From: Zooko O'Whielacronx Date: Wed, 31 Dec 2008 22:42:26 +0000 (-0700) Subject: immutable: storage servers accept any size shares now X-Git-Url: https://git.rkrishnan.org/components/com_hotproperty/simplejson/running.html?a=commitdiff_plain;h=6c4019ec33e7a253695bf7314dabfb39343b0ddc;p=tahoe-lafs%2Ftahoe-lafs.git immutable: storage servers accept any size shares now Nathan Wilcox observed that the storage server can rely on the size of the share file combined with the count of leases to unambiguously identify the location of the leases. This means that it can hold any size share data, even though the field nominally used to hold the size of the share data is only 32 bits wide. With this patch, the storage server still writes the "size of the share data" field (just in case the server gets downgraded to an earlier version which requires that field, or the share file gets moved to another server which is of an earlier vintage), but it doesn't use it. Also, with this patch, the server no longer rejects requests to write shares which are >= 2^32 bytes in size, and it no longer rejects attempts to read such shares. This fixes http://allmydata.org/trac/tahoe/ticket/346 (increase share-size field to 8 bytes, remove 12GiB filesize limit), although there remains open a question of how clients know that a given server can handle large shares (by using the new versioning scheme, probably). Note that share size is also limited by another factor -- how big of a file we can store on the local filesystem on the server. Currently allmydata.com typically uses ext3 and I think we typically have block size = 4 KiB, which means that the largest file is about 2 TiB. Also, the hard drives themselves are only 1 TB, so the largest share is definitely slightly less than 1 TB, which means (when K == 3), the largest file is less than 3 TB. This patch also refactors the creation of new sharefiles so that only a single fopen() is used. This patch also helps with the unit-testing of repairer, since formerly it was unclear what repairer should expect to find if the "share data size" field was corrupted (some corruptions would have no effect, others would cause failure to download). Now it is clear that repairer is not required to notice if this field is corrupted since it has no effect on download. :-) --- diff --git a/src/allmydata/storage.py b/src/allmydata/storage.py index cc142382..bff0431b 100644 --- a/src/allmydata/storage.py +++ b/src/allmydata/storage.py @@ -33,15 +33,21 @@ NUM_RE=re.compile("^[0-9]+$") # The share file has the following layout: # 0x00: share file version number, four bytes, current version is 1 -# 0x04: share data length, four bytes big-endian = A -# 0x08: number of leases, four bytes big-endian +# 0x04: share data length, four bytes big-endian = A # See Footnote 1 below. +# 0x08: number of leases, four bytes big-endian (unused) # 0x0c: beginning of share data (see immutable.layout.WriteBucketProxy) -# A+0x0c = B: first lease. Lease format is: -# B+0x00: owner number, 4 bytes big-endian, 0 is reserved for no-owner -# B+0x04: renew secret, 32 bytes (SHA256) -# B+0x24: cancel secret, 32 bytes (SHA256) -# B+0x44: expiration time, 4 bytes big-endian seconds-since-epoch -# B+0x48: next lease, or end of record +# A+0x0c = B: first lease. Lease format is: (unused) XXX +# B+0x00: owner number, 4 bytes big-endian, 0 is reserved for no-owner (unused) XXX +# B+0x04: renew secret, 32 bytes (SHA256) (unused) XXX +# B+0x24: cancel secret, 32 bytes (SHA256) (unused) XXX +# B+0x44: expiration time, 4 bytes big-endian seconds-since-epoch (unused) XXX +# B+0x48: next lease, or end of record (unused) XXX + +# Footnote 1: as of Tahoe v1.3.0 this field is not used by storage servers, but it is still +# filled in by storage servers in case the storage server software gets downgraded from >= Tahoe +# v1.3.0 to < Tahoe v1.3.0, or the share file is moved from one storage server to another. The +# value stored in this field is truncated, so If the actual share data length is >= 2**32, then +# the value stored in this field will be the actual share data length modulo 2**32. def si_b2a(storageindex): return base32.b2a(storageindex) @@ -95,22 +101,40 @@ class LeaseInfo: class ShareFile: LEASE_SIZE = struct.calcsize(">L32s32sL") - def __init__(self, filename): + def __init__(self, filename, max_size=None, create=False): + """ If max_size is not None then I won't allow more than max_size to be written to me. If create=True and max_size must not be None. """ + precondition((max_size is not None) or (not create), max_size, create) self.home = filename - f = open(self.home, 'rb') - (version, size, num_leases) = struct.unpack(">LLL", f.read(0xc)) - assert version == 1, version - self._size = size - self._num_leases = num_leases + self._max_size = max_size + if create: + # touch the file, so later callers will see that we're working on it. + # Also construct the metadata. + assert not os.path.exists(self.home) + fileutil.make_dirs(os.path.dirname(self.home)) + f = open(self.home, 'wb') + # The second field -- share data length -- is no longer used as of Tahoe v1.3.0, but + # we continue to write it in there in case someone downgrades a storage server from + # >= Tahoe-1.3.0 to < Tahoe-1.3.0, or moves a share file from one server to another, + # etc. + f.write(struct.pack(">LLL", 1, max_size % 4294967295, 0)) + f.close() + self._lease_offset = max_size + 0x0c + self._num_leases = 0 + else: + f = open(self.home, 'rb') + filesize = os.path.getsize(self.home) + (version, unused, num_leases) = struct.unpack(">LLL", f.read(0xc)) + f.close() + assert version == 1, version + self._num_leases = num_leases + self._lease_offset = filesize - (num_leases * self.LEASE_SIZE) self._data_offset = 0xc - self._lease_offset = 0xc + self._size def unlink(self): os.unlink(self.home) def read_share_data(self, offset, length): precondition(offset >= 0) - precondition(offset+length <= self._size) f = open(self.home, 'rb') f.seek(self._data_offset+offset) return f.read(length) @@ -118,7 +142,8 @@ class ShareFile: def write_share_data(self, offset, data): length = len(data) precondition(offset >= 0, offset) - precondition(offset+length <= self._size, offset+length, self._size) + if self._max_size is not None and offset+length > self._max_size: + raise DataTooLargeError(self._max_size, offset, length) f = open(self.home, 'rb+') real_offset = self._data_offset+offset f.seek(real_offset) @@ -148,7 +173,7 @@ class ShareFile: """Yields (ownernum, renew_secret, cancel_secret, expiration_time) for all leases.""" f = open(self.home, 'rb') - (version, size, num_leases) = struct.unpack(">LLL", f.read(0xc)) + (version, unused, num_leases) = struct.unpack(">LLL", f.read(0xc)) f.seek(self._lease_offset) for i in range(num_leases): data = f.read(self.LEASE_SIZE) @@ -221,30 +246,22 @@ class ShareFile: class BucketWriter(Referenceable): implements(RIBucketWriter) - def __init__(self, ss, incominghome, finalhome, size, lease_info, canary): + def __init__(self, ss, incominghome, finalhome, max_size, lease_info, canary): self.ss = ss self.incominghome = incominghome self.finalhome = finalhome - self._size = size + self._max_size = max_size # don't allow the client to write more than this self._canary = canary self._disconnect_marker = canary.notifyOnDisconnect(self._disconnected) self.closed = False self.throw_out_all_data = False - # touch the file, so later callers will see that we're working on it. - assert not os.path.exists(incominghome) - fileutil.make_dirs(os.path.dirname(incominghome)) - # Also construct the metadata. - f = open(incominghome, 'wb') - precondition(size < 2**32) # v1 container format: 4-byte size field - f.write(struct.pack(">LLL", 1, size, 0)) - f.close() - self._sharefile = ShareFile(incominghome) + self._sharefile = ShareFile(incominghome, create=True, max_size=max_size) # also, add our lease to the file now, so that other ones can be # added by simultaneous uploaders self._sharefile.add_lease(lease_info) def allocated_size(self): - return self._size + return self._max_size def remote_write(self, offset, data): start = time.time() @@ -896,7 +913,7 @@ class StorageServer(service.MultiService, Referenceable): stats["storage_server.disk_used"] = disk_used stats["storage_server.disk_avail"] = disk_avail except AttributeError: - # os.statvfs is only available on unix + # os.statvfs is available only on unix pass stats["storage_server.accepting_immutable_shares"] = int(writeable) return stats @@ -953,7 +970,7 @@ class StorageServer(service.MultiService, Referenceable): renew_secret, cancel_secret, expire_time, self.my_nodeid) - space_per_bucket = allocated_size + max_space_per_bucket = allocated_size remaining_space = self.get_available_space() limited = remaining_space is not None @@ -986,16 +1003,16 @@ class StorageServer(service.MultiService, Referenceable): # occurs while the first is still in progress, the second # uploader will use different storage servers. pass - elif (not limited) or (remaining_space >= space_per_bucket): + elif (not limited) or (remaining_space >= max_space_per_bucket): # ok! we need to create the new share file. bw = BucketWriter(self, incominghome, finalhome, - space_per_bucket, lease_info, canary) + max_space_per_bucket, lease_info, canary) if self.no_storage: bw.throw_out_all_data = True bucketwriters[shnum] = bw self._active_writers[bw] = 1 if limited: - remaining_space -= space_per_bucket + remaining_space -= max_space_per_bucket else: # bummer! not enough space to accept this bucket pass diff --git a/src/allmydata/test/test_storage.py b/src/allmydata/test/test_storage.py index 0f153ad4..84f5a60c 100644 --- a/src/allmydata/test/test_storage.py +++ b/src/allmydata/test/test_storage.py @@ -255,6 +255,22 @@ class Server(unittest.TestCase): renew_secret, cancel_secret, sharenums, size, canary) + def test_large_share(self): + ss = self.create("test_large_share") + + already,writers = self.allocate(ss, "allocate", [0,1,2], 2**32+2) + self.failUnlessEqual(already, set()) + self.failUnlessEqual(set(writers.keys()), set([0,1,2])) + + shnum, bucket = writers.items()[0] + # This test is going to hammer your filesystem if it doesn't make a sparse file for this. :-( + bucket.remote_write(2**32, "ab") + bucket.remote_close() + + readers = ss.remote_get_buckets("allocate") + reader = readers[shnum] + self.failUnlessEqual(reader.remote_read(2**32, 2), "ab") + def test_dont_overfill_dirs(self): """ This test asserts that if you add a second share whose storage index