17
17
"""Core compression logic for compressing streams of related files."""
19
from __future__ import absolute_import
24
from bzrlib.lazy_import import lazy_import
25
lazy_import(globals(), """
26
26
from bzrlib import (
31
30
graph as _mod_graph,
39
from bzrlib.repofmt import pack_repo
40
from bzrlib.i18n import gettext
43
37
from bzrlib.btree_index import BTreeBuilder
44
38
from bzrlib.lru_cache import LRUSizeCache
39
from bzrlib.tsort import topo_sort
45
40
from bzrlib.versionedfile import (
48
42
AbsentContentFactory,
49
43
ChunkedContentFactory,
50
44
FulltextContentFactory,
51
VersionedFilesWithFallbacks,
54
48
# Minimum number of uncompressed bytes to try fetch at once when retrieving
55
49
# groupcompress blocks.
52
_USE_LZMA = False and (pylzma is not None)
58
54
# osutils.sha_string('')
59
55
_null_sha1 = 'da39a3ee5e6b4b0d3255bfef95601890afd80709'
139
135
self._content = ''.join(self._content_chunks)
140
136
self._content_chunks = None
141
137
if self._content is None:
142
# We join self._z_content_chunks here, because if we are
143
# decompressing, then it is *very* likely that we have a single
145
if self._z_content_chunks is None:
138
if self._z_content is None:
146
139
raise AssertionError('No content to decompress')
147
z_content = ''.join(self._z_content_chunks)
140
if self._z_content == '':
149
141
self._content = ''
150
142
elif self._compressor_name == 'lzma':
151
143
# We don't do partial lzma decomp yet
153
self._content = pylzma.decompress(z_content)
144
self._content = pylzma.decompress(self._z_content)
154
145
elif self._compressor_name == 'zlib':
155
146
# Start a zlib decompressor
156
147
if num_bytes * 4 > self._content_length * 3:
157
148
# If we are requesting more that 3/4ths of the content,
158
149
# just extract the whole thing in a single pass
159
150
num_bytes = self._content_length
160
self._content = zlib.decompress(z_content)
151
self._content = zlib.decompress(self._z_content)
162
153
self._z_content_decompressor = zlib.decompressobj()
163
154
# Seed the decompressor with the uncompressed bytes, so
164
155
# that the rest of the code is simplified
165
156
self._content = self._z_content_decompressor.decompress(
166
z_content, num_bytes + _ZLIB_DECOMP_WINDOW)
157
self._z_content, num_bytes + _ZLIB_DECOMP_WINDOW)
167
158
if not self._z_content_decompressor.unconsumed_tail:
168
159
self._z_content_decompressor = None
216
207
# XXX: Define some GCCorrupt error ?
217
208
raise AssertionError('Invalid bytes: (%d) != %d + %d' %
218
209
(len(bytes), pos, self._z_content_length))
219
self._z_content_chunks = (bytes[pos:],)
222
def _z_content(self):
223
"""Return z_content_chunks as a simple string.
225
Meant only to be used by the test suite.
227
if self._z_content_chunks is not None:
228
return ''.join(self._z_content_chunks)
210
self._z_content = bytes[pos:]
232
213
def from_bytes(cls, bytes):
288
269
self._content_length = length
289
270
self._content_chunks = content_chunks
290
271
self._content = None
291
self._z_content_chunks = None
272
self._z_content = None
293
274
def set_content(self, content):
294
275
"""Set the content of this block."""
295
276
self._content_length = len(content)
296
277
self._content = content
297
self._z_content_chunks = None
299
def _create_z_content_from_chunks(self, chunks):
278
self._z_content = None
280
def _create_z_content_using_lzma(self):
281
if self._content_chunks is not None:
282
self._content = ''.join(self._content_chunks)
283
self._content_chunks = None
284
if self._content is None:
285
raise AssertionError('Nothing to compress')
286
self._z_content = pylzma.compress(self._content)
287
self._z_content_length = len(self._z_content)
289
def _create_z_content_from_chunks(self):
300
290
compressor = zlib.compressobj(zlib.Z_DEFAULT_COMPRESSION)
301
# Peak in this point is 1 fulltext, 1 compressed text, + zlib overhead
302
# (measured peak is maybe 30MB over the above...)
303
compressed_chunks = map(compressor.compress, chunks)
291
compressed_chunks = map(compressor.compress, self._content_chunks)
304
292
compressed_chunks.append(compressor.flush())
305
# Ignore empty chunks
306
self._z_content_chunks = [c for c in compressed_chunks if c]
307
self._z_content_length = sum(map(len, self._z_content_chunks))
293
self._z_content = ''.join(compressed_chunks)
294
self._z_content_length = len(self._z_content)
309
296
def _create_z_content(self):
310
if self._z_content_chunks is not None:
297
if self._z_content is not None:
300
self._create_z_content_using_lzma()
312
302
if self._content_chunks is not None:
313
chunks = self._content_chunks
315
chunks = (self._content,)
316
self._create_z_content_from_chunks(chunks)
303
self._create_z_content_from_chunks()
305
self._z_content = zlib.compress(self._content)
306
self._z_content_length = len(self._z_content)
319
"""Create the byte stream as a series of 'chunks'"""
309
"""Encode the information into a byte stream."""
320
310
self._create_z_content()
321
header = self.GCB_HEADER
322
chunks = ['%s%d\n%d\n'
323
% (header, self._z_content_length, self._content_length),
312
header = self.GCB_LZ_HEADER
314
header = self.GCB_HEADER
316
'%d\n%d\n' % (self._z_content_length, self._content_length),
325
chunks.extend(self._z_content_chunks)
326
total_len = sum(map(len, chunks))
327
return total_len, chunks
330
"""Encode the information into a byte stream."""
331
total_len, chunks = self.to_chunks()
332
319
return ''.join(chunks)
334
321
def _dump(self, include_text=False):
448
435
# Grab and cache the raw bytes for this entry
449
436
# and break the ref-cycle with _manager since we don't need it
452
self._manager._prepare_for_extract()
453
except zlib.error as value:
454
raise errors.DecompressCorruption("zlib: " + str(value))
438
self._manager._prepare_for_extract()
455
439
block = self._manager._block
456
440
self._bytes = block.extract(self.key, self._start, self._end)
457
441
# There are code paths that first extract as fulltext, and then
476
460
_full_enough_block_size = 3*1024*1024 # size at which we won't repack
477
461
_full_enough_mixed_block_size = 2*768*1024 # 1.5MB
479
def __init__(self, block, get_compressor_settings=None):
463
def __init__(self, block):
480
464
self._block = block
481
465
# We need to preserve the ordering
482
466
self._factories = []
483
467
self._last_byte = 0
484
self._get_settings = get_compressor_settings
485
self._compressor_settings = None
487
def _get_compressor_settings(self):
488
if self._compressor_settings is not None:
489
return self._compressor_settings
491
if self._get_settings is not None:
492
settings = self._get_settings()
494
vf = GroupCompressVersionedFiles
495
settings = vf._DEFAULT_COMPRESSOR_SETTINGS
496
self._compressor_settings = settings
497
return self._compressor_settings
499
469
def add_factory(self, key, parents, start, end):
500
470
if not self._factories:
533
503
new_block.set_content(self._block._content[:last_byte])
534
504
self._block = new_block
536
def _make_group_compressor(self):
537
return GroupCompressor(self._get_compressor_settings())
539
506
def _rebuild_block(self):
540
507
"""Create a new GroupCompressBlock with only the referenced texts."""
541
compressor = self._make_group_compressor()
508
compressor = GroupCompressor()
542
509
tstart = time.time()
543
510
old_length = self._block._content_length
556
523
# block? It seems hard to come up with a method that it would
557
524
# expand, since we do full compression again. Perhaps based on a
558
525
# request that ends up poorly ordered?
559
# TODO: If the content would have expanded, then we would want to
560
# handle a case where we need to split the block.
561
# Now that we have a user-tweakable option
562
# (max_bytes_to_index), it is possible that one person set it
563
# to a very low value, causing poor compression.
564
526
delta = time.time() - tstart
565
527
self._block = new_block
566
528
trace.mutter('creating new compressed block on-the-fly in %.3fs'
717
679
z_header_bytes = zlib.compress(header_bytes)
719
681
z_header_bytes_len = len(z_header_bytes)
720
block_bytes_len, block_chunks = self._block.to_chunks()
682
block_bytes = self._block.to_bytes()
721
683
lines.append('%d\n%d\n%d\n' % (z_header_bytes_len, header_bytes_len,
723
685
lines.append(z_header_bytes)
724
lines.extend(block_chunks)
725
del z_header_bytes, block_chunks
726
# TODO: This is a point where we will double the memory consumption. To
727
# avoid this, we probably have to switch to a 'chunked' api
686
lines.append(block_bytes)
687
del z_header_bytes, block_bytes
728
688
return ''.join(lines)
731
691
def from_bytes(cls, bytes):
732
692
# TODO: This does extra string copying, probably better to do it a
733
# different way. At a minimum this creates 2 copies of the
735
694
(storage_kind, z_header_len, header_len,
736
695
block_len, rest) = bytes.split('\n', 4)
900
855
After calling this, the compressor should no longer be used
857
# TODO: this causes us to 'bloat' to 2x the size of content in the
858
# group. This has an impact for 'commit' of large objects.
859
# One possibility is to use self._content_chunks, and be lazy and
860
# only fill out self._content as a full string when we actually
861
# need it. That would at least drop the peak memory consumption
862
# for 'commit' down to ~1x the size of the largest file, at a
863
# cost of increased complexity within this code. 2x is still <<
864
# 3x the size of the largest file, so we are doing ok.
902
865
self._block.set_chunked_content(self.chunks, self.endpoint)
903
866
self.chunks = None
904
867
self._delta_index = None
923
886
class PythonGroupCompressor(_CommonGroupCompressor):
925
def __init__(self, settings=None):
926
889
"""Create a GroupCompressor.
928
891
Used only if the pyrex version is not available.
930
super(PythonGroupCompressor, self).__init__(settings)
893
super(PythonGroupCompressor, self).__init__()
931
894
self._delta_index = LinesDeltaIndex([])
932
895
# The actual content is managed by LinesDeltaIndex
933
896
self.chunks = self._delta_index.lines
971
934
It contains code very similar to SequenceMatcher because of having a similar
972
935
task. However some key differences apply:
974
* there is no junk, we want a minimal edit not a human readable diff.
975
* we don't filter very common lines (because we don't know where a good
976
range will start, and after the first text we want to be emitting minmal
978
* we chain the left side, not the right side
979
* we incrementally update the adjacency matrix as new lines are provided.
980
* we look for matches in all of the left side, so the routine which does
981
the analagous task of find_longest_match does not need to filter on the
936
- there is no junk, we want a minimal edit not a human readable diff.
937
- we don't filter very common lines (because we don't know where a good
938
range will start, and after the first text we want to be emitting minmal
940
- we chain the left side, not the right side
941
- we incrementally update the adjacency matrix as new lines are provided.
942
- we look for matches in all of the left side, so the routine which does
943
the analagous task of find_longest_match does not need to filter on the
985
def __init__(self, settings=None):
986
super(PyrexGroupCompressor, self).__init__(settings)
987
max_bytes_to_index = self._settings.get('max_bytes_to_index', 0)
988
self._delta_index = DeltaIndex(max_bytes_to_index=max_bytes_to_index)
948
super(PyrexGroupCompressor, self).__init__()
949
self._delta_index = DeltaIndex()
990
951
def _compress(self, key, bytes, max_delta_size, soft=False):
991
952
"""see _CommonGroupCompressor._compress"""
1066
1027
index = _GCGraphIndex(graph_index, lambda:True, parents=parents,
1067
1028
add_callback=graph_index.add_nodes,
1068
1029
inconsistency_fatal=inconsistency_fatal)
1069
access = pack_repo._DirectPackAccess({})
1030
access = knit._DirectPackAccess({})
1070
1031
access.set_writer(writer, graph_index, (transport, 'newpack'))
1071
1032
result = GroupCompressVersionedFiles(index, access, delta)
1072
1033
result.stream = stream
1083
1044
class _BatchingBlockFetcher(object):
1084
1045
"""Fetch group compress blocks in batches.
1086
1047
:ivar total_bytes: int of expected number of bytes needed to fetch the
1087
1048
currently pending batch.
1090
def __init__(self, gcvf, locations, get_compressor_settings=None):
1051
def __init__(self, gcvf, locations):
1091
1052
self.gcvf = gcvf
1092
1053
self.locations = locations
1176
1136
memos_to_get_stack.pop()
1178
1138
block = self.batch_memos[read_memo]
1179
self.manager = _LazyGroupContentManager(block,
1180
get_compressor_settings=self._get_compressor_settings)
1139
self.manager = _LazyGroupContentManager(block)
1181
1140
self.last_read_memo = read_memo
1182
1141
start, end = index_memo[3:5]
1183
1142
self.manager.add_factory(key, parents, start, end)
1190
1149
self.total_bytes = 0
1193
class GroupCompressVersionedFiles(VersionedFilesWithFallbacks):
1152
class GroupCompressVersionedFiles(VersionedFiles):
1194
1153
"""A group-compress based VersionedFiles implementation."""
1196
# This controls how the GroupCompress DeltaIndex works. Basically, we
1197
# compute hash pointers into the source blocks (so hash(text) => text).
1198
# However each of these references costs some memory in trade against a
1199
# more accurate match result. For very large files, they either are
1200
# pre-compressed and change in bulk whenever they change, or change in just
1201
# local blocks. Either way, 'improved resolution' is not very helpful,
1202
# versus running out of memory trying to track everything. The default max
1203
# gives 100% sampling of a 1MB file.
1204
_DEFAULT_MAX_BYTES_TO_INDEX = 1024 * 1024
1205
_DEFAULT_COMPRESSOR_SETTINGS = {'max_bytes_to_index':
1206
_DEFAULT_MAX_BYTES_TO_INDEX}
1208
def __init__(self, index, access, delta=True, _unadded_refs=None,
1155
def __init__(self, index, access, delta=True, _unadded_refs=None):
1210
1156
"""Create a GroupCompressVersionedFiles object.
1212
1158
:param index: The index object storing access and graph data.
1213
1159
:param access: The access object storing raw data.
1214
1160
:param delta: Whether to delta compress or just entropy compress.
1215
1161
:param _unadded_refs: private parameter, don't use.
1216
:param _group_cache: private parameter, don't use.
1218
1163
self._index = index
1219
1164
self._access = access
1221
1166
if _unadded_refs is None:
1222
1167
_unadded_refs = {}
1223
1168
self._unadded_refs = _unadded_refs
1224
if _group_cache is None:
1225
_group_cache = LRUSizeCache(max_size=50*1024*1024)
1226
self._group_cache = _group_cache
1227
self._immediate_fallback_vfs = []
1228
self._max_bytes_to_index = None
1169
self._group_cache = LRUSizeCache(max_size=50*1024*1024)
1170
self._fallback_vfs = []
1230
1172
def without_fallbacks(self):
1231
1173
"""Return a clone of this object without any fallbacks configured."""
1232
1174
return GroupCompressVersionedFiles(self._index, self._access,
1233
self._delta, _unadded_refs=dict(self._unadded_refs),
1234
_group_cache=self._group_cache)
1175
self._delta, _unadded_refs=dict(self._unadded_refs))
1236
1177
def add_lines(self, key, parents, lines, parent_texts=None,
1237
1178
left_matching_blocks=None, nostore_sha=None, random_id=False,
1241
1182
:param key: The key tuple of the text to add.
1242
1183
:param parents: The parents key tuples of the text to add.
1243
1184
:param lines: A list of lines. Each line must be a bytestring. And all
1244
of them except the last must be terminated with \\n and contain no
1245
other \\n's. The last line may either contain no \\n's or a single
1246
terminating \\n. If the lines list does meet this constraint the
1247
add routine may error or may succeed - but you will be unable to
1248
read the data back accurately. (Checking the lines have been split
1185
of them except the last must be terminated with \n and contain no
1186
other \n's. The last line may either contain no \n's or a single
1187
terminating \n. If the lines list does meet this constraint the add
1188
routine may error or may succeed - but you will be unable to read
1189
the data back accurately. (Checking the lines have been split
1249
1190
correctly is expensive and extremely unlikely to catch bugs so it
1250
1191
is not done at runtime unless check_content is True.)
1251
1192
:param parent_texts: An optional dictionary containing the opaque
1346
1287
self._check_lines_not_unicode(lines)
1347
1288
self._check_lines_are_lines(lines)
1290
def get_known_graph_ancestry(self, keys):
1291
"""Get a KnownGraph instance with the ancestry of keys."""
1292
# Note that this is identical to
1293
# KnitVersionedFiles.get_known_graph_ancestry, but they don't share
1295
parent_map, missing_keys = self._index.find_ancestry(keys)
1296
for fallback in self._fallback_vfs:
1297
if not missing_keys:
1299
(f_parent_map, f_missing_keys) = fallback._index.find_ancestry(
1301
parent_map.update(f_parent_map)
1302
missing_keys = f_missing_keys
1303
kg = _mod_graph.KnownGraph(parent_map)
1349
1306
def get_parent_map(self, keys):
1350
1307
"""Get a map of the graph parents of keys.
1490
1447
The returned objects should be in the order defined by 'ordering',
1491
1448
which can weave between different sources.
1493
1449
:param ordering: Must be one of 'topological' or 'groupcompress'
1494
1450
:return: List of [(source, [keys])] tuples, such that all keys are in
1495
1451
the defined order, regardless of source.
1497
1453
if ordering == 'topological':
1498
present_keys = tsort.topo_sort(parent_map)
1454
present_keys = topo_sort(parent_map)
1500
1456
# ordering == 'groupcompress'
1501
1457
# XXX: This only optimizes for the target ordering. We may need
1590
1546
# - we encounter an unadded ref, or
1591
1547
# - we run out of keys, or
1592
1548
# - the total bytes to retrieve for this batch > BATCH_SIZE
1593
batcher = _BatchingBlockFetcher(self, locations,
1594
get_compressor_settings=self._get_compressor_settings)
1549
batcher = _BatchingBlockFetcher(self, locations)
1595
1550
for source, keys in source_keys:
1596
1551
if source is self:
1597
1552
for key in keys:
1643
1598
for _ in self._insert_record_stream(stream, random_id=False):
1646
def _get_compressor_settings(self):
1647
if self._max_bytes_to_index is None:
1648
# TODO: VersionedFiles don't know about their containing
1649
# repository, so they don't have much of an idea about their
1650
# location. So for now, this is only a global option.
1651
c = config.GlobalConfig()
1652
val = c.get_user_option('bzr.groupcompress.max_bytes_to_index')
1656
except ValueError, e:
1657
trace.warning('Value for '
1658
'"bzr.groupcompress.max_bytes_to_index"'
1659
' %r is not an integer'
1663
val = self._DEFAULT_MAX_BYTES_TO_INDEX
1664
self._max_bytes_to_index = val
1665
return {'max_bytes_to_index': self._max_bytes_to_index}
1667
def _make_group_compressor(self):
1668
return GroupCompressor(self._get_compressor_settings())
1670
1601
def _insert_record_stream(self, stream, random_id=False, nostore_sha=None,
1671
1602
reuse_blocks=True):
1672
1603
"""Internal core to insert a record stream into this container.
1696
1627
# This will go up to fulltexts for gc to gc fetching, which isn't
1698
self._compressor = self._make_group_compressor()
1629
self._compressor = GroupCompressor()
1699
1630
self._unadded_refs = {}
1700
1631
keys_to_add = []
1702
bytes_len, chunks = self._compressor.flush().to_chunks()
1703
self._compressor = self._make_group_compressor()
1704
# Note: At this point we still have 1 copy of the fulltext (in
1705
# record and the var 'bytes'), and this generates 2 copies of
1706
# the compressed text (one for bytes, one in chunks)
1707
# TODO: Push 'chunks' down into the _access api, so that we don't
1708
# have to double compressed memory here
1709
# TODO: Figure out how to indicate that we would be happy to free
1710
# the fulltext content at this point. Note that sometimes we
1711
# will want it later (streaming CHK pages), but most of the
1712
# time we won't (everything else)
1713
bytes = ''.join(chunks)
1633
bytes = self._compressor.flush().to_bytes()
1715
1634
index, start, length = self._access.add_raw_records(
1716
1635
[(None, len(bytes))], bytes)[0]
1736
1656
raise errors.RevisionNotPresent(record.key, self)
1738
1658
if record.key in inserted_keys:
1739
trace.note(gettext('Insert claimed random_id=True,'
1740
' but then inserted %r two times'), record.key)
1659
trace.note('Insert claimed random_id=True,'
1660
' but then inserted %r two times', record.key)
1742
1662
inserted_keys.add(record.key)
1743
1663
if reuse_blocks:
1882
1802
"""See VersionedFiles.keys."""
1883
1803
if 'evil' in debug.debug_flags:
1884
1804
trace.mutter_callsite(2, "keys scales with size of history")
1885
sources = [self._index] + self._immediate_fallback_vfs
1805
sources = [self._index] + self._fallback_vfs
1887
1807
for source in sources:
1888
1808
result.update(source.keys())
1892
class _GCBuildDetails(object):
1893
"""A blob of data about the build details.
1895
This stores the minimal data, which then allows compatibility with the old
1896
api, without taking as much memory.
1899
__slots__ = ('_index', '_group_start', '_group_end', '_basis_end',
1900
'_delta_end', '_parents')
1903
compression_parent = None
1905
def __init__(self, parents, position_info):
1906
self._parents = parents
1907
(self._index, self._group_start, self._group_end, self._basis_end,
1908
self._delta_end) = position_info
1911
return '%s(%s, %s)' % (self.__class__.__name__,
1912
self.index_memo, self._parents)
1915
def index_memo(self):
1916
return (self._index, self._group_start, self._group_end,
1917
self._basis_end, self._delta_end)
1920
def record_details(self):
1921
return static_tuple.StaticTuple(self.method, None)
1923
def __getitem__(self, offset):
1924
"""Compatibility thunk to act like a tuple."""
1926
return self.index_memo
1928
return self.compression_parent # Always None
1930
return self._parents
1932
return self.record_details
1934
raise IndexError('offset out of range')
1940
1812
class _GCGraphIndex(object):
1941
1813
"""Mapper from GroupCompressVersionedFiles needs into GraphIndex storage."""
2117
1989
:param keys: An iterable of keys.
2118
1990
:return: A dict of key:
2119
1991
(index_memo, compression_parent, parents, record_details).
2121
* index_memo: opaque structure to pass to read_records to extract
2123
* compression_parent: Content that this record is built upon, may
2125
* parents: Logical parents of this node
2126
* record_details: extra information about the content which needs
2127
to be passed to Factory.parse_record
1993
opaque structure to pass to read_records to extract the raw
1996
Content that this record is built upon, may be None
1998
Logical parents of this node
2000
extra information about the content which needs to be passed to
2001
Factory.parse_record
2129
2003
self._check_read()
2158
2033
# each, or about 7MB. Note that it might be even more when you consider
2159
2034
# how PyInt is allocated in separate slabs. And you can't return a slab
2160
2035
# to the OS if even 1 int on it is in use. Note though that Python uses
2161
# a LIFO when re-using PyInt slots, which might cause more
2036
# a LIFO when re-using PyInt slots, which probably causes more
2162
2037
# fragmentation.
2163
2038
start = int(bits[0])
2164
2039
start = self._int_cache.setdefault(start, start)
added
removed
bzrlib/groupcompress.py
