1 import os, re, weakref, struct, time
3 from foolscap import Referenceable
4 from twisted.application import service
6 from zope.interface import implements
7 from allmydata.interfaces import RIStorageServer, IStatsProducer
8 from allmydata.util import base32, fileutil, log, time_format
9 import allmydata # for __full_version__
11 from allmydata.storage.common import si_b2a, si_a2b, storage_index_to_dir
12 _pyflakes_hush = [si_b2a, si_a2b, storage_index_to_dir] # re-exported
13 from allmydata.storage.lease import LeaseInfo
14 from allmydata.storage.mutable import MutableShareFile, EmptyShare, \
15 create_mutable_sharefile
16 from allmydata.storage.immutable import ShareFile, BucketWriter, BucketReader
17 from allmydata.storage.crawler import BucketCountingCrawler
20 # storage/shares/incoming
21 # incoming/ holds temp dirs named $START/$STORAGEINDEX/$SHARENUM which will
22 # be moved to storage/shares/$START/$STORAGEINDEX/$SHARENUM upon success
23 # storage/shares/$START/$STORAGEINDEX
24 # storage/shares/$START/$STORAGEINDEX/$SHARENUM
26 # Where "$START" denotes the first 10 bits worth of $STORAGEINDEX (that's 2
29 # $SHARENUM matches this regex:
30 NUM_RE=re.compile("^[0-9]+$")
34 class StorageServer(service.MultiService, Referenceable):
35 implements(RIStorageServer, IStatsProducer)
38 def __init__(self, storedir, nodeid, reserved_space=0,
39 discard_storage=False, readonly_storage=False,
41 service.MultiService.__init__(self)
42 assert isinstance(nodeid, str)
43 assert len(nodeid) == 20
44 self.my_nodeid = nodeid
45 self.storedir = storedir
46 sharedir = os.path.join(storedir, "shares")
47 fileutil.make_dirs(sharedir)
48 self.sharedir = sharedir
49 # we don't actually create the corruption-advisory dir until necessary
50 self.corruption_advisory_dir = os.path.join(storedir,
51 "corruption-advisories")
52 self.reserved_space = int(reserved_space)
53 self.no_storage = discard_storage
54 self.readonly_storage = readonly_storage
55 self.stats_provider = stats_provider
56 if self.stats_provider:
57 self.stats_provider.register_producer(self)
58 self.incomingdir = os.path.join(sharedir, 'incoming')
59 self._clean_incomplete()
60 fileutil.make_dirs(self.incomingdir)
61 self._active_writers = weakref.WeakKeyDictionary()
62 lp = log.msg("StorageServer created", facility="tahoe.storage")
65 if self.get_available_space() is None:
66 log.msg("warning: [storage]reserved_space= is set, but this platform does not support statvfs(2), so this reservation cannot be honored",
67 umin="0wZ27w", level=log.UNUSUAL)
69 self.latencies = {"allocate": [], # immutable
74 "writev": [], # mutable
76 "add-lease": [], # both
81 statefile = os.path.join(storedir, "bucket_counter.state")
82 self.bucket_counter = BucketCountingCrawler(self, statefile)
83 self.bucket_counter.setServiceParent(self)
85 def count(self, name, delta=1):
86 if self.stats_provider:
87 self.stats_provider.count("storage_server." + name, delta)
89 def add_latency(self, category, latency):
90 a = self.latencies[category]
93 self.latencies[category] = a[-1000:]
95 def get_latencies(self):
96 """Return a dict, indexed by category, that contains a dict of
97 latency numbers for each category. Each dict will contain the
98 following keys: mean, 01_0_percentile, 10_0_percentile,
99 50_0_percentile (median), 90_0_percentile, 95_0_percentile,
100 99_0_percentile, 99_9_percentile. If no samples have been collected
101 for the given category, then that category name will not be present
102 in the return value."""
103 # note that Amazon's Dynamo paper says they use 99.9% percentile.
105 for category in self.latencies:
106 if not self.latencies[category]:
109 samples = self.latencies[category][:]
112 stats["mean"] = sum(samples) / count
113 stats["01_0_percentile"] = samples[int(0.01 * count)]
114 stats["10_0_percentile"] = samples[int(0.1 * count)]
115 stats["50_0_percentile"] = samples[int(0.5 * count)]
116 stats["90_0_percentile"] = samples[int(0.9 * count)]
117 stats["95_0_percentile"] = samples[int(0.95 * count)]
118 stats["99_0_percentile"] = samples[int(0.99 * count)]
119 stats["99_9_percentile"] = samples[int(0.999 * count)]
120 output[category] = stats
123 def log(self, *args, **kwargs):
124 if "facility" not in kwargs:
125 kwargs["facility"] = "tahoe.storage"
126 return log.msg(*args, **kwargs)
128 def _clean_incomplete(self):
129 fileutil.rm_dir(self.incomingdir)
131 def do_statvfs(self):
132 return os.statvfs(self.storedir)
135 # remember: RIStatsProvider requires that our return dict
136 # contains numeric values.
137 stats = { 'storage_server.allocated': self.allocated_size(), }
138 stats["storage_server.reserved_space"] = self.reserved_space
139 for category,ld in self.get_latencies().items():
140 for name,v in ld.items():
141 stats['storage_server.latencies.%s.%s' % (category, name)] = v
143 if self.readonly_storage:
146 s = self.do_statvfs()
147 disk_total = s.f_bsize * s.f_blocks
148 disk_used = s.f_bsize * (s.f_blocks - s.f_bfree)
149 # spacetime predictors should look at the slope of disk_used.
150 disk_free_for_root = s.f_bsize * s.f_bfree
151 disk_free_for_nonroot = s.f_bsize * s.f_bavail
153 # include our local policy here: if we stop accepting shares when
154 # the available space drops below 1GB, then include that fact in
156 disk_avail = disk_free_for_nonroot - self.reserved_space
157 disk_avail = max(disk_avail, 0)
158 if self.readonly_storage:
163 # spacetime predictors should use disk_avail / (d(disk_used)/dt)
164 stats["storage_server.disk_total"] = disk_total
165 stats["storage_server.disk_used"] = disk_used
166 stats["storage_server.disk_free_for_root"] = disk_free_for_root
167 stats["storage_server.disk_free_for_nonroot"] = disk_free_for_nonroot
168 stats["storage_server.disk_avail"] = disk_avail
169 except AttributeError:
170 # os.statvfs is available only on unix
172 stats["storage_server.accepting_immutable_shares"] = int(writeable)
173 s = self.bucket_counter.get_state()
174 bucket_count = s.get("last-complete-bucket-count")
176 stats["storage_server.total_bucket_count"] = bucket_count
180 def stat_disk(self, d):
182 # s.f_bavail: available to non-root users
183 disk_avail = s.f_bsize * s.f_bavail
186 def get_available_space(self):
187 # returns None if it cannot be measured (windows)
189 disk_avail = self.stat_disk(self.storedir)
190 disk_avail -= self.reserved_space
191 except AttributeError:
193 if self.readonly_storage:
197 def allocated_size(self):
199 for bw in self._active_writers:
200 space += bw.allocated_size()
203 def remote_get_version(self):
204 remaining_space = self.get_available_space()
205 if remaining_space is None:
206 # we're on a platform that doesn't have 'df', so make a vague
208 remaining_space = 2**64
209 version = { "http://allmydata.org/tahoe/protocols/storage/v1" :
210 { "maximum-immutable-share-size": remaining_space,
211 "tolerates-immutable-read-overrun": True,
212 "delete-mutable-shares-with-zero-length-writev": True,
214 "application-version": str(allmydata.__full_version__),
218 def remote_allocate_buckets(self, storage_index,
219 renew_secret, cancel_secret,
220 sharenums, allocated_size,
221 canary, owner_num=0):
222 # owner_num is not for clients to set, but rather it should be
223 # curried into the PersonalStorageServer instance that is dedicated
224 # to a particular owner.
226 self.count("allocate")
228 bucketwriters = {} # k: shnum, v: BucketWriter
229 si_dir = storage_index_to_dir(storage_index)
230 si_s = si_b2a(storage_index)
232 log.msg("storage: allocate_buckets %s" % si_s)
234 # in this implementation, the lease information (including secrets)
235 # goes into the share files themselves. It could also be put into a
236 # separate database. Note that the lease should not be added until
237 # the BucketWriter has been closed.
238 expire_time = time.time() + 31*24*60*60
239 lease_info = LeaseInfo(owner_num,
240 renew_secret, cancel_secret,
241 expire_time, self.my_nodeid)
243 max_space_per_bucket = allocated_size
245 remaining_space = self.get_available_space()
246 limited = remaining_space is not None
248 # this is a bit conservative, since some of this allocated_size()
249 # has already been written to disk, where it will show up in
250 # get_available_space.
251 remaining_space -= self.allocated_size()
253 # fill alreadygot with all shares that we have, not just the ones
254 # they asked about: this will save them a lot of work. Add or update
255 # leases for all of them: if they want us to hold shares for this
256 # file, they'll want us to hold leases for this file.
257 for (shnum, fn) in self._get_bucket_shares(storage_index):
258 alreadygot.add(shnum)
260 sf.add_or_renew_lease(lease_info)
262 # self.readonly_storage causes remaining_space=0
264 for shnum in sharenums:
265 incominghome = os.path.join(self.incomingdir, si_dir, "%d" % shnum)
266 finalhome = os.path.join(self.sharedir, si_dir, "%d" % shnum)
267 if os.path.exists(finalhome):
268 # great! we already have it. easy.
270 elif os.path.exists(incominghome):
271 # Note that we don't create BucketWriters for shnums that
272 # have a partial share (in incoming/), so if a second upload
273 # occurs while the first is still in progress, the second
274 # uploader will use different storage servers.
276 elif (not limited) or (remaining_space >= max_space_per_bucket):
277 # ok! we need to create the new share file.
278 bw = BucketWriter(self, incominghome, finalhome,
279 max_space_per_bucket, lease_info, canary)
281 bw.throw_out_all_data = True
282 bucketwriters[shnum] = bw
283 self._active_writers[bw] = 1
285 remaining_space -= max_space_per_bucket
287 # bummer! not enough space to accept this bucket
291 fileutil.make_dirs(os.path.join(self.sharedir, si_dir))
293 self.add_latency("allocate", time.time() - start)
294 return alreadygot, bucketwriters
296 def _iter_share_files(self, storage_index):
297 for shnum, filename in self._get_bucket_shares(storage_index):
298 f = open(filename, 'rb')
301 if header[:32] == MutableShareFile.MAGIC:
302 sf = MutableShareFile(filename, self)
303 # note: if the share has been migrated, the renew_lease()
304 # call will throw an exception, with information to help the
305 # client update the lease.
306 elif header[:4] == struct.pack(">L", 1):
307 sf = ShareFile(filename)
309 continue # non-sharefile
312 def remote_add_lease(self, storage_index, renew_secret, cancel_secret,
315 self.count("add-lease")
316 new_expire_time = time.time() + 31*24*60*60
317 lease_info = LeaseInfo(owner_num,
318 renew_secret, cancel_secret,
319 new_expire_time, self.my_nodeid)
320 for sf in self._iter_share_files(storage_index):
321 sf.add_or_renew_lease(lease_info)
322 self.add_latency("add-lease", time.time() - start)
325 def remote_renew_lease(self, storage_index, renew_secret):
328 new_expire_time = time.time() + 31*24*60*60
329 found_buckets = False
330 for sf in self._iter_share_files(storage_index):
332 sf.renew_lease(renew_secret, new_expire_time)
333 self.add_latency("renew", time.time() - start)
334 if not found_buckets:
335 raise IndexError("no such lease to renew")
337 def remote_cancel_lease(self, storage_index, cancel_secret):
341 total_space_freed = 0
342 found_buckets = False
343 for sf in self._iter_share_files(storage_index):
344 # note: if we can't find a lease on one share, we won't bother
345 # looking in the others. Unless something broke internally
346 # (perhaps we ran out of disk space while adding a lease), the
347 # leases on all shares will be identical.
349 # this raises IndexError if the lease wasn't present XXXX
350 total_space_freed += sf.cancel_lease(cancel_secret)
353 storagedir = os.path.join(self.sharedir,
354 storage_index_to_dir(storage_index))
355 if not os.listdir(storagedir):
358 if self.stats_provider:
359 self.stats_provider.count('storage_server.bytes_freed',
361 self.add_latency("cancel", time.time() - start)
362 if not found_buckets:
363 raise IndexError("no such storage index")
365 def bucket_writer_closed(self, bw, consumed_size):
366 if self.stats_provider:
367 self.stats_provider.count('storage_server.bytes_added', consumed_size)
368 del self._active_writers[bw]
370 def _get_bucket_shares(self, storage_index):
371 """Return a list of (shnum, pathname) tuples for files that hold
372 shares for this storage_index. In each tuple, 'shnum' will always be
373 the integer form of the last component of 'pathname'."""
374 storagedir = os.path.join(self.sharedir, storage_index_to_dir(storage_index))
376 for f in os.listdir(storagedir):
378 filename = os.path.join(storagedir, f)
379 yield (int(f), filename)
381 # Commonly caused by there being no buckets at all.
384 def remote_get_buckets(self, storage_index):
387 si_s = si_b2a(storage_index)
388 log.msg("storage: get_buckets %s" % si_s)
389 bucketreaders = {} # k: sharenum, v: BucketReader
390 for shnum, filename in self._get_bucket_shares(storage_index):
391 bucketreaders[shnum] = BucketReader(self, filename,
392 storage_index, shnum)
393 self.add_latency("get", time.time() - start)
396 def get_leases(self, storage_index):
397 """Provide an iterator that yields all of the leases attached to this
398 bucket. Each lease is returned as a tuple of (owner_num,
399 renew_secret, cancel_secret, expiration_time).
401 This method is not for client use.
404 # since all shares get the same lease data, we just grab the leases
405 # from the first share
407 shnum, filename = self._get_bucket_shares(storage_index).next()
408 sf = ShareFile(filename)
409 return sf.iter_leases()
410 except StopIteration:
413 def remote_slot_testv_and_readv_and_writev(self, storage_index,
415 test_and_write_vectors,
419 si_s = si_b2a(storage_index)
420 lp = log.msg("storage: slot_writev %s" % si_s)
421 si_dir = storage_index_to_dir(storage_index)
422 (write_enabler, renew_secret, cancel_secret) = secrets
423 # shares exist if there is a file for them
424 bucketdir = os.path.join(self.sharedir, si_dir)
426 if os.path.isdir(bucketdir):
427 for sharenum_s in os.listdir(bucketdir):
429 sharenum = int(sharenum_s)
432 filename = os.path.join(bucketdir, sharenum_s)
433 msf = MutableShareFile(filename, self)
434 msf.check_write_enabler(write_enabler, si_s)
435 shares[sharenum] = msf
436 # write_enabler is good for all existing shares.
438 # Now evaluate test vectors.
440 for sharenum in test_and_write_vectors:
441 (testv, datav, new_length) = test_and_write_vectors[sharenum]
442 if sharenum in shares:
443 if not shares[sharenum].check_testv(testv):
444 self.log("testv failed: [%d]: %r" % (sharenum, testv))
445 testv_is_good = False
448 # compare the vectors against an empty share, in which all
449 # reads return empty strings.
450 if not EmptyShare().check_testv(testv):
451 self.log("testv failed (empty): [%d] %r" % (sharenum,
453 testv_is_good = False
456 # now gather the read vectors, before we do any writes
458 for sharenum, share in shares.items():
459 read_data[sharenum] = share.readv(read_vector)
462 expire_time = time.time() + 31*24*60*60 # one month
463 lease_info = LeaseInfo(ownerid,
464 renew_secret, cancel_secret,
465 expire_time, self.my_nodeid)
468 # now apply the write vectors
469 for sharenum in test_and_write_vectors:
470 (testv, datav, new_length) = test_and_write_vectors[sharenum]
472 if sharenum in shares:
473 shares[sharenum].unlink()
475 if sharenum not in shares:
476 # allocate a new share
477 allocated_size = 2000 # arbitrary, really
478 share = self._allocate_slot_share(bucketdir, secrets,
482 shares[sharenum] = share
483 shares[sharenum].writev(datav, new_length)
484 # and update the lease
485 shares[sharenum].add_or_renew_lease(lease_info)
488 # delete empty bucket directories
489 if not os.listdir(bucketdir):
494 self.add_latency("writev", time.time() - start)
495 return (testv_is_good, read_data)
497 def _allocate_slot_share(self, bucketdir, secrets, sharenum,
498 allocated_size, owner_num=0):
499 (write_enabler, renew_secret, cancel_secret) = secrets
500 my_nodeid = self.my_nodeid
501 fileutil.make_dirs(bucketdir)
502 filename = os.path.join(bucketdir, "%d" % sharenum)
503 share = create_mutable_sharefile(filename, my_nodeid, write_enabler,
507 def remote_slot_readv(self, storage_index, shares, readv):
510 si_s = si_b2a(storage_index)
511 lp = log.msg("storage: slot_readv %s %s" % (si_s, shares),
512 facility="tahoe.storage", level=log.OPERATIONAL)
513 si_dir = storage_index_to_dir(storage_index)
514 # shares exist if there is a file for them
515 bucketdir = os.path.join(self.sharedir, si_dir)
516 if not os.path.isdir(bucketdir):
517 self.add_latency("readv", time.time() - start)
520 for sharenum_s in os.listdir(bucketdir):
522 sharenum = int(sharenum_s)
525 if sharenum in shares or not shares:
526 filename = os.path.join(bucketdir, sharenum_s)
527 msf = MutableShareFile(filename, self)
528 datavs[sharenum] = msf.readv(readv)
529 log.msg("returning shares %s" % (datavs.keys(),),
530 facility="tahoe.storage", level=log.NOISY, parent=lp)
531 self.add_latency("readv", time.time() - start)
534 def remote_advise_corrupt_share(self, share_type, storage_index, shnum,
536 fileutil.make_dirs(self.corruption_advisory_dir)
537 now = time_format.iso_utc(sep="T")
538 si_s = base32.b2a(storage_index)
539 # windows can't handle colons in the filename
540 fn = os.path.join(self.corruption_advisory_dir,
541 "%s--%s-%d" % (now, si_s, shnum)).replace(":","")
543 f.write("report: Share Corruption\n")
544 f.write("type: %s\n" % share_type)
545 f.write("storage_index: %s\n" % si_s)
546 f.write("share_number: %d\n" % shnum)
551 log.msg(format=("client claims corruption in (%(share_type)s) " +
552 "%(si)s-%(shnum)d: %(reason)s"),
553 share_type=share_type, si=si_s, shnum=shnum, reason=reason,
554 level=log.SCARY, umid="SGx2fA")