22
from bzrlib.lazy_import import lazy_import
23
lazy_import(globals(), """
24
26
from bzrlib import (
29
30
graph as _mod_graph,
37
from bzrlib.repofmt import pack_repo
38
from bzrlib.i18n import gettext
41
37
from bzrlib.btree_index import BTreeBuilder
42
38
from bzrlib.lru_cache import LRUSizeCache
39
from bzrlib.tsort import topo_sort
43
40
from bzrlib.versionedfile import (
46
42
AbsentContentFactory,
47
43
ChunkedContentFactory,
48
44
FulltextContentFactory,
49
VersionedFilesWithFallbacks,
52
48
# Minimum number of uncompressed bytes to try fetch at once when retrieving
53
49
# groupcompress blocks.
52
_USE_LZMA = False and (pylzma is not None)
56
54
# osutils.sha_string('')
57
55
_null_sha1 = 'da39a3ee5e6b4b0d3255bfef95601890afd80709'
137
135
self._content = ''.join(self._content_chunks)
138
136
self._content_chunks = None
139
137
if self._content is None:
140
# We join self._z_content_chunks here, because if we are
141
# decompressing, then it is *very* likely that we have a single
143
if self._z_content_chunks is None:
138
if self._z_content is None:
144
139
raise AssertionError('No content to decompress')
145
z_content = ''.join(self._z_content_chunks)
140
if self._z_content == '':
147
141
self._content = ''
148
142
elif self._compressor_name == 'lzma':
149
143
# We don't do partial lzma decomp yet
151
self._content = pylzma.decompress(z_content)
144
self._content = pylzma.decompress(self._z_content)
152
145
elif self._compressor_name == 'zlib':
153
146
# Start a zlib decompressor
154
147
if num_bytes * 4 > self._content_length * 3:
155
148
# If we are requesting more that 3/4ths of the content,
156
149
# just extract the whole thing in a single pass
157
150
num_bytes = self._content_length
158
self._content = zlib.decompress(z_content)
151
self._content = zlib.decompress(self._z_content)
160
153
self._z_content_decompressor = zlib.decompressobj()
161
154
# Seed the decompressor with the uncompressed bytes, so
162
155
# that the rest of the code is simplified
163
156
self._content = self._z_content_decompressor.decompress(
164
z_content, num_bytes + _ZLIB_DECOMP_WINDOW)
157
self._z_content, num_bytes + _ZLIB_DECOMP_WINDOW)
165
158
if not self._z_content_decompressor.unconsumed_tail:
166
159
self._z_content_decompressor = None
214
207
# XXX: Define some GCCorrupt error ?
215
208
raise AssertionError('Invalid bytes: (%d) != %d + %d' %
216
209
(len(bytes), pos, self._z_content_length))
217
self._z_content_chunks = (bytes[pos:],)
220
def _z_content(self):
221
"""Return z_content_chunks as a simple string.
223
Meant only to be used by the test suite.
225
if self._z_content_chunks is not None:
226
return ''.join(self._z_content_chunks)
210
self._z_content = bytes[pos:]
230
213
def from_bytes(cls, bytes):
286
269
self._content_length = length
287
270
self._content_chunks = content_chunks
288
271
self._content = None
289
self._z_content_chunks = None
272
self._z_content = None
291
274
def set_content(self, content):
292
275
"""Set the content of this block."""
293
276
self._content_length = len(content)
294
277
self._content = content
295
self._z_content_chunks = None
297
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):
298
290
compressor = zlib.compressobj(zlib.Z_DEFAULT_COMPRESSION)
299
# Peak in this point is 1 fulltext, 1 compressed text, + zlib overhead
300
# (measured peak is maybe 30MB over the above...)
301
compressed_chunks = map(compressor.compress, chunks)
291
compressed_chunks = map(compressor.compress, self._content_chunks)
302
292
compressed_chunks.append(compressor.flush())
303
# Ignore empty chunks
304
self._z_content_chunks = [c for c in compressed_chunks if c]
305
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)
307
296
def _create_z_content(self):
308
if self._z_content_chunks is not None:
297
if self._z_content is not None:
300
self._create_z_content_using_lzma()
310
302
if self._content_chunks is not None:
311
chunks = self._content_chunks
313
chunks = (self._content,)
314
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)
317
"""Create the byte stream as a series of 'chunks'"""
309
"""Encode the information into a byte stream."""
318
310
self._create_z_content()
319
header = self.GCB_HEADER
320
chunks = ['%s%d\n%d\n'
321
% (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),
323
chunks.extend(self._z_content_chunks)
324
total_len = sum(map(len, chunks))
325
return total_len, chunks
328
"""Encode the information into a byte stream."""
329
total_len, chunks = self.to_chunks()
330
319
return ''.join(chunks)
332
321
def _dump(self, include_text=False):
446
435
# Grab and cache the raw bytes for this entry
447
436
# and break the ref-cycle with _manager since we don't need it
450
self._manager._prepare_for_extract()
451
except zlib.error as value:
452
raise errors.DecompressCorruption("zlib: " + str(value))
438
self._manager._prepare_for_extract()
453
439
block = self._manager._block
454
440
self._bytes = block.extract(self.key, self._start, self._end)
455
441
# There are code paths that first extract as fulltext, and then
474
460
_full_enough_block_size = 3*1024*1024 # size at which we won't repack
475
461
_full_enough_mixed_block_size = 2*768*1024 # 1.5MB
477
def __init__(self, block, get_compressor_settings=None):
463
def __init__(self, block):
478
464
self._block = block
479
465
# We need to preserve the ordering
480
466
self._factories = []
481
467
self._last_byte = 0
482
self._get_settings = get_compressor_settings
483
self._compressor_settings = None
485
def _get_compressor_settings(self):
486
if self._compressor_settings is not None:
487
return self._compressor_settings
489
if self._get_settings is not None:
490
settings = self._get_settings()
492
vf = GroupCompressVersionedFiles
493
settings = vf._DEFAULT_COMPRESSOR_SETTINGS
494
self._compressor_settings = settings
495
return self._compressor_settings
497
469
def add_factory(self, key, parents, start, end):
498
470
if not self._factories:
531
503
new_block.set_content(self._block._content[:last_byte])
532
504
self._block = new_block
534
def _make_group_compressor(self):
535
return GroupCompressor(self._get_compressor_settings())
537
506
def _rebuild_block(self):
538
507
"""Create a new GroupCompressBlock with only the referenced texts."""
539
compressor = self._make_group_compressor()
508
compressor = GroupCompressor()
540
509
tstart = time.time()
541
510
old_length = self._block._content_length
554
523
# block? It seems hard to come up with a method that it would
555
524
# expand, since we do full compression again. Perhaps based on a
556
525
# request that ends up poorly ordered?
557
# TODO: If the content would have expanded, then we would want to
558
# handle a case where we need to split the block.
559
# Now that we have a user-tweakable option
560
# (max_bytes_to_index), it is possible that one person set it
561
# to a very low value, causing poor compression.
562
526
delta = time.time() - tstart
563
527
self._block = new_block
564
528
trace.mutter('creating new compressed block on-the-fly in %.3fs'
715
679
z_header_bytes = zlib.compress(header_bytes)
717
681
z_header_bytes_len = len(z_header_bytes)
718
block_bytes_len, block_chunks = self._block.to_chunks()
682
block_bytes = self._block.to_bytes()
719
683
lines.append('%d\n%d\n%d\n' % (z_header_bytes_len, header_bytes_len,
721
685
lines.append(z_header_bytes)
722
lines.extend(block_chunks)
723
del z_header_bytes, block_chunks
724
# TODO: This is a point where we will double the memory consumption. To
725
# avoid this, we probably have to switch to a 'chunked' api
686
lines.append(block_bytes)
687
del z_header_bytes, block_bytes
726
688
return ''.join(lines)
729
691
def from_bytes(cls, bytes):
730
692
# TODO: This does extra string copying, probably better to do it a
731
# different way. At a minimum this creates 2 copies of the
733
694
(storage_kind, z_header_len, header_len,
734
695
block_len, rest) = bytes.split('\n', 4)
898
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.
900
865
self._block.set_chunked_content(self.chunks, self.endpoint)
901
866
self.chunks = None
902
867
self._delta_index = None
921
886
class PythonGroupCompressor(_CommonGroupCompressor):
923
def __init__(self, settings=None):
924
889
"""Create a GroupCompressor.
926
891
Used only if the pyrex version is not available.
928
super(PythonGroupCompressor, self).__init__(settings)
893
super(PythonGroupCompressor, self).__init__()
929
894
self._delta_index = LinesDeltaIndex([])
930
895
# The actual content is managed by LinesDeltaIndex
931
896
self.chunks = self._delta_index.lines
969
934
It contains code very similar to SequenceMatcher because of having a similar
970
935
task. However some key differences apply:
972
* there is no junk, we want a minimal edit not a human readable diff.
973
* we don't filter very common lines (because we don't know where a good
974
range will start, and after the first text we want to be emitting minmal
976
* we chain the left side, not the right side
977
* we incrementally update the adjacency matrix as new lines are provided.
978
* we look for matches in all of the left side, so the routine which does
979
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
983
def __init__(self, settings=None):
984
super(PyrexGroupCompressor, self).__init__(settings)
985
max_bytes_to_index = self._settings.get('max_bytes_to_index', 0)
986
self._delta_index = DeltaIndex(max_bytes_to_index=max_bytes_to_index)
948
super(PyrexGroupCompressor, self).__init__()
949
self._delta_index = DeltaIndex()
988
951
def _compress(self, key, bytes, max_delta_size, soft=False):
989
952
"""see _CommonGroupCompressor._compress"""
1064
1027
index = _GCGraphIndex(graph_index, lambda:True, parents=parents,
1065
1028
add_callback=graph_index.add_nodes,
1066
1029
inconsistency_fatal=inconsistency_fatal)
1067
access = pack_repo._DirectPackAccess({})
1030
access = knit._DirectPackAccess({})
1068
1031
access.set_writer(writer, graph_index, (transport, 'newpack'))
1069
1032
result = GroupCompressVersionedFiles(index, access, delta)
1070
1033
result.stream = stream
1081
1044
class _BatchingBlockFetcher(object):
1082
1045
"""Fetch group compress blocks in batches.
1084
1047
:ivar total_bytes: int of expected number of bytes needed to fetch the
1085
1048
currently pending batch.
1088
def __init__(self, gcvf, locations, get_compressor_settings=None):
1051
def __init__(self, gcvf, locations):
1089
1052
self.gcvf = gcvf
1090
1053
self.locations = locations
1174
1136
memos_to_get_stack.pop()
1176
1138
block = self.batch_memos[read_memo]
1177
self.manager = _LazyGroupContentManager(block,
1178
get_compressor_settings=self._get_compressor_settings)
1139
self.manager = _LazyGroupContentManager(block)
1179
1140
self.last_read_memo = read_memo
1180
1141
start, end = index_memo[3:5]
1181
1142
self.manager.add_factory(key, parents, start, end)
1188
1149
self.total_bytes = 0
1191
class GroupCompressVersionedFiles(VersionedFilesWithFallbacks):
1152
class GroupCompressVersionedFiles(VersionedFiles):
1192
1153
"""A group-compress based VersionedFiles implementation."""
1194
# This controls how the GroupCompress DeltaIndex works. Basically, we
1195
# compute hash pointers into the source blocks (so hash(text) => text).
1196
# However each of these references costs some memory in trade against a
1197
# more accurate match result. For very large files, they either are
1198
# pre-compressed and change in bulk whenever they change, or change in just
1199
# local blocks. Either way, 'improved resolution' is not very helpful,
1200
# versus running out of memory trying to track everything. The default max
1201
# gives 100% sampling of a 1MB file.
1202
_DEFAULT_MAX_BYTES_TO_INDEX = 1024 * 1024
1203
_DEFAULT_COMPRESSOR_SETTINGS = {'max_bytes_to_index':
1204
_DEFAULT_MAX_BYTES_TO_INDEX}
1206
def __init__(self, index, access, delta=True, _unadded_refs=None,
1155
def __init__(self, index, access, delta=True, _unadded_refs=None):
1208
1156
"""Create a GroupCompressVersionedFiles object.
1210
1158
:param index: The index object storing access and graph data.
1211
1159
:param access: The access object storing raw data.
1212
1160
:param delta: Whether to delta compress or just entropy compress.
1213
1161
:param _unadded_refs: private parameter, don't use.
1214
:param _group_cache: private parameter, don't use.
1216
1163
self._index = index
1217
1164
self._access = access
1219
1166
if _unadded_refs is None:
1220
1167
_unadded_refs = {}
1221
1168
self._unadded_refs = _unadded_refs
1222
if _group_cache is None:
1223
_group_cache = LRUSizeCache(max_size=50*1024*1024)
1224
self._group_cache = _group_cache
1225
self._immediate_fallback_vfs = []
1226
self._max_bytes_to_index = None
1169
self._group_cache = LRUSizeCache(max_size=50*1024*1024)
1170
self._fallback_vfs = []
1228
1172
def without_fallbacks(self):
1229
1173
"""Return a clone of this object without any fallbacks configured."""
1230
1174
return GroupCompressVersionedFiles(self._index, self._access,
1231
self._delta, _unadded_refs=dict(self._unadded_refs),
1232
_group_cache=self._group_cache)
1175
self._delta, _unadded_refs=dict(self._unadded_refs))
1234
1177
def add_lines(self, key, parents, lines, parent_texts=None,
1235
1178
left_matching_blocks=None, nostore_sha=None, random_id=False,
1239
1182
:param key: The key tuple of the text to add.
1240
1183
:param parents: The parents key tuples of the text to add.
1241
1184
:param lines: A list of lines. Each line must be a bytestring. And all
1242
of them except the last must be terminated with \\n and contain no
1243
other \\n's. The last line may either contain no \\n's or a single
1244
terminating \\n. If the lines list does meet this constraint the
1245
add routine may error or may succeed - but you will be unable to
1246
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
1247
1190
correctly is expensive and extremely unlikely to catch bugs so it
1248
1191
is not done at runtime unless check_content is True.)
1249
1192
:param parent_texts: An optional dictionary containing the opaque
1344
1287
self._check_lines_not_unicode(lines)
1345
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)
1347
1306
def get_parent_map(self, keys):
1348
1307
"""Get a map of the graph parents of keys.
1488
1447
The returned objects should be in the order defined by 'ordering',
1489
1448
which can weave between different sources.
1491
1449
:param ordering: Must be one of 'topological' or 'groupcompress'
1492
1450
:return: List of [(source, [keys])] tuples, such that all keys are in
1493
1451
the defined order, regardless of source.
1495
1453
if ordering == 'topological':
1496
present_keys = tsort.topo_sort(parent_map)
1454
present_keys = topo_sort(parent_map)
1498
1456
# ordering == 'groupcompress'
1499
1457
# XXX: This only optimizes for the target ordering. We may need
1588
1546
# - we encounter an unadded ref, or
1589
1547
# - we run out of keys, or
1590
1548
# - the total bytes to retrieve for this batch > BATCH_SIZE
1591
batcher = _BatchingBlockFetcher(self, locations,
1592
get_compressor_settings=self._get_compressor_settings)
1549
batcher = _BatchingBlockFetcher(self, locations)
1593
1550
for source, keys in source_keys:
1594
1551
if source is self:
1595
1552
for key in keys:
1641
1598
for _ in self._insert_record_stream(stream, random_id=False):
1644
def _get_compressor_settings(self):
1645
if self._max_bytes_to_index is None:
1646
# TODO: VersionedFiles don't know about their containing
1647
# repository, so they don't have much of an idea about their
1648
# location. So for now, this is only a global option.
1649
c = config.GlobalConfig()
1650
val = c.get_user_option('bzr.groupcompress.max_bytes_to_index')
1654
except ValueError, e:
1655
trace.warning('Value for '
1656
'"bzr.groupcompress.max_bytes_to_index"'
1657
' %r is not an integer'
1661
val = self._DEFAULT_MAX_BYTES_TO_INDEX
1662
self._max_bytes_to_index = val
1663
return {'max_bytes_to_index': self._max_bytes_to_index}
1665
def _make_group_compressor(self):
1666
return GroupCompressor(self._get_compressor_settings())
1668
1601
def _insert_record_stream(self, stream, random_id=False, nostore_sha=None,
1669
1602
reuse_blocks=True):
1670
1603
"""Internal core to insert a record stream into this container.
1694
1627
# This will go up to fulltexts for gc to gc fetching, which isn't
1696
self._compressor = self._make_group_compressor()
1629
self._compressor = GroupCompressor()
1697
1630
self._unadded_refs = {}
1698
1631
keys_to_add = []
1700
bytes_len, chunks = self._compressor.flush().to_chunks()
1701
self._compressor = self._make_group_compressor()
1702
# Note: At this point we still have 1 copy of the fulltext (in
1703
# record and the var 'bytes'), and this generates 2 copies of
1704
# the compressed text (one for bytes, one in chunks)
1705
# TODO: Push 'chunks' down into the _access api, so that we don't
1706
# have to double compressed memory here
1707
# TODO: Figure out how to indicate that we would be happy to free
1708
# the fulltext content at this point. Note that sometimes we
1709
# will want it later (streaming CHK pages), but most of the
1710
# time we won't (everything else)
1711
bytes = ''.join(chunks)
1633
bytes = self._compressor.flush().to_bytes()
1713
1634
index, start, length = self._access.add_raw_records(
1714
1635
[(None, len(bytes))], bytes)[0]
1734
1656
raise errors.RevisionNotPresent(record.key, self)
1736
1658
if record.key in inserted_keys:
1737
trace.note(gettext('Insert claimed random_id=True,'
1738
' 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)
1740
1662
inserted_keys.add(record.key)
1741
1663
if reuse_blocks:
1880
1802
"""See VersionedFiles.keys."""
1881
1803
if 'evil' in debug.debug_flags:
1882
1804
trace.mutter_callsite(2, "keys scales with size of history")
1883
sources = [self._index] + self._immediate_fallback_vfs
1805
sources = [self._index] + self._fallback_vfs
1885
1807
for source in sources:
1886
1808
result.update(source.keys())
1890
class _GCBuildDetails(object):
1891
"""A blob of data about the build details.
1893
This stores the minimal data, which then allows compatibility with the old
1894
api, without taking as much memory.
1897
__slots__ = ('_index', '_group_start', '_group_end', '_basis_end',
1898
'_delta_end', '_parents')
1901
compression_parent = None
1903
def __init__(self, parents, position_info):
1904
self._parents = parents
1905
(self._index, self._group_start, self._group_end, self._basis_end,
1906
self._delta_end) = position_info
1909
return '%s(%s, %s)' % (self.__class__.__name__,
1910
self.index_memo, self._parents)
1913
def index_memo(self):
1914
return (self._index, self._group_start, self._group_end,
1915
self._basis_end, self._delta_end)
1918
def record_details(self):
1919
return static_tuple.StaticTuple(self.method, None)
1921
def __getitem__(self, offset):
1922
"""Compatibility thunk to act like a tuple."""
1924
return self.index_memo
1926
return self.compression_parent # Always None
1928
return self._parents
1930
return self.record_details
1932
raise IndexError('offset out of range')
1938
1812
class _GCGraphIndex(object):
1939
1813
"""Mapper from GroupCompressVersionedFiles needs into GraphIndex storage."""
2115
1989
:param keys: An iterable of keys.
2116
1990
:return: A dict of key:
2117
1991
(index_memo, compression_parent, parents, record_details).
2119
* index_memo: opaque structure to pass to read_records to extract
2121
* compression_parent: Content that this record is built upon, may
2123
* parents: Logical parents of this node
2124
* record_details: extra information about the content which needs
2125
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
2127
2003
self._check_read()
2156
2033
# each, or about 7MB. Note that it might be even more when you consider
2157
2034
# how PyInt is allocated in separate slabs. And you can't return a slab
2158
2035
# to the OS if even 1 int on it is in use. Note though that Python uses
2159
# a LIFO when re-using PyInt slots, which might cause more
2036
# a LIFO when re-using PyInt slots, which probably causes more
2160
2037
# fragmentation.
2161
2038
start = int(bits[0])
2162
2039
start = self._int_cache.setdefault(start, start)
added
removed
bzrlib/groupcompress.py
