1 """High-level interface for working with LBS archives.
3 This module provides an easy interface for reading from and manipulating
4 various parts of an LBS archive:
5 - listing the snapshots and segments present
6 - reading segment contents
7 - parsing snapshot descriptors and snapshot metadata logs
8 - reading and maintaining the local object database
11 from __future__ import division
12 import os, re, sha, tarfile, tempfile, thread
13 from pysqlite2 import dbapi2 as sqlite3
15 # The largest supported snapshot format that can be understood.
16 FORMAT_VERSION = (0, 6) # LBS Snapshot v0.6
18 # Maximum number of nested indirect references allowed in a snapshot.
19 MAX_RECURSION_DEPTH = 3
21 # All segments which have been accessed this session.
22 accessed_segments = set()
25 """A class which merely acts as a data container.
27 Instances of this class (or its subclasses) are merely used to store data
28 in various attributes. No methods are provided.
32 return "<%s %s>" % (self.__class__, self.__dict__)
34 CHECKSUM_ALGORITHMS = {
38 class ChecksumCreator:
39 """Compute an LBS checksum for provided data.
41 The algorithm used is selectable, but currently defaults to sha1.
44 def __init__(self, algorithm='sha1'):
45 self.algorithm = algorithm
46 self.hash = CHECKSUM_ALGORITHMS[algorithm]()
48 def update(self, data):
49 self.hash.update(data)
53 return "%s=%s" % (self.algorithm, self.hash.hexdigest())
55 class ChecksumVerifier:
56 """Verify whether a checksum from a snapshot matches the supplied data."""
58 def __init__(self, checksumstr):
59 """Create an object to check the supplied checksum."""
61 (algo, checksum) = checksumstr.split("=", 1)
62 self.checksum = checksum
63 self.hash = CHECKSUM_ALGORITHMS[algo]()
65 def update(self, data):
66 self.hash.update(data)
69 """Return a boolean indicating whether the checksum matches."""
71 result = self.hash.hexdigest()
72 return result == self.checksum
74 class LowlevelDataStore:
75 """Access to the backup store containing segments and snapshot descriptors.
77 Instances of this class are used to get direct filesystem-level access to
78 the backup data. To read a backup, a caller will ordinarily not care about
79 direct access to backup segments, but will instead merely need to access
80 objects from those segments. The ObjectStore class provides a suitable
81 wrapper around a DataStore to give this high-level access.
84 def __init__(self, path):
87 # Low-level filesystem access. These methods could be overwritten to
88 # provide access to remote data stores.
89 def lowlevel_list(self):
90 """Get a listing of files stored."""
92 return os.listdir(self.path)
94 def lowlevel_open(self, filename):
95 """Return a file-like object for reading data from the given file."""
97 return open(os.path.join(self.path, filename), 'rb')
99 def lowlevel_stat(self, filename):
100 """Return a dictionary of information about the given file.
102 Currently, the only defined field is 'size', giving the size of the
106 stat = os.stat(os.path.join(self.path, filename))
107 return {'size': stat.st_size}
109 # Slightly higher-level list methods.
110 def list_snapshots(self):
111 for f in self.lowlevel_list():
112 m = re.match(r"^snapshot-(.*)\.lbs$", f)
116 def list_segments(self):
117 for f in self.lowlevel_list():
118 m = re.match(r"^([0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12})(\.\S+)?$", f)
123 def __init__(self, data_store):
124 self.store = data_store
129 def get_cachedir(self):
130 if self.cachedir is None:
131 self.cachedir = tempfile.mkdtemp(".lbs")
135 if self.cachedir is not None:
136 # TODO: Avoid use of system, make this safer
137 os.system("rm -rf " + self.cachedir)
141 def parse_ref(refstr):
142 m = re.match(r"^zero\[(\d+)\]$", refstr)
144 return ("zero", None, None, (0, int(m.group(1))))
146 m = re.match(r"^([-0-9a-f]+)\/([0-9a-f]+)(\(\S+\))?(\[((\d+)\+)?(\d+)\])?$", refstr)
151 checksum = m.group(3)
154 if checksum is not None:
155 checksum = checksum.lstrip("(").rstrip(")")
157 if slice is not None:
158 if m.group(5) is None:
160 slice = (0, int(m.group(7)))
162 slice = (int(m.group(6)), int(m.group(7)))
164 return (segment, object, checksum, slice)
166 def get_segment(self, segment):
167 accessed_segments.add(segment)
168 raw = self.store.lowlevel_open(segment + ".tar.gpg")
170 (input, output) = os.popen2("lbs-filter-gpg --decrypt")
171 def copy_thread(src, dst):
174 block = src.read(BLOCK_SIZE)
175 if len(block) == 0: break
179 thread.start_new_thread(copy_thread, (raw, input))
182 def load_segment(self, segment):
183 seg = tarfile.open(segment, 'r|', self.get_segment(segment))
185 data_obj = seg.extractfile(item)
186 path = item.name.split('/')
187 if len(path) == 2 and path[0] == segment:
188 yield (path[1], data_obj.read())
190 def load_snapshot(self, snapshot):
191 file = self.store.lowlevel_open("snapshot-" + snapshot + ".lbs")
192 return file.read().splitlines(True)
194 def extract_segment(self, segment):
195 segdir = os.path.join(self.get_cachedir(), segment)
197 for (object, data) in self.load_segment(segment):
198 f = open(os.path.join(segdir, object), 'wb')
202 def load_object(self, segment, object):
203 accessed_segments.add(segment)
204 path = os.path.join(self.get_cachedir(), segment, object)
205 if not os.access(path, os.R_OK):
206 self.extract_segment(segment)
207 if segment in self.lru_list: self.lru_list.remove(segment)
208 self.lru_list.append(segment)
209 while len(self.lru_list) > self.CACHE_SIZE:
210 os.system("rm -rf " + os.path.join(self.cachedir, self.lru_list[0]))
211 self.lru_list = self.lru_list[1:]
212 return open(path, 'rb').read()
214 def get(self, refstr):
215 """Fetch the given object and return it.
217 The input should be an object reference, in string form.
220 (segment, object, checksum, slice) = self.parse_ref(refstr)
222 if segment == "zero":
223 return "\0" * slice[1]
225 data = self.load_object(segment, object)
227 if checksum is not None:
228 verifier = ChecksumVerifier(checksum)
229 verifier.update(data)
230 if not verifier.valid():
233 if slice is not None:
234 (start, length) = slice
235 data = data[start:start+length]
236 if len(data) != length: raise IndexError
240 def parse(lines, terminate=None):
241 """Generic parser for RFC822-style "Key: Value" data streams.
243 This parser can be used to read metadata logs and snapshot root descriptor
246 lines must be an iterable object which yields a sequence of lines of input.
248 If terminate is specified, it is used as a predicate to determine when to
249 stop reading input lines.
256 # Strip off a trailing newline, if present
257 if len(l) > 0 and l[-1] == "\n":
260 if terminate is not None and terminate(l):
261 if len(dict) > 0: yield dict
266 m = re.match(r"^(\w+):\s*(.*)$", l)
268 dict[m.group(1)] = m.group(2)
269 last_key = m.group(1)
270 elif len(l) > 0 and l[0].isspace() and last_key is not None:
275 if len(dict) > 0: yield dict
277 def parse_full(lines):
279 return parse(lines).next()
280 except StopIteration:
283 def parse_metadata_version(s):
284 """Convert a string with the snapshot version format to a tuple."""
286 m = re.match(r"^LBS Snapshot v(\d+(\.\d+)*)$", s)
290 return tuple([int(d) for d in m.group(1).split(".")])
292 def read_metadata(object_store, root):
293 """Iterate through all lines in the metadata log, following references."""
295 # Stack for keeping track of recursion when following references to
296 # portions of the log. The last entry in the stack corresponds to the
297 # object currently being parsed. Each entry is a list of lines which have
298 # been reversed, so that popping successive lines from the end of each list
299 # will return lines of the metadata log in order.
302 def follow_ref(refstr):
303 if len(stack) >= MAX_RECURSION_DEPTH: raise OverflowError
304 lines = object_store.get(refstr).splitlines(True)
310 while len(stack) > 0:
317 # An indirect reference which we must follow?
318 if len(line) > 0 and line[0] == '@':
326 """Metadata for a single file (or directory or...) from a snapshot."""
328 # Functions for parsing various datatypes that can appear in a metadata log
332 """Decode an integer, expressed in decimal, octal, or hexadecimal."""
333 if s.startswith("0x"):
335 elif s.startswith("0"):
342 """Decode a URI-encoded (%xx escapes) string."""
343 def hex_decode(m): return chr(int(m.group(1), 16))
344 return re.sub(r"%([0-9a-f]{2})", hex_decode, s)
348 """An unecoded string."""
353 """Decode a user/group to a tuple of uid/gid followed by name."""
355 uid = MetadataItem.decode_int(items[0])
358 if items[1].startswith("(") and items[1].endswith(")"):
359 name = MetadataItem.decode_str(items[1][1:-1])
363 def decode_device(s):
364 """Decode a device major/minor number."""
365 (major, minor) = map(MetadataItem.decode_int, s.split("/"))
366 return (major, minor)
370 def __init__(self, fields, object_store):
371 """Initialize from a dictionary of key/value pairs from metadata log."""
374 self.object_store = object_store
376 self.items = self.Items()
377 for (k, v) in fields.items():
378 if k in self.field_types:
379 decoder = self.field_types[k]
380 setattr(self.items, k, decoder(v))
384 """Return an iterator for the data blocks that make up a file."""
386 # This traverses the list of blocks that make up a file, following
387 # indirect references. It is implemented in much the same way as
388 # read_metadata, so see that function for details of the technique.
390 objects = self.fields['data'].split()
394 def follow_ref(refstr):
395 if len(stack) >= MAX_RECURSION_DEPTH: raise OverflowError
396 objects = self.object_store.get(refstr).split()
398 stack.append(objects)
400 while len(stack) > 0:
407 # An indirect reference which we must follow?
408 if len(ref) > 0 and ref[0] == '@':
413 # Description of fields that might appear, and how they should be parsed.
414 MetadataItem.field_types = {
415 'name': MetadataItem.decode_str,
416 'type': MetadataItem.raw_str,
417 'mode': MetadataItem.decode_int,
418 'device': MetadataItem.decode_device,
419 'user': MetadataItem.decode_user,
420 'group': MetadataItem.decode_user,
421 'ctime': MetadataItem.decode_int,
422 'mtime': MetadataItem.decode_int,
423 'links': MetadataItem.decode_int,
424 'inode': MetadataItem.raw_str,
425 'checksum': MetadataItem.decode_str,
426 'size': MetadataItem.decode_int,
427 'contents': MetadataItem.decode_str,
428 'target': MetadataItem.decode_str,
431 def iterate_metadata(object_store, root):
432 for d in parse(read_metadata(object_store, root), lambda l: len(l) == 0):
433 yield MetadataItem(d, object_store)
436 """Access to the local database of snapshot contents and object checksums.
438 The local database is consulted when creating a snapshot to determine what
439 data can be re-used from old snapshots. Segment cleaning is performed by
440 manipulating the data in the local database; the local database also
441 includes enough data to guide the segment cleaning process.
444 def __init__(self, path, dbname="localdb.sqlite"):
445 self.db_connection = sqlite3.connect(path + "/" + dbname)
447 # Low-level database access. Use these methods when there isn't a
448 # higher-level interface available. Exception: do, however, remember to
449 # use the commit() method after making changes to make sure they are
450 # actually saved, even when going through higher-level interfaces.
452 "Commit any pending changes to the local database."
453 self.db_connection.commit()
456 "Roll back any pending changes to the local database."
457 self.db_connection.rollback()
460 "Return a DB-API cursor for directly accessing the local database."
461 return self.db_connection.cursor()
463 def list_schemes(self):
464 """Return the list of snapshots found in the local database.
466 The returned value is a list of tuples (id, scheme, name, time, intent).
470 cur.execute("select distinct scheme from snapshots")
471 schemes = [row[0] for row in cur.fetchall()]
475 def garbage_collect(self, scheme, intent=1.0):
476 """Delete entries from old snapshots from the database.
478 Only snapshots with the specified scheme name will be deleted. If
479 intent is given, it gives the intended next snapshot type, to determine
480 how aggressively to clean (for example, intent=7 could be used if the
481 next snapshot will be a weekly snapshot).
486 # Find the id of the last snapshot to be created. This is used for
487 # measuring time in a way: we record this value in each segment we
488 # expire on this run, and then on a future run can tell if there have
489 # been intervening backups made.
490 cur.execute("select max(snapshotid) from snapshots")
491 last_snapshotid = cur.fetchone()[0]
493 # Get the list of old snapshots for this scheme. Delete all the old
494 # ones. Rules for what to keep:
495 # - Always keep the most recent snapshot.
496 # - If snapshot X is younger than Y, and X has higher intent, then Y
498 cur.execute("""select snapshotid, name, intent,
499 julianday('now') - timestamp as age
500 from snapshots where scheme = ?
501 order by age""", (scheme,))
505 for (id, name, snap_intent, snap_age) in cur.fetchall():
507 if snap_intent < max_intent:
508 # Delete small-intent snapshots if there is a more recent
509 # large-intent snapshot.
511 elif snap_intent == intent:
512 # Delete previous snapshots with the specified intent level.
515 if can_delete and not first:
516 print "Delete snapshot %d (%s)" % (id, name)
517 cur.execute("delete from snapshots where snapshotid = ?",
520 max_intent = max(max_intent, snap_intent)
522 # Delete entries in the segments_used table which are for non-existent
524 cur.execute("""delete from segments_used
525 where snapshotid not in
526 (select snapshotid from snapshots)""")
528 # Find segments which contain no objects used by any current snapshots,
529 # and delete them from the segment table.
530 cur.execute("""delete from segments where segmentid not in
531 (select segmentid from segments_used)""")
533 # Delete unused objects in the block_index table. By "unused", we mean
534 # any object which was stored in a segment which has been deleted, and
535 # any object in a segment which was marked for cleaning and has had
536 # cleaning performed already (the expired time is less than the current
537 # largest snapshot id).
538 cur.execute("""delete from block_index
539 where segmentid not in (select segmentid from segments)
540 or segmentid in (select segmentid from segments
541 where expire_time < ?)""",
545 class SegmentInfo(Struct): pass
547 def get_segment_cleaning_list(self, age_boost=0.0):
548 """Return a list of all current segments with information for cleaning.
550 Return all segments which are currently known in the local database
551 (there might be other, older segments in the archive itself), and
552 return usage statistics for each to help decide which segments to
555 The returned list will be sorted by estimated cleaning benefit, with
556 segments that are best to clean at the start of the list.
558 If specified, the age_boost parameter (measured in days) will added to
559 the age of each segment, as a way of adjusting the benefit computation
560 before a long-lived snapshot is taken (for example, age_boost might be
561 set to 7 when cleaning prior to taking a weekly snapshot).
566 cur.execute("""select segmentid, used, size, mtime,
567 julianday('now') - mtime as age from segment_info
568 where expire_time is null""")
570 info = self.SegmentInfo()
572 info.used_bytes = row[1]
573 info.size_bytes = row[2]
575 info.age_days = row[4]
577 # If age is not available for whatever reason, treat it as 0.0.
578 if info.age_days is None:
581 # Benefit calculation: u is the estimated fraction of each segment
582 # which is utilized (bytes belonging to objects still in use
583 # divided by total size; this doesn't take compression or storage
584 # overhead into account, but should give a reasonable estimate).
586 # The total benefit is a heuristic that combines several factors:
587 # the amount of space that can be reclaimed (1 - u), an ageing
588 # factor (info.age_days) that favors cleaning old segments to young
589 # ones and also is more likely to clean segments that will be
590 # rewritten for long-lived snapshots (age_boost), and finally a
591 # penalty factor for the cost of re-uploading data (u + 0.1).
592 u = info.used_bytes / info.size_bytes
593 info.cleaning_benefit \
594 = (1 - u) * (info.age_days + age_boost) / (u + 0.1)
596 segments.append(info)
598 segments.sort(cmp, key=lambda s: s.cleaning_benefit, reverse=True)
601 def mark_segment_expired(self, segment):
602 """Mark a segment for cleaning in the local database.
604 The segment parameter should be either a SegmentInfo object or an
605 integer segment id. Objects in the given segment will be marked as
606 expired, which means that any future snapshots that would re-use those
607 objects will instead write out a new copy of the object, and thus no
608 future snapshots will depend upon the given segment.
611 if isinstance(segment, int):
613 elif isinstance(segment, self.SegmentInfo):
616 raise TypeError("Invalid segment: %s, must be of type int or SegmentInfo, not %s" % (segment, type(segment)))
619 cur.execute("select max(snapshotid) from snapshots")
620 last_snapshotid = cur.fetchone()[0]
621 cur.execute("update segments set expire_time = ? where segmentid = ?",
622 (last_snapshotid, id))
623 cur.execute("update block_index set expired = 0 where segmentid = ?",
626 def balance_expired_objects(self):
627 """Analyze expired objects in segments to be cleaned and group by age.
629 Update the block_index table of the local database to group expired
630 objects by age. The exact number of buckets and the cutoffs for each
631 are dynamically determined. Calling this function after marking
632 segments expired will help in the segment cleaning process, by ensuring
633 that when active objects from clean segments are rewritten, they will
634 be placed into new segments roughly grouped by age.
637 # The expired column of the block_index table is used when generating a
638 # new LBS snapshot. A null value indicates that an object may be
639 # re-used. Otherwise, an object must be written into a new segment if
640 # needed. Objects with distinct expired values will be written into
641 # distinct segments, to allow for some grouping by age. The value 0 is
642 # somewhat special in that it indicates any rewritten objects can be
643 # placed in the same segment as completely new objects; this can be
644 # used for very young objects which have been expired, or objects not
645 # expected to be encountered.
647 # In the balancing process, all objects which are not used in any
648 # current snapshots will have expired set to 0. Objects which have
649 # been seen will be sorted by age and will have expired values set to
650 # 0, 1, 2, and so on based on age (with younger objects being assigned
651 # lower values). The number of buckets and the age cutoffs is
652 # determined by looking at the distribution of block ages.
656 # Mark all expired objects with expired = 0; these objects will later
657 # have values set to indicate groupings of objects when repacking.
658 cur.execute("""update block_index set expired = 0
659 where expired is not null""")
661 # We will want to aim for at least one full segment for each bucket
662 # that we eventually create, but don't know how many bytes that should
663 # be due to compression. So compute the average number of bytes in
664 # each expired segment as a rough estimate for the minimum size of each
665 # bucket. (This estimate could be thrown off by many not-fully-packed
666 # segments, but for now don't worry too much about that.) If we can't
667 # compute an average, it's probably because there are no expired
668 # segments, so we have no more work to do.
669 cur.execute("""select avg(size) from segments
671 (select distinct segmentid from block_index
672 where expired is not null)""")
673 segment_size_estimate = cur.fetchone()[0]
674 if not segment_size_estimate:
677 # Next, extract distribution of expired objects (number and size) by
678 # age. Save the timestamp for "now" so that the classification of
679 # blocks into age buckets will not change later in the function, after
680 # time has passed. Set any timestamps in the future to now, so we are
681 # guaranteed that for the rest of this function, age is always
683 cur.execute("select julianday('now')")
684 now = cur.fetchone()[0]
686 cur.execute("""update block_index set timestamp = ?
687 where timestamp > ? and expired is not null""",
690 cur.execute("""select round(? - timestamp) as age, count(*), sum(size)
691 from block_index where expired = 0
692 group by age order by age""", (now,))
693 distribution = cur.fetchall()
695 # Start to determine the buckets for expired objects. Heuristics used:
696 # - An upper bound on the number of buckets is given by the number of
697 # segments we estimate it will take to store all data. In fact,
698 # aim for a couple of segments per bucket.
699 # - Place very young objects in bucket 0 (place with new objects)
700 # unless there are enough of them to warrant a separate bucket.
701 # - Try not to create unnecessarily many buckets, since fewer buckets
702 # will allow repacked data to be grouped based on spatial locality
703 # (while more buckets will group by temporal locality). We want a
706 total_bytes = sum([i[2] for i in distribution])
707 target_buckets = 2 * (total_bytes / segment_size_estimate) ** 0.4
708 min_size = 1.5 * segment_size_estimate
709 target_size = max(2 * segment_size_estimate,
710 total_bytes / target_buckets)
712 print "segment_size:", segment_size_estimate
713 print "distribution:", distribution
714 print "total_bytes:", total_bytes
715 print "target_buckets:", target_buckets
716 print "min, target size:", min_size, target_size
718 # Chosen cutoffs. Each bucket consists of objects with age greater
719 # than one cutoff value, but not greater than the next largest cutoff.
722 # Starting with the oldest objects, begin grouping together into
723 # buckets of size at least target_size bytes.
724 distribution.reverse()
726 min_age_bucket = False
727 for (age, items, size) in distribution:
728 if bucket_size >= target_size \
729 or (age < MIN_AGE and not min_age_bucket):
730 if bucket_size < target_size and len(cutoffs) > 0:
737 min_age_bucket = True
739 # The last (youngest) bucket will be group 0, unless it has enough data
740 # to be of size min_size by itself, or there happen to be no objects
741 # less than MIN_AGE at all.
742 if bucket_size >= min_size or not min_age_bucket:
746 print "cutoffs:", cutoffs
748 # Update the database to assign each object to the appropriate bucket.
750 for i in range(len(cutoffs)):
751 cur.execute("""update block_index set expired = ?
752 where round(? - timestamp) > ?
753 and expired is not null""",
754 (i, now, cutoffs[i]))