17
17
"""Core compression logic for compressing streams of related files."""
19
from __future__ import absolute_import
21
from future_builtins import map
25
from .lazy_import import lazy_import
26
lazy_import(globals(), """
30
32
graph as _mod_graph,
37
from bzrlib.btree_index import BTreeBuilder
38
from bzrlib.lru_cache import LRUSizeCache
39
from bzrlib.tsort import topo_sort
40
from bzrlib.versionedfile import (
40
from breezy.repofmt import pack_repo
41
from breezy.i18n import gettext
44
from .btree_index import BTreeBuilder
45
from .lru_cache import LRUSizeCache
46
from .versionedfile import (
42
49
AbsentContentFactory,
43
50
ChunkedContentFactory,
44
51
FulltextContentFactory,
52
VersionedFilesWithFallbacks,
48
55
# Minimum number of uncompressed bytes to try fetch at once when retrieving
49
56
# groupcompress blocks.
52
_USE_LZMA = False and (pylzma is not None)
54
59
# osutils.sha_string('')
55
60
_null_sha1 = 'da39a3ee5e6b4b0d3255bfef95601890afd80709'
135
140
self._content = ''.join(self._content_chunks)
136
141
self._content_chunks = None
137
142
if self._content is None:
138
if self._z_content is None:
143
# We join self._z_content_chunks here, because if we are
144
# decompressing, then it is *very* likely that we have a single
146
if self._z_content_chunks is None:
139
147
raise AssertionError('No content to decompress')
140
if self._z_content == '':
148
z_content = ''.join(self._z_content_chunks)
141
150
self._content = ''
142
151
elif self._compressor_name == 'lzma':
143
152
# We don't do partial lzma decomp yet
144
self._content = pylzma.decompress(self._z_content)
154
self._content = pylzma.decompress(z_content)
145
155
elif self._compressor_name == 'zlib':
146
156
# Start a zlib decompressor
147
157
if num_bytes * 4 > self._content_length * 3:
148
158
# If we are requesting more that 3/4ths of the content,
149
159
# just extract the whole thing in a single pass
150
160
num_bytes = self._content_length
151
self._content = zlib.decompress(self._z_content)
161
self._content = zlib.decompress(z_content)
153
163
self._z_content_decompressor = zlib.decompressobj()
154
164
# Seed the decompressor with the uncompressed bytes, so
155
165
# that the rest of the code is simplified
156
166
self._content = self._z_content_decompressor.decompress(
157
self._z_content, num_bytes + _ZLIB_DECOMP_WINDOW)
167
z_content, num_bytes + _ZLIB_DECOMP_WINDOW)
158
168
if not self._z_content_decompressor.unconsumed_tail:
159
169
self._z_content_decompressor = None
207
217
# XXX: Define some GCCorrupt error ?
208
218
raise AssertionError('Invalid bytes: (%d) != %d + %d' %
209
219
(len(bytes), pos, self._z_content_length))
210
self._z_content = bytes[pos:]
220
self._z_content_chunks = (bytes[pos:],)
223
def _z_content(self):
224
"""Return z_content_chunks as a simple string.
226
Meant only to be used by the test suite.
228
if self._z_content_chunks is not None:
229
return ''.join(self._z_content_chunks)
213
233
def from_bytes(cls, bytes):
269
289
self._content_length = length
270
290
self._content_chunks = content_chunks
271
291
self._content = None
272
self._z_content = None
292
self._z_content_chunks = None
274
294
def set_content(self, content):
275
295
"""Set the content of this block."""
276
296
self._content_length = len(content)
277
297
self._content = content
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):
298
self._z_content_chunks = None
300
def _create_z_content_from_chunks(self, chunks):
290
301
compressor = zlib.compressobj(zlib.Z_DEFAULT_COMPRESSION)
291
compressed_chunks = map(compressor.compress, self._content_chunks)
302
# Peak in this point is 1 fulltext, 1 compressed text, + zlib overhead
303
# (measured peak is maybe 30MB over the above...)
304
compressed_chunks = list(map(compressor.compress, chunks))
292
305
compressed_chunks.append(compressor.flush())
293
self._z_content = ''.join(compressed_chunks)
294
self._z_content_length = len(self._z_content)
306
# Ignore empty chunks
307
self._z_content_chunks = [c for c in compressed_chunks if c]
308
self._z_content_length = sum(map(len, self._z_content_chunks))
296
310
def _create_z_content(self):
297
if self._z_content is not None:
300
self._create_z_content_using_lzma()
311
if self._z_content_chunks is not None:
302
313
if self._content_chunks is not None:
303
self._create_z_content_from_chunks()
305
self._z_content = zlib.compress(self._content)
306
self._z_content_length = len(self._z_content)
314
chunks = self._content_chunks
316
chunks = (self._content,)
317
self._create_z_content_from_chunks(chunks)
320
"""Create the byte stream as a series of 'chunks'"""
321
self._create_z_content()
322
header = self.GCB_HEADER
323
chunks = ['%s%d\n%d\n'
324
% (header, self._z_content_length, self._content_length),
326
chunks.extend(self._z_content_chunks)
327
total_len = sum(map(len, chunks))
328
return total_len, chunks
308
330
def to_bytes(self):
309
331
"""Encode the information into a byte stream."""
310
self._create_z_content()
312
header = self.GCB_LZ_HEADER
314
header = self.GCB_HEADER
316
'%d\n%d\n' % (self._z_content_length, self._content_length),
332
total_len, chunks = self.to_chunks()
319
333
return ''.join(chunks)
321
335
def _dump(self, include_text=False):
435
449
# Grab and cache the raw bytes for this entry
436
450
# and break the ref-cycle with _manager since we don't need it
438
self._manager._prepare_for_extract()
453
self._manager._prepare_for_extract()
454
except zlib.error as value:
455
raise errors.DecompressCorruption("zlib: " + str(value))
439
456
block = self._manager._block
440
457
self._bytes = block.extract(self.key, self._start, self._end)
441
458
# There are code paths that first extract as fulltext, and then
460
477
_full_enough_block_size = 3*1024*1024 # size at which we won't repack
461
478
_full_enough_mixed_block_size = 2*768*1024 # 1.5MB
463
def __init__(self, block):
480
def __init__(self, block, get_compressor_settings=None):
464
481
self._block = block
465
482
# We need to preserve the ordering
466
483
self._factories = []
467
484
self._last_byte = 0
485
self._get_settings = get_compressor_settings
486
self._compressor_settings = None
488
def _get_compressor_settings(self):
489
if self._compressor_settings is not None:
490
return self._compressor_settings
492
if self._get_settings is not None:
493
settings = self._get_settings()
495
vf = GroupCompressVersionedFiles
496
settings = vf._DEFAULT_COMPRESSOR_SETTINGS
497
self._compressor_settings = settings
498
return self._compressor_settings
469
500
def add_factory(self, key, parents, start, end):
470
501
if not self._factories:
503
534
new_block.set_content(self._block._content[:last_byte])
504
535
self._block = new_block
537
def _make_group_compressor(self):
538
return GroupCompressor(self._get_compressor_settings())
506
540
def _rebuild_block(self):
507
541
"""Create a new GroupCompressBlock with only the referenced texts."""
508
compressor = GroupCompressor()
542
compressor = self._make_group_compressor()
509
543
tstart = time.time()
510
544
old_length = self._block._content_length
523
557
# block? It seems hard to come up with a method that it would
524
558
# expand, since we do full compression again. Perhaps based on a
525
559
# request that ends up poorly ordered?
560
# TODO: If the content would have expanded, then we would want to
561
# handle a case where we need to split the block.
562
# Now that we have a user-tweakable option
563
# (max_bytes_to_index), it is possible that one person set it
564
# to a very low value, causing poor compression.
526
565
delta = time.time() - tstart
527
566
self._block = new_block
528
567
trace.mutter('creating new compressed block on-the-fly in %.3fs'
679
718
z_header_bytes = zlib.compress(header_bytes)
681
720
z_header_bytes_len = len(z_header_bytes)
682
block_bytes = self._block.to_bytes()
721
block_bytes_len, block_chunks = self._block.to_chunks()
683
722
lines.append('%d\n%d\n%d\n' % (z_header_bytes_len, header_bytes_len,
685
724
lines.append(z_header_bytes)
686
lines.append(block_bytes)
687
del z_header_bytes, block_bytes
725
lines.extend(block_chunks)
726
del z_header_bytes, block_chunks
727
# TODO: This is a point where we will double the memory consumption. To
728
# avoid this, we probably have to switch to a 'chunked' api
688
729
return ''.join(lines)
691
732
def from_bytes(cls, bytes):
692
733
# TODO: This does extra string copying, probably better to do it a
734
# different way. At a minimum this creates 2 copies of the
694
736
(storage_kind, z_header_len, header_len,
695
737
block_len, rest) = bytes.split('\n', 4)
855
901
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.
865
903
self._block.set_chunked_content(self.chunks, self.endpoint)
866
904
self.chunks = None
867
905
self._delta_index = None
886
924
class PythonGroupCompressor(_CommonGroupCompressor):
926
def __init__(self, settings=None):
889
927
"""Create a GroupCompressor.
891
929
Used only if the pyrex version is not available.
893
super(PythonGroupCompressor, self).__init__()
931
super(PythonGroupCompressor, self).__init__(settings)
894
932
self._delta_index = LinesDeltaIndex([])
895
933
# The actual content is managed by LinesDeltaIndex
896
934
self.chunks = self._delta_index.lines
934
972
It contains code very similar to SequenceMatcher because of having a similar
935
973
task. However some key differences apply:
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
975
* there is no junk, we want a minimal edit not a human readable diff.
976
* we don't filter very common lines (because we don't know where a good
977
range will start, and after the first text we want to be emitting minmal
979
* we chain the left side, not the right side
980
* we incrementally update the adjacency matrix as new lines are provided.
981
* we look for matches in all of the left side, so the routine which does
982
the analagous task of find_longest_match does not need to filter on the
948
super(PyrexGroupCompressor, self).__init__()
949
self._delta_index = DeltaIndex()
986
def __init__(self, settings=None):
987
super(PyrexGroupCompressor, self).__init__(settings)
988
max_bytes_to_index = self._settings.get('max_bytes_to_index', 0)
989
self._delta_index = DeltaIndex(max_bytes_to_index=max_bytes_to_index)
951
991
def _compress(self, key, bytes, max_delta_size, soft=False):
952
992
"""see _CommonGroupCompressor._compress"""
1027
1067
index = _GCGraphIndex(graph_index, lambda:True, parents=parents,
1028
1068
add_callback=graph_index.add_nodes,
1029
1069
inconsistency_fatal=inconsistency_fatal)
1030
access = knit._DirectPackAccess({})
1070
access = pack_repo._DirectPackAccess({})
1031
1071
access.set_writer(writer, graph_index, (transport, 'newpack'))
1032
1072
result = GroupCompressVersionedFiles(index, access, delta)
1033
1073
result.stream = stream
1044
1084
class _BatchingBlockFetcher(object):
1045
1085
"""Fetch group compress blocks in batches.
1047
1087
:ivar total_bytes: int of expected number of bytes needed to fetch the
1048
1088
currently pending batch.
1051
def __init__(self, gcvf, locations):
1091
def __init__(self, gcvf, locations, get_compressor_settings=None):
1052
1092
self.gcvf = gcvf
1053
1093
self.locations = locations
1102
1143
def yield_factories(self, full_flush=False):
1103
1144
"""Yield factories for keys added since the last yield. They will be
1104
1145
returned in the order they were added via add_key.
1106
1147
:param full_flush: by default, some results may not be returned in case
1107
1148
they can be part of the next batch. If full_flush is True, then
1108
1149
all results are returned.
1136
1177
memos_to_get_stack.pop()
1138
1179
block = self.batch_memos[read_memo]
1139
self.manager = _LazyGroupContentManager(block)
1180
self.manager = _LazyGroupContentManager(block,
1181
get_compressor_settings=self._get_compressor_settings)
1140
1182
self.last_read_memo = read_memo
1141
1183
start, end = index_memo[3:5]
1142
1184
self.manager.add_factory(key, parents, start, end)
1149
1191
self.total_bytes = 0
1152
class GroupCompressVersionedFiles(VersionedFiles):
1194
class GroupCompressVersionedFiles(VersionedFilesWithFallbacks):
1153
1195
"""A group-compress based VersionedFiles implementation."""
1155
def __init__(self, index, access, delta=True, _unadded_refs=None):
1197
# This controls how the GroupCompress DeltaIndex works. Basically, we
1198
# compute hash pointers into the source blocks (so hash(text) => text).
1199
# However each of these references costs some memory in trade against a
1200
# more accurate match result. For very large files, they either are
1201
# pre-compressed and change in bulk whenever they change, or change in just
1202
# local blocks. Either way, 'improved resolution' is not very helpful,
1203
# versus running out of memory trying to track everything. The default max
1204
# gives 100% sampling of a 1MB file.
1205
_DEFAULT_MAX_BYTES_TO_INDEX = 1024 * 1024
1206
_DEFAULT_COMPRESSOR_SETTINGS = {'max_bytes_to_index':
1207
_DEFAULT_MAX_BYTES_TO_INDEX}
1209
def __init__(self, index, access, delta=True, _unadded_refs=None,
1156
1211
"""Create a GroupCompressVersionedFiles object.
1158
1213
:param index: The index object storing access and graph data.
1159
1214
:param access: The access object storing raw data.
1160
1215
:param delta: Whether to delta compress or just entropy compress.
1161
1216
:param _unadded_refs: private parameter, don't use.
1217
:param _group_cache: private parameter, don't use.
1163
1219
self._index = index
1164
1220
self._access = access
1166
1222
if _unadded_refs is None:
1167
1223
_unadded_refs = {}
1168
1224
self._unadded_refs = _unadded_refs
1169
self._group_cache = LRUSizeCache(max_size=50*1024*1024)
1170
self._fallback_vfs = []
1225
if _group_cache is None:
1226
_group_cache = LRUSizeCache(max_size=50*1024*1024)
1227
self._group_cache = _group_cache
1228
self._immediate_fallback_vfs = []
1229
self._max_bytes_to_index = None
1172
1231
def without_fallbacks(self):
1173
1232
"""Return a clone of this object without any fallbacks configured."""
1174
1233
return GroupCompressVersionedFiles(self._index, self._access,
1175
self._delta, _unadded_refs=dict(self._unadded_refs))
1234
self._delta, _unadded_refs=dict(self._unadded_refs),
1235
_group_cache=self._group_cache)
1177
1237
def add_lines(self, key, parents, lines, parent_texts=None,
1178
1238
left_matching_blocks=None, nostore_sha=None, random_id=False,
1182
1242
:param key: The key tuple of the text to add.
1183
1243
:param parents: The parents key tuples of the text to add.
1184
1244
:param lines: A list of lines. Each line must be a bytestring. And all
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
1245
of them except the last must be terminated with \\n and contain no
1246
other \\n's. The last line may either contain no \\n's or a single
1247
terminating \\n. If the lines list does meet this constraint the
1248
add routine may error or may succeed - but you will be unable to
1249
read the data back accurately. (Checking the lines have been split
1190
1250
correctly is expensive and extremely unlikely to catch bugs so it
1191
1251
is not done at runtime unless check_content is True.)
1192
1252
:param parent_texts: An optional dictionary containing the opaque
1287
1347
self._check_lines_not_unicode(lines)
1288
1348
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)
1306
1350
def get_parent_map(self, keys):
1307
1351
"""Get a map of the graph parents of keys.
1366
1410
yield read_memo, cached[read_memo]
1367
1411
except KeyError:
1368
1412
# Read the block, and cache it.
1369
zdata = raw_records.next()
1413
zdata = next(raw_records)
1370
1414
block = GroupCompressBlock.from_bytes(zdata)
1371
1415
self._group_cache[read_memo] = block
1372
1416
cached[read_memo] = block
1447
1491
The returned objects should be in the order defined by 'ordering',
1448
1492
which can weave between different sources.
1449
1494
:param ordering: Must be one of 'topological' or 'groupcompress'
1450
1495
:return: List of [(source, [keys])] tuples, such that all keys are in
1451
1496
the defined order, regardless of source.
1453
1498
if ordering == 'topological':
1454
present_keys = topo_sort(parent_map)
1499
present_keys = tsort.topo_sort(parent_map)
1456
1501
# ordering == 'groupcompress'
1457
1502
# XXX: This only optimizes for the target ordering. We may need
1546
1591
# - we encounter an unadded ref, or
1547
1592
# - we run out of keys, or
1548
1593
# - the total bytes to retrieve for this batch > BATCH_SIZE
1549
batcher = _BatchingBlockFetcher(self, locations)
1594
batcher = _BatchingBlockFetcher(self, locations,
1595
get_compressor_settings=self._get_compressor_settings)
1550
1596
for source, keys in source_keys:
1551
1597
if source is self:
1552
1598
for key in keys:
1598
1644
for _ in self._insert_record_stream(stream, random_id=False):
1647
def _get_compressor_settings(self):
1648
if self._max_bytes_to_index is None:
1649
# TODO: VersionedFiles don't know about their containing
1650
# repository, so they don't have much of an idea about their
1651
# location. So for now, this is only a global option.
1652
c = config.GlobalConfig()
1653
val = c.get_user_option('bzr.groupcompress.max_bytes_to_index')
1657
except ValueError as e:
1658
trace.warning('Value for '
1659
'"bzr.groupcompress.max_bytes_to_index"'
1660
' %r is not an integer'
1664
val = self._DEFAULT_MAX_BYTES_TO_INDEX
1665
self._max_bytes_to_index = val
1666
return {'max_bytes_to_index': self._max_bytes_to_index}
1668
def _make_group_compressor(self):
1669
return GroupCompressor(self._get_compressor_settings())
1601
1671
def _insert_record_stream(self, stream, random_id=False, nostore_sha=None,
1602
1672
reuse_blocks=True):
1603
1673
"""Internal core to insert a record stream into this container.
1627
1697
# This will go up to fulltexts for gc to gc fetching, which isn't
1629
self._compressor = GroupCompressor()
1699
self._compressor = self._make_group_compressor()
1630
1700
self._unadded_refs = {}
1631
1701
keys_to_add = []
1633
bytes = self._compressor.flush().to_bytes()
1634
self._compressor = GroupCompressor()
1703
bytes_len, chunks = self._compressor.flush().to_chunks()
1704
self._compressor = self._make_group_compressor()
1705
# Note: At this point we still have 1 copy of the fulltext (in
1706
# record and the var 'bytes'), and this generates 2 copies of
1707
# the compressed text (one for bytes, one in chunks)
1708
# TODO: Push 'chunks' down into the _access api, so that we don't
1709
# have to double compressed memory here
1710
# TODO: Figure out how to indicate that we would be happy to free
1711
# the fulltext content at this point. Note that sometimes we
1712
# will want it later (streaming CHK pages), but most of the
1713
# time we won't (everything else)
1714
bytes = ''.join(chunks)
1635
1716
index, start, length = self._access.add_raw_records(
1636
1717
[(None, len(bytes))], bytes)[0]
1656
1737
raise errors.RevisionNotPresent(record.key, self)
1658
1739
if record.key in inserted_keys:
1659
trace.note('Insert claimed random_id=True,'
1660
' but then inserted %r two times', record.key)
1740
trace.note(gettext('Insert claimed random_id=True,'
1741
' but then inserted %r two times'), record.key)
1662
1743
inserted_keys.add(record.key)
1663
1744
if reuse_blocks:
1802
1883
"""See VersionedFiles.keys."""
1803
1884
if 'evil' in debug.debug_flags:
1804
1885
trace.mutter_callsite(2, "keys scales with size of history")
1805
sources = [self._index] + self._fallback_vfs
1886
sources = [self._index] + self._immediate_fallback_vfs
1807
1888
for source in sources:
1808
1889
result.update(source.keys())
1893
class _GCBuildDetails(object):
1894
"""A blob of data about the build details.
1896
This stores the minimal data, which then allows compatibility with the old
1897
api, without taking as much memory.
1900
__slots__ = ('_index', '_group_start', '_group_end', '_basis_end',
1901
'_delta_end', '_parents')
1904
compression_parent = None
1906
def __init__(self, parents, position_info):
1907
self._parents = parents
1908
(self._index, self._group_start, self._group_end, self._basis_end,
1909
self._delta_end) = position_info
1912
return '%s(%s, %s)' % (self.__class__.__name__,
1913
self.index_memo, self._parents)
1916
def index_memo(self):
1917
return (self._index, self._group_start, self._group_end,
1918
self._basis_end, self._delta_end)
1921
def record_details(self):
1922
return static_tuple.StaticTuple(self.method, None)
1924
def __getitem__(self, offset):
1925
"""Compatibility thunk to act like a tuple."""
1927
return self.index_memo
1929
return self.compression_parent # Always None
1931
return self._parents
1933
return self.record_details
1935
raise IndexError('offset out of range')
1812
1941
class _GCGraphIndex(object):
1813
1942
"""Mapper from GroupCompressVersionedFiles needs into GraphIndex storage."""
1817
1946
inconsistency_fatal=True, track_new_keys=False):
1818
1947
"""Construct a _GCGraphIndex on a graph_index.
1820
:param graph_index: An implementation of bzrlib.index.GraphIndex.
1949
:param graph_index: An implementation of breezy.index.GraphIndex.
1821
1950
:param is_locked: A callback, returns True if the index is locked and
1823
1952
:param parents: If True, record knits parents, if not do not record
1989
2118
:param keys: An iterable of keys.
1990
2119
:return: A dict of key:
1991
2120
(index_memo, compression_parent, parents, record_details).
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
2122
* index_memo: opaque structure to pass to read_records to extract
2124
* compression_parent: Content that this record is built upon, may
2126
* parents: Logical parents of this node
2127
* record_details: extra information about the content which needs
2128
to be passed to Factory.parse_record
2003
2130
self._check_read()
2033
2159
# each, or about 7MB. Note that it might be even more when you consider
2034
2160
# how PyInt is allocated in separate slabs. And you can't return a slab
2035
2161
# to the OS if even 1 int on it is in use. Note though that Python uses
2036
# a LIFO when re-using PyInt slots, which probably causes more
2162
# a LIFO when re-using PyInt slots, which might cause more
2037
2163
# fragmentation.
2038
2164
start = int(bits[0])
2039
2165
start = self._int_cache.setdefault(start, start)
added
removed
breezy/groupcompress.py
