "contents2" * 1000000 # about 9MiB 
+            d.addCallback(lambda ignored:
+                n.overwrite(bigger_contents))
+            d.addCallback(lambda ignored:
+                n.download_best_version())
+            d.addCallback(lambda data:
+                self.failUnlessEqual(data, bigger_contents))
+            return d
+        d.addCallback(_created)
+        return d
+
+
     def test_create_with_initial_contents(self):
         d = self.nodemaker.create_mutable_file("contents 1")
         def _created(n):
hunk ./src/allmydata/test/test_mutable.py 1147
 
 
     def _test_corrupt_all(self, offset, substring,
-                          should_succeed=False, corrupt_early=True,
-                          failure_checker=None):
+                          should_succeed=False,
+                          corrupt_early=True,
+                          failure_checker=None,
+                          fetch_privkey=False):
         d = defer.succeed(None)
         if corrupt_early:
             d.addCallback(corrupt, self._storage, offset)
hunk ./src/allmydata/test/test_mutable.py 1167
                     self.failUnlessIn(substring, "".join(allproblems))
                 return servermap
             if should_succeed:
-                d1 = self._fn.download_version(servermap, ver)
+                d1 = self._fn.download_version(servermap, ver,
+                                               fetch_privkey)
                 d1.addCallback(lambda new_contents:
                                self.failUnlessEqual(new_contents, self.CONTENTS))
             else:
hunk ./src/allmydata/test/test_mutable.py 1175
                 d1 = self.shouldFail(NotEnoughSharesError,
                                      "_corrupt_all(offset=%s)" % (offset,),
                                      substring,
-                                     self._fn.download_version, servermap, ver)
+                                     self._fn.download_version, servermap,
+                                                                ver,
+                                                                fetch_privkey)
             if failure_checker:
                 d1.addCallback(failure_checker)
             d1.addCallback(lambda res: servermap)
hunk ./src/allmydata/test/test_mutable.py 1186
         return d
 
     def test_corrupt_all_verbyte(self):
-        # when the version byte is not 0, we hit an UnknownVersionError error
-        # in unpack_share().
+        # when the version byte is not 0 or 1, we hit an UnknownVersionError
+        # error in unpack_share().
         d = self._test_corrupt_all(0, "UnknownVersionError")
         def _check_servermap(servermap):
             # and the dump should mention the problems
hunk ./src/allmydata/test/test_mutable.py 1193
             s = StringIO()
             dump = servermap.dump(s).getvalue()
-            self.failUnless("10 PROBLEMS" in dump, dump)
+            self.failUnless("30 PROBLEMS" in dump, dump)
         d.addCallback(_check_servermap)
         return d
 
hunk ./src/allmydata/test/test_mutable.py 1263
         return self._test_corrupt_all("enc_privkey", None, should_succeed=True)
 
 
+    def test_corrupt_all_encprivkey_late(self):
+        # this should work for the same reason as above, but we corrupt 
+        # after the servermap update to exercise the error handling
+        # code.
+        # We need to remove the privkey from the node, or the retrieve
+        # process won't know to update it.
+        self._fn._privkey = None
+        return self._test_corrupt_all("enc_privkey",
+                                      None, # this shouldn't fail
+                                      should_succeed=True,
+                                      corrupt_early=False,
+                                      fetch_privkey=True)
+
+
     def test_corrupt_all_seqnum_late(self):
         # corrupting the seqnum between mapupdate and retrieve should result
         # in NotEnoughSharesError, since each share will look invalid
hunk ./src/allmydata/test/test_mutable.py 1283
         def _check(res):
             f = res[0]
             self.failUnless(f.check(NotEnoughSharesError))
-            self.failUnless("someone wrote to the data since we read the servermap" in str(f))
+            self.failUnless("uncoordinated write" in str(f))
         return self._test_corrupt_all(1, "ran out of peers",
                                       corrupt_early=False,
                                       failure_checker=_check)
hunk ./src/allmydata/test/test_mutable.py 1333
                       self.failUnlessEqual(new_contents, self.CONTENTS))
         return d
 
-    def test_corrupt_some(self):
-        # corrupt the data of first five shares (so the servermap thinks
-        # they're good but retrieve marks them as bad), so that the
-        # MODE_READ set of 6 will be insufficient, forcing node.download to
-        # retry with more servers.
-        corrupt(None, self._storage, "share_data", range(5))
-        d = self.make_servermap()
+
+    def _test_corrupt_some(self, offset, mdmf=False):
+        if mdmf:
+            d = self.publish_mdmf()
+        else:
+            d = defer.succeed(None)
+        d.addCallback(lambda ignored:
+            corrupt(None, self._storage, offset, range(5)))
+        d.addCallback(lambda ignored:
+            self.make_servermap())
         def _do_retrieve(servermap):
             ver = servermap.best_recoverable_version()
             self.failUnless(ver)
hunk ./src/allmydata/test/test_mutable.py 1349
             return self._fn.download_best_version()
         d.addCallback(_do_retrieve)
         d.addCallback(lambda new_contents:
-                      self.failUnlessEqual(new_contents, self.CONTENTS))
+            self.failUnlessEqual(new_contents, self.CONTENTS))
         return d
 
hunk ./src/allmydata/test/test_mutable.py 1352
+
+    def test_corrupt_some(self):
+        # corrupt the data of first five shares (so the servermap thinks
+        # they're good but retrieve marks them as bad), so that the
+        # MODE_READ set of 6 will be insufficient, forcing node.download to
+        # retry with more servers.
+        return self._test_corrupt_some("share_data")
+
+
     def test_download_fails(self):
         d = corrupt(None, self._storage, "signature")
         d.addCallback(lambda ignored:
hunk ./src/allmydata/test/test_mutable.py 1366
             self.shouldFail(UnrecoverableFileError, "test_download_anyway",
                             "no recoverable versions",
-                            self._fn.download_best_version)
+                            self._fn.download_best_version))
         return d
 
 
hunk ./src/allmydata/test/test_mutable.py 1370
+
+    def test_corrupt_mdmf_block_hash_tree(self):
+        d = self.publish_mdmf()
+        d.addCallback(lambda ignored:
+            self._test_corrupt_all(("block_hash_tree", 12 * 32),
+                                   "block hash tree failure",
+                                   corrupt_early=False,
+                                   should_succeed=False))
+        return d
+
+
+    def test_corrupt_mdmf_block_hash_tree_late(self):
+        d = self.publish_mdmf()
+        d.addCallback(lambda ignored:
+            self._test_corrupt_all(("block_hash_tree", 12 * 32),
+                                   "block hash tree failure",
+                                   corrupt_early=True,
+                                   should_succeed=False))
+        return d
+
+
+    def test_corrupt_mdmf_share_data(self):
+        d = self.publish_mdmf()
+        d.addCallback(lambda ignored:
+            # TODO: Find out what the block size is and corrupt a
+            # specific block, rather than just guessing.
+            self._test_corrupt_all(("share_data", 12 * 40),
+                                    "block hash tree failure",
+                                    corrupt_early=True,
+                                    should_succeed=False))
+        return d
+
+
+    def test_corrupt_some_mdmf(self):
+        return self._test_corrupt_some(("share_data", 12 * 40),
+                                       mdmf=True)
+
+
 class CheckerMixin:
     def check_good(self, r, where):
         self.failUnless(r.is_healthy(), where)
hunk ./src/allmydata/test/test_mutable.py 2116
             d.addCallback(lambda res:
                           self.shouldFail(NotEnoughSharesError,
                                           "test_retrieve_surprise",
-                                          "ran out of peers: have 0 shares (k=3)",
+                                          "ran out of peers: have 0 of 1",
                                           n.download_version,
                                           self.old_map,
                                           self.old_map.best_recoverable_version(),
hunk ./src/allmydata/test/test_mutable.py 2125
         d.addCallback(_created)
         return d
 
+
     def test_unexpected_shares(self):
         # upload the file, take a servermap, shut down one of the servers,
         # upload it again (causing shares to appear on a new server), then
hunk ./src/allmydata/test/test_mutable.py 2329
         self.basedir = "mutable/Problems/test_privkey_query_missing"
         self.set_up_grid(num_servers=20)
         nm = self.g.clients[0].nodemaker
-        LARGE = "These are Larger contents" * 2000 # about 50KB
+        LARGE = "These are Larger contents" * 2000 # about 50KiB
         nm._node_cache = DevNullDictionary() # disable the nodecache
 
         d = nm.create_mutable_file(LARGE)
hunk ./src/allmydata/test/test_mutable.py 2342
         d.addCallback(_created)
         d.addCallback(lambda res: self.n2.get_servermap(MODE_WRITE))
         return d
+
+
+    def test_block_and_hash_query_error(self):
+        # This tests for what happens when a query to a remote server
+        # fails in either the hash validation step or the block getting
+        # step (because of batching, this is the same actual query).
+        # We need to have the storage server persist up until the point
+        # that its prefix is validated, then suddenly die. This
+        # exercises some exception handling code in Retrieve.
+        self.basedir = "mutable/Problems/test_block_and_hash_query_error"
+        self.set_up_grid(num_servers=20)
+        nm = self.g.clients[0].nodemaker
+        CONTENTS = "contents" * 2000
+        d = nm.create_mutable_file(CONTENTS)
+        def _created(node):
+            self._node = node
+        d.addCallback(_created)
+        d.addCallback(lambda ignored:
+            self._node.get_servermap(MODE_READ))
+        def _then(servermap):
+            # we have our servermap. Now we set up the servers like the
+            # tests above -- the first one that gets a read call should
+            # start throwing errors, but only after returning its prefix
+            # for validation. Since we'll download without fetching the
+            # private key, the next query to the remote server will be
+            # for either a block and salt or for hashes, either of which
+            # will exercise the error handling code.
+            killer = FirstServerGetsKilled()
+            for (serverid, ss) in nm.storage_broker.get_all_servers():
+                ss.post_call_notifier = killer.notify
+            ver = servermap.best_recoverable_version()
+            assert ver
+            return self._node.download_version(servermap, ver)
+        d.addCallback(_then)
+        d.addCallback(lambda data:
+            self.failUnlessEqual(data, CONTENTS))
+        return d
}

Context:

[docs: about.html link to home page early on, and be decentralized storage instead of cloud storage this time around
zooko@zooko.com**20100619065318
 Ignore-this: dc6db03f696e5b6d2848699e754d8053
] 
[docs: update about.html, especially to have a non-broken link to quickstart.html, and also to comment out the broken links to "for Paranoids" and "for Corporates"
zooko@zooko.com**20100619065124
 Ignore-this: e292c7f51c337a84ebfeb366fbd24d6c
] 
[TAG allmydata-tahoe-1.7.0
zooko@zooko.com**20100619052631
 Ignore-this: d21e27afe6d85e2e3ba6a3292ba2be1
] 
Patch bundle hash:
7595cd3e7a002ae1c9c0f6ba7e21f39a7f4477e1