162
164
# 'unconsumed_tail'
164
166
# Do we have enough bytes already?
165
if num_bytes is not None and len(self._content) >= num_bytes:
167
if num_bytes is None and self._z_content_decompressor is None:
168
# We must have already decompressed everything
167
if len(self._content) >= num_bytes:
170
169
# If we got this far, and don't have a decompressor, something is wrong
171
170
if self._z_content_decompressor is None:
172
171
raise AssertionError(
173
172
'No decompressor to decompress %d bytes' % num_bytes)
174
173
remaining_decomp = self._z_content_decompressor.unconsumed_tail
175
if num_bytes is None:
177
# We don't know how much is left, but we'll decompress it all
178
self._content += self._z_content_decompressor.decompress(
180
# Note: There's what I consider a bug in zlib.decompressobj
181
# If you pass back in the entire unconsumed_tail, only
182
# this time you don't pass a max-size, it doesn't
183
# change the unconsumed_tail back to None/''.
184
# However, we know we are done with the whole stream
185
self._z_content_decompressor = None
186
# XXX: Why is this the only place in this routine we set this?
187
self._content_length = len(self._content)
189
if not remaining_decomp:
190
raise AssertionError('Nothing left to decompress')
191
needed_bytes = num_bytes - len(self._content)
192
# We always set max_size to 32kB over the minimum needed, so that
193
# zlib will give us as much as we really want.
194
# TODO: If this isn't good enough, we could make a loop here,
195
# that keeps expanding the request until we get enough
196
self._content += self._z_content_decompressor.decompress(
197
remaining_decomp, needed_bytes + _ZLIB_DECOMP_WINDOW)
198
if len(self._content) < num_bytes:
199
raise AssertionError('%d bytes wanted, only %d available'
200
% (num_bytes, len(self._content)))
201
if not self._z_content_decompressor.unconsumed_tail:
202
# The stream is finished
203
self._z_content_decompressor = None
174
if not remaining_decomp:
175
raise AssertionError('Nothing left to decompress')
176
needed_bytes = num_bytes - len(self._content)
177
# We always set max_size to 32kB over the minimum needed, so that
178
# zlib will give us as much as we really want.
179
# TODO: If this isn't good enough, we could make a loop here,
180
# that keeps expanding the request until we get enough
181
self._content += self._z_content_decompressor.decompress(
182
remaining_decomp, needed_bytes + _ZLIB_DECOMP_WINDOW)
183
if len(self._content) < num_bytes:
184
raise AssertionError('%d bytes wanted, only %d available'
185
% (num_bytes, len(self._content)))
186
if not self._z_content_decompressor.unconsumed_tail:
187
# The stream is finished
188
self._z_content_decompressor = None
205
190
def _parse_bytes(self, bytes, pos):
206
191
"""Read the various lengths from the header.
545
537
# time (self._block._content) is a little expensive.
546
538
self._block._ensure_content(self._last_byte)
548
def _check_rebuild_block(self):
540
def _check_rebuild_action(self):
549
541
"""Check to see if our block should be repacked."""
550
542
total_bytes_used = 0
551
543
last_byte_used = 0
552
544
for factory in self._factories:
553
545
total_bytes_used += factory._end - factory._start
554
last_byte_used = max(last_byte_used, factory._end)
555
# If we are using most of the bytes from the block, we have nothing
556
# else to check (currently more that 1/2)
546
if last_byte_used < factory._end:
547
last_byte_used = factory._end
548
# If we are using more than half of the bytes from the block, we have
549
# nothing else to check
557
550
if total_bytes_used * 2 >= self._block._content_length:
559
# Can we just strip off the trailing bytes? If we are going to be
560
# transmitting more than 50% of the front of the content, go ahead
551
return None, last_byte_used, total_bytes_used
552
# We are using less than 50% of the content. Is the content we are
553
# using at the beginning of the block? If so, we can just trim the
554
# tail, rather than rebuilding from scratch.
561
555
if total_bytes_used * 2 > last_byte_used:
562
self._trim_block(last_byte_used)
556
return 'trim', last_byte_used, total_bytes_used
565
558
# We are using a small amount of the data, and it isn't just packed
566
559
# nicely at the front, so rebuild the content.
573
566
# expanding many deltas into fulltexts, as well.
574
567
# If we build a cheap enough 'strip', then we could try a strip,
575
568
# if that expands the content, we then rebuild.
576
self._rebuild_block()
569
return 'rebuild', last_byte_used, total_bytes_used
571
def check_is_well_utilized(self):
572
"""Is the current block considered 'well utilized'?
574
This heuristic asks if the current block considers itself to be a fully
575
developed group, rather than just a loose collection of data.
577
if len(self._factories) == 1:
578
# A block of length 1 could be improved by combining with other
579
# groups - don't look deeper. Even larger than max size groups
580
# could compress well with adjacent versions of the same thing.
582
action, last_byte_used, total_bytes_used = self._check_rebuild_action()
583
block_size = self._block._content_length
584
if total_bytes_used < block_size * self._max_cut_fraction:
585
# This block wants to trim itself small enough that we want to
586
# consider it under-utilized.
588
# TODO: This code is meant to be the twin of _insert_record_stream's
589
# 'start_new_block' logic. It would probably be better to factor
590
# out that logic into a shared location, so that it stays
592
# We currently assume a block is properly utilized whenever it is >75%
593
# of the size of a 'full' block. In normal operation, a block is
594
# considered full when it hits 4MB of same-file content. So any block
595
# >3MB is 'full enough'.
596
# The only time this isn't true is when a given block has large-object
597
# content. (a single file >4MB, etc.)
598
# Under these circumstances, we allow a block to grow to
599
# 2 x largest_content. Which means that if a given block had a large
600
# object, it may actually be under-utilized. However, given that this
601
# is 'pack-on-the-fly' it is probably reasonable to not repack large
602
# content blobs on-the-fly. Note that because we return False for all
603
# 1-item blobs, we will repack them; we may wish to reevaluate our
604
# treatment of large object blobs in the future.
605
if block_size >= self._full_enough_block_size:
607
# If a block is <3MB, it still may be considered 'full' if it contains
608
# mixed content. The current rule is 2MB of mixed content is considered
609
# full. So check to see if this block contains mixed content, and
610
# set the threshold appropriately.
612
for factory in self._factories:
613
prefix = factory.key[:-1]
614
if common_prefix is None:
615
common_prefix = prefix
616
elif prefix != common_prefix:
617
# Mixed content, check the size appropriately
618
if block_size >= self._full_enough_mixed_block_size:
621
# The content failed both the mixed check and the single-content check
622
# so obviously it is not fully utilized
623
# TODO: there is one other constraint that isn't being checked
624
# namely, that the entries in the block are in the appropriate
625
# order. For example, you could insert the entries in exactly
626
# reverse groupcompress order, and we would think that is ok.
627
# (all the right objects are in one group, and it is fully
628
# utilized, etc.) For now, we assume that case is rare,
629
# especially since we should always fetch in 'groupcompress'
633
def _check_rebuild_block(self):
634
action, last_byte_used, total_bytes_used = self._check_rebuild_action()
638
self._trim_block(last_byte_used)
639
elif action == 'rebuild':
640
self._rebuild_block()
642
raise ValueError('unknown rebuild action: %r' % (action,))
578
644
def _wire_bytes(self):
579
645
"""Return a byte stream suitable for transmitting over the wire."""
975
1041
versioned_files.stream.close()
1044
class _BatchingBlockFetcher(object):
1045
"""Fetch group compress blocks in batches.
1047
:ivar total_bytes: int of expected number of bytes needed to fetch the
1048
currently pending batch.
1051
def __init__(self, gcvf, locations):
1053
self.locations = locations
1055
self.batch_memos = {}
1056
self.memos_to_get = []
1057
self.total_bytes = 0
1058
self.last_read_memo = None
1061
def add_key(self, key):
1062
"""Add another to key to fetch.
1064
:return: The estimated number of bytes needed to fetch the batch so
1067
self.keys.append(key)
1068
index_memo, _, _, _ = self.locations[key]
1069
read_memo = index_memo[0:3]
1070
# Three possibilities for this read_memo:
1071
# - it's already part of this batch; or
1072
# - it's not yet part of this batch, but is already cached; or
1073
# - it's not yet part of this batch and will need to be fetched.
1074
if read_memo in self.batch_memos:
1075
# This read memo is already in this batch.
1076
return self.total_bytes
1078
cached_block = self.gcvf._group_cache[read_memo]
1080
# This read memo is new to this batch, and the data isn't cached
1082
self.batch_memos[read_memo] = None
1083
self.memos_to_get.append(read_memo)
1084
byte_length = read_memo[2]
1085
self.total_bytes += byte_length
1087
# This read memo is new to this batch, but cached.
1088
# Keep a reference to the cached block in batch_memos because it's
1089
# certain that we'll use it when this batch is processed, but
1090
# there's a risk that it would fall out of _group_cache between now
1092
self.batch_memos[read_memo] = cached_block
1093
return self.total_bytes
1095
def _flush_manager(self):
1096
if self.manager is not None:
1097
for factory in self.manager.get_record_stream():
1100
self.last_read_memo = None
1102
def yield_factories(self, full_flush=False):
1103
"""Yield factories for keys added since the last yield. They will be
1104
returned in the order they were added via add_key.
1106
:param full_flush: by default, some results may not be returned in case
1107
they can be part of the next batch. If full_flush is True, then
1108
all results are returned.
1110
if self.manager is None and not self.keys:
1112
# Fetch all memos in this batch.
1113
blocks = self.gcvf._get_blocks(self.memos_to_get)
1114
# Turn blocks into factories and yield them.
1115
memos_to_get_stack = list(self.memos_to_get)
1116
memos_to_get_stack.reverse()
1117
for key in self.keys:
1118
index_memo, _, parents, _ = self.locations[key]
1119
read_memo = index_memo[:3]
1120
if self.last_read_memo != read_memo:
1121
# We are starting a new block. If we have a
1122
# manager, we have found everything that fits for
1123
# now, so yield records
1124
for factory in self._flush_manager():
1126
# Now start a new manager.
1127
if memos_to_get_stack and memos_to_get_stack[-1] == read_memo:
1128
# The next block from _get_blocks will be the block we
1130
block_read_memo, block = blocks.next()
1131
if block_read_memo != read_memo:
1132
raise AssertionError(
1133
"block_read_memo out of sync with read_memo"
1134
"(%r != %r)" % (block_read_memo, read_memo))
1135
self.batch_memos[read_memo] = block
1136
memos_to_get_stack.pop()
1138
block = self.batch_memos[read_memo]
1139
self.manager = _LazyGroupContentManager(block)
1140
self.last_read_memo = read_memo
1141
start, end = index_memo[3:5]
1142
self.manager.add_factory(key, parents, start, end)
1144
for factory in self._flush_manager():
1147
self.batch_memos.clear()
1148
del self.memos_to_get[:]
1149
self.total_bytes = 0
978
1152
class GroupCompressVersionedFiles(VersionedFiles):
979
1153
"""A group-compress based VersionedFiles implementation."""
981
def __init__(self, index, access, delta=True):
1155
def __init__(self, index, access, delta=True, _unadded_refs=None):
982
1156
"""Create a GroupCompressVersionedFiles object.
984
1158
:param index: The index object storing access and graph data.
985
1159
:param access: The access object storing raw data.
986
1160
:param delta: Whether to delta compress or just entropy compress.
1161
:param _unadded_refs: private parameter, don't use.
988
1163
self._index = index
989
1164
self._access = access
990
1165
self._delta = delta
991
self._unadded_refs = {}
1166
if _unadded_refs is None:
1168
self._unadded_refs = _unadded_refs
992
1169
self._group_cache = LRUSizeCache(max_size=50*1024*1024)
993
1170
self._fallback_vfs = []
1172
def without_fallbacks(self):
1173
"""Return a clone of this object without any fallbacks configured."""
1174
return GroupCompressVersionedFiles(self._index, self._access,
1175
self._delta, _unadded_refs=dict(self._unadded_refs))
995
1177
def add_lines(self, key, parents, lines, parent_texts=None,
996
1178
left_matching_blocks=None, nostore_sha=None, random_id=False,
997
1179
check_content=True):
1075
1257
def get_annotator(self):
1076
1258
return annotate.Annotator(self)
1078
def check(self, progress_bar=None):
1260
def check(self, progress_bar=None, keys=None):
1079
1261
"""See VersionedFiles.check()."""
1081
for record in self.get_record_stream(keys, 'unordered', True):
1082
record.get_bytes_as('fulltext')
1264
for record in self.get_record_stream(keys, 'unordered', True):
1265
record.get_bytes_as('fulltext')
1267
return self.get_record_stream(keys, 'unordered', True)
1269
def clear_cache(self):
1270
"""See VersionedFiles.clear_cache()"""
1271
self._group_cache.clear()
1272
self._index._graph_index.clear_cache()
1273
self._index._int_cache.clear()
1084
1275
def _check_add(self, key, lines, random_id, check_content):
1085
1276
"""check that version_id and lines are safe to add."""
1128
1335
missing.difference_update(set(new_result))
1129
1336
return result, source_results
1131
def _get_block(self, index_memo):
1132
read_memo = index_memo[0:3]
1135
block = self._group_cache[read_memo]
1138
zdata = self._access.get_raw_records([read_memo]).next()
1139
# decompress - whole thing - this is not a bug, as it
1140
# permits caching. We might want to store the partially
1141
# decompresed group and decompress object, so that recent
1142
# texts are not penalised by big groups.
1143
block = GroupCompressBlock.from_bytes(zdata)
1144
self._group_cache[read_memo] = block
1146
# print len(zdata), len(plain)
1147
# parse - requires split_lines, better to have byte offsets
1148
# here (but not by much - we only split the region for the
1149
# recipe, and we often want to end up with lines anyway.
1338
def _get_blocks(self, read_memos):
1339
"""Get GroupCompressBlocks for the given read_memos.
1341
:returns: a series of (read_memo, block) pairs, in the order they were
1345
for read_memo in read_memos:
1347
block = self._group_cache[read_memo]
1351
cached[read_memo] = block
1353
not_cached_seen = set()
1354
for read_memo in read_memos:
1355
if read_memo in cached:
1356
# Don't fetch what we already have
1358
if read_memo in not_cached_seen:
1359
# Don't try to fetch the same data twice
1361
not_cached.append(read_memo)
1362
not_cached_seen.add(read_memo)
1363
raw_records = self._access.get_raw_records(not_cached)
1364
for read_memo in read_memos:
1366
yield read_memo, cached[read_memo]
1368
# Read the block, and cache it.
1369
zdata = raw_records.next()
1370
block = GroupCompressBlock.from_bytes(zdata)
1371
self._group_cache[read_memo] = block
1372
cached[read_memo] = block
1373
yield read_memo, block
1152
1375
def get_missing_compression_parent_keys(self):
1153
1376
"""Return the keys of missing compression parents.
1319
1542
unadded_keys, source_result)
1320
1543
for key in missing:
1321
1544
yield AbsentContentFactory(key)
1323
last_read_memo = None
1324
# TODO: This works fairly well at batching up existing groups into a
1325
# streamable format, and possibly allowing for taking one big
1326
# group and splitting it when it isn't fully utilized.
1327
# However, it doesn't allow us to find under-utilized groups and
1328
# combine them into a bigger group on the fly.
1329
# (Consider the issue with how chk_map inserts texts
1330
# one-at-a-time.) This could be done at insert_record_stream()
1331
# time, but it probably would decrease the number of
1332
# bytes-on-the-wire for fetch.
1545
# Batch up as many keys as we can until either:
1546
# - we encounter an unadded ref, or
1547
# - we run out of keys, or
1548
# - the total bytes to retrieve for this batch > BATCH_SIZE
1549
batcher = _BatchingBlockFetcher(self, locations)
1333
1550
for source, keys in source_keys:
1334
1551
if source is self:
1335
1552
for key in keys:
1336
1553
if key in self._unadded_refs:
1337
if manager is not None:
1338
for factory in manager.get_record_stream():
1340
last_read_memo = manager = None
1554
# Flush batch, then yield unadded ref from
1556
for factory in batcher.yield_factories(full_flush=True):
1341
1558
bytes, sha1 = self._compressor.extract(key)
1342
1559
parents = self._unadded_refs[key]
1343
1560
yield FulltextContentFactory(key, parents, sha1, bytes)
1345
index_memo, _, parents, (method, _) = locations[key]
1346
read_memo = index_memo[0:3]
1347
if last_read_memo != read_memo:
1348
# We are starting a new block. If we have a
1349
# manager, we have found everything that fits for
1350
# now, so yield records
1351
if manager is not None:
1352
for factory in manager.get_record_stream():
1354
# Now start a new manager
1355
block = self._get_block(index_memo)
1356
manager = _LazyGroupContentManager(block)
1357
last_read_memo = read_memo
1358
start, end = index_memo[3:5]
1359
manager.add_factory(key, parents, start, end)
1562
if batcher.add_key(key) > BATCH_SIZE:
1563
# Ok, this batch is big enough. Yield some results.
1564
for factory in batcher.yield_factories():
1361
if manager is not None:
1362
for factory in manager.get_record_stream():
1364
last_read_memo = manager = None
1567
for factory in batcher.yield_factories(full_flush=True):
1365
1569
for record in source.get_record_stream(keys, ordering,
1366
1570
include_delta_closure):
1368
if manager is not None:
1369
for factory in manager.get_record_stream():
1572
for factory in batcher.yield_factories(full_flush=True):
1372
1575
def get_sha1s(self, keys):
1373
1576
"""See VersionedFiles.get_sha1s()."""
1459
1663
if reuse_blocks:
1460
1664
# If the reuse_blocks flag is set, check to see if we can just
1461
1665
# copy a groupcompress block as-is.
1666
# We only check on the first record (groupcompress-block) not
1667
# on all of the (groupcompress-block-ref) entries.
1668
# The reuse_this_block flag is then kept for as long as
1669
if record.storage_kind == 'groupcompress-block':
1670
# Check to see if we really want to re-use this block
1671
insert_manager = record._manager
1672
reuse_this_block = insert_manager.check_is_well_utilized()
1674
reuse_this_block = False
1675
if reuse_this_block:
1676
# We still want to reuse this block
1462
1677
if record.storage_kind == 'groupcompress-block':
1463
1678
# Insert the raw block into the target repo
1464
1679
insert_manager = record._manager
1465
insert_manager._check_rebuild_block()
1466
1680
bytes = record._manager._block.to_bytes()
1467
1681
_, start, length = self._access.add_raw_records(
1468
1682
[(None, len(bytes))], bytes)[0]
1787
2026
"""Convert an index value to position details."""
1788
2027
bits = node[2].split(' ')
1789
2028
# It would be nice not to read the entire gzip.
2029
# start and stop are put into _int_cache because they are very common.
2030
# They define the 'group' that an entry is in, and many groups can have
2031
# thousands of objects.
2032
# Branching Launchpad, for example, saves ~600k integers, at 12 bytes
2033
# each, or about 7MB. Note that it might be even more when you consider
2034
# how PyInt is allocated in separate slabs. And you can't return a slab
2035
# 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
1790
2038
start = int(bits[0])
2039
start = self._int_cache.setdefault(start, start)
1791
2040
stop = int(bits[1])
2041
stop = self._int_cache.setdefault(stop, stop)
1792
2042
basis_end = int(bits[2])
1793
2043
delta_end = int(bits[3])
1794
return node[0], start, stop, basis_end, delta_end
2044
# We can't use StaticTuple here, because node[0] is a BTreeGraphIndex
2046
return (node[0], start, stop, basis_end, delta_end)
1796
2048
def scan_unvalidated_index(self, graph_index):
1797
2049
"""Inform this _GCGraphIndex that there is an unvalidated index.
1799
2051
This allows this _GCGraphIndex to keep track of any missing
1800
2052
compression parents we may want to have filled in to make those
2053
indices valid. It also allows _GCGraphIndex to track any new keys.
1803
2055
:param graph_index: A GraphIndex
1805
if self._key_dependencies is not None:
1806
# Add parent refs from graph_index (and discard parent refs that
1807
# the graph_index has).
1808
add_refs = self._key_dependencies.add_references
1809
for node in graph_index.iter_all_entries():
1810
add_refs(node[1], node[3][0])
2057
key_dependencies = self._key_dependencies
2058
if key_dependencies is None:
2060
for node in graph_index.iter_all_entries():
2061
# Add parent refs from graph_index (and discard parent refs
2062
# that the graph_index has).
2063
key_dependencies.add_references(node[1], node[3][0])
1814
2066
from bzrlib._groupcompress_py import (