1
# Copyright (C) 2008, 2009 Canonical Ltd
3
# This program is free software; you can redistribute it and/or modify
4
# it under the terms of the GNU General Public License as published by
5
# the Free Software Foundation; either version 2 of the License, or
6
# (at your option) any later version.
8
# This program is distributed in the hope that it will be useful,
9
# but WITHOUT ANY WARRANTY; without even the implied warranty of
10
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11
# GNU General Public License for more details.
13
# You should have received a copy of the GNU General Public License
14
# along with this program; if not, write to the Free Software
15
# Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
17
"""Core compression logic for compressing streams of related files."""
36
from bzrlib.btree_index import BTreeBuilder
37
from bzrlib.lru_cache import LRUSizeCache
38
from bzrlib.tsort import topo_sort
39
from bzrlib.versionedfile import (
42
ChunkedContentFactory,
43
FulltextContentFactory,
47
# Minimum number of uncompressed bytes to try fetch at once when retrieving
48
# groupcompress blocks.
51
_USE_LZMA = False and (pylzma is not None)
53
# osutils.sha_string('')
54
_null_sha1 = 'da39a3ee5e6b4b0d3255bfef95601890afd80709'
56
def sort_gc_optimal(parent_map):
57
"""Sort and group the keys in parent_map into groupcompress order.
59
groupcompress is defined (currently) as reverse-topological order, grouped
62
:return: A sorted-list of keys
64
# groupcompress ordering is approximately reverse topological,
65
# properly grouped by file-id.
67
for key, value in parent_map.iteritems():
68
if isinstance(key, str) or len(key) == 1:
73
per_prefix_map[prefix][key] = value
75
per_prefix_map[prefix] = {key: value}
78
for prefix in sorted(per_prefix_map):
79
present_keys.extend(reversed(topo_sort(per_prefix_map[prefix])))
83
# The max zlib window size is 32kB, so if we set 'max_size' output of the
84
# decompressor to the requested bytes + 32kB, then we should guarantee
85
# num_bytes coming out.
86
_ZLIB_DECOMP_WINDOW = 32*1024
88
class GroupCompressBlock(object):
89
"""An object which maintains the internal structure of the compressed data.
91
This tracks the meta info (start of text, length, type, etc.)
94
# Group Compress Block v1 Zlib
95
GCB_HEADER = 'gcb1z\n'
96
# Group Compress Block v1 Lzma
97
GCB_LZ_HEADER = 'gcb1l\n'
98
GCB_KNOWN_HEADERS = (GCB_HEADER, GCB_LZ_HEADER)
101
# map by key? or just order in file?
102
self._compressor_name = None
103
self._z_content = None
104
self._z_content_decompressor = None
105
self._z_content_length = None
106
self._content_length = None
108
self._content_chunks = None
111
# This is the maximum number of bytes this object will reference if
112
# everything is decompressed. However, if we decompress less than
113
# everything... (this would cause some problems for LRUSizeCache)
114
return self._content_length + self._z_content_length
116
def _ensure_content(self, num_bytes=None):
117
"""Make sure that content has been expanded enough.
119
:param num_bytes: Ensure that we have extracted at least num_bytes of
120
content. If None, consume everything
122
# TODO: If we re-use the same content block at different times during
123
# get_record_stream(), it is possible that the first pass will
124
# get inserted, triggering an extract/_ensure_content() which
125
# will get rid of _z_content. And then the next use of the block
126
# will try to access _z_content (to send it over the wire), and
127
# fail because it is already extracted. Consider never releasing
128
# _z_content because of this.
129
if num_bytes is None:
130
num_bytes = self._content_length
131
elif (self._content_length is not None
132
and num_bytes > self._content_length):
133
raise AssertionError(
134
'requested num_bytes (%d) > content length (%d)'
135
% (num_bytes, self._content_length))
136
# Expand the content if required
137
if self._content is None:
138
if self._content_chunks is not None:
139
self._content = ''.join(self._content_chunks)
140
self._content_chunks = None
141
if self._content is None:
142
if self._z_content is None:
143
raise AssertionError('No content to decompress')
144
if self._z_content == '':
146
elif self._compressor_name == 'lzma':
147
# We don't do partial lzma decomp yet
148
self._content = pylzma.decompress(self._z_content)
149
elif self._compressor_name == 'zlib':
150
# Start a zlib decompressor
151
if num_bytes is None:
152
self._content = zlib.decompress(self._z_content)
154
self._z_content_decompressor = zlib.decompressobj()
155
# Seed the decompressor with the uncompressed bytes, so
156
# that the rest of the code is simplified
157
self._content = self._z_content_decompressor.decompress(
158
self._z_content, num_bytes + _ZLIB_DECOMP_WINDOW)
160
raise AssertionError('Unknown compressor: %r'
161
% self._compressor_name)
162
# Any bytes remaining to be decompressed will be in the decompressors
165
# Do we have enough bytes already?
166
if num_bytes is not None and len(self._content) >= num_bytes:
168
if num_bytes is None and self._z_content_decompressor is None:
169
# We must have already decompressed everything
171
# If we got this far, and don't have a decompressor, something is wrong
172
if self._z_content_decompressor is None:
173
raise AssertionError(
174
'No decompressor to decompress %d bytes' % num_bytes)
175
remaining_decomp = self._z_content_decompressor.unconsumed_tail
176
if num_bytes is None:
178
# We don't know how much is left, but we'll decompress it all
179
self._content += self._z_content_decompressor.decompress(
181
# Note: There's what I consider a bug in zlib.decompressobj
182
# If you pass back in the entire unconsumed_tail, only
183
# this time you don't pass a max-size, it doesn't
184
# change the unconsumed_tail back to None/''.
185
# However, we know we are done with the whole stream
186
self._z_content_decompressor = None
187
# XXX: Why is this the only place in this routine we set this?
188
self._content_length = len(self._content)
190
if not remaining_decomp:
191
raise AssertionError('Nothing left to decompress')
192
needed_bytes = num_bytes - len(self._content)
193
# We always set max_size to 32kB over the minimum needed, so that
194
# zlib will give us as much as we really want.
195
# TODO: If this isn't good enough, we could make a loop here,
196
# that keeps expanding the request until we get enough
197
self._content += self._z_content_decompressor.decompress(
198
remaining_decomp, needed_bytes + _ZLIB_DECOMP_WINDOW)
199
if len(self._content) < num_bytes:
200
raise AssertionError('%d bytes wanted, only %d available'
201
% (num_bytes, len(self._content)))
202
if not self._z_content_decompressor.unconsumed_tail:
203
# The stream is finished
204
self._z_content_decompressor = None
206
def _parse_bytes(self, bytes, pos):
207
"""Read the various lengths from the header.
209
This also populates the various 'compressed' buffers.
211
:return: The position in bytes just after the last newline
213
# At present, we have 2 integers for the compressed and uncompressed
214
# content. In base10 (ascii) 14 bytes can represent > 1TB, so to avoid
215
# checking too far, cap the search to 14 bytes.
216
pos2 = bytes.index('\n', pos, pos + 14)
217
self._z_content_length = int(bytes[pos:pos2])
219
pos2 = bytes.index('\n', pos, pos + 14)
220
self._content_length = int(bytes[pos:pos2])
222
if len(bytes) != (pos + self._z_content_length):
223
# XXX: Define some GCCorrupt error ?
224
raise AssertionError('Invalid bytes: (%d) != %d + %d' %
225
(len(bytes), pos, self._z_content_length))
226
self._z_content = bytes[pos:]
229
def from_bytes(cls, bytes):
231
if bytes[:6] not in cls.GCB_KNOWN_HEADERS:
232
raise ValueError('bytes did not start with any of %r'
233
% (cls.GCB_KNOWN_HEADERS,))
234
# XXX: why not testing the whole header ?
236
out._compressor_name = 'zlib'
237
elif bytes[4] == 'l':
238
out._compressor_name = 'lzma'
240
raise ValueError('unknown compressor: %r' % (bytes,))
241
out._parse_bytes(bytes, 6)
244
def extract(self, key, start, end, sha1=None):
245
"""Extract the text for a specific key.
247
:param key: The label used for this content
248
:param sha1: TODO (should we validate only when sha1 is supplied?)
249
:return: The bytes for the content
251
if start == end == 0:
253
self._ensure_content(end)
254
# The bytes are 'f' or 'd' for the type, then a variable-length
255
# base128 integer for the content size, then the actual content
256
# We know that the variable-length integer won't be longer than 5
257
# bytes (it takes 5 bytes to encode 2^32)
258
c = self._content[start]
263
raise ValueError('Unknown content control code: %s'
266
content_len, len_len = decode_base128_int(
267
self._content[start + 1:start + 6])
268
content_start = start + 1 + len_len
269
if end != content_start + content_len:
270
raise ValueError('end != len according to field header'
271
' %s != %s' % (end, content_start + content_len))
273
bytes = self._content[content_start:end]
275
bytes = apply_delta_to_source(self._content, content_start, end)
278
def set_chunked_content(self, content_chunks, length):
279
"""Set the content of this block to the given chunks."""
280
# If we have lots of short lines, it is may be more efficient to join
281
# the content ahead of time. If the content is <10MiB, we don't really
282
# care about the extra memory consumption, so we can just pack it and
283
# be done. However, timing showed 18s => 17.9s for repacking 1k revs of
284
# mysql, which is below the noise margin
285
self._content_length = length
286
self._content_chunks = content_chunks
288
self._z_content = None
290
def set_content(self, content):
291
"""Set the content of this block."""
292
self._content_length = len(content)
293
self._content = content
294
self._z_content = None
296
def _create_z_content_using_lzma(self):
297
if self._content_chunks is not None:
298
self._content = ''.join(self._content_chunks)
299
self._content_chunks = None
300
if self._content is None:
301
raise AssertionError('Nothing to compress')
302
self._z_content = pylzma.compress(self._content)
303
self._z_content_length = len(self._z_content)
305
def _create_z_content_from_chunks(self):
306
compressor = zlib.compressobj(zlib.Z_DEFAULT_COMPRESSION)
307
compressed_chunks = map(compressor.compress, self._content_chunks)
308
compressed_chunks.append(compressor.flush())
309
self._z_content = ''.join(compressed_chunks)
310
self._z_content_length = len(self._z_content)
312
def _create_z_content(self):
313
if self._z_content is not None:
316
self._create_z_content_using_lzma()
318
if self._content_chunks is not None:
319
self._create_z_content_from_chunks()
321
self._z_content = zlib.compress(self._content)
322
self._z_content_length = len(self._z_content)
325
"""Encode the information into a byte stream."""
326
self._create_z_content()
328
header = self.GCB_LZ_HEADER
330
header = self.GCB_HEADER
332
'%d\n%d\n' % (self._z_content_length, self._content_length),
335
return ''.join(chunks)
337
def _dump(self, include_text=False):
338
"""Take this block, and spit out a human-readable structure.
340
:param include_text: Inserts also include text bits, chose whether you
341
want this displayed in the dump or not.
342
:return: A dump of the given block. The layout is something like:
343
[('f', length), ('d', delta_length, text_length, [delta_info])]
344
delta_info := [('i', num_bytes, text), ('c', offset, num_bytes),
347
self._ensure_content()
350
while pos < self._content_length:
351
kind = self._content[pos]
353
if kind not in ('f', 'd'):
354
raise ValueError('invalid kind character: %r' % (kind,))
355
content_len, len_len = decode_base128_int(
356
self._content[pos:pos + 5])
358
if content_len + pos > self._content_length:
359
raise ValueError('invalid content_len %d for record @ pos %d'
360
% (content_len, pos - len_len - 1))
361
if kind == 'f': # Fulltext
363
text = self._content[pos:pos+content_len]
364
result.append(('f', content_len, text))
366
result.append(('f', content_len))
367
elif kind == 'd': # Delta
368
delta_content = self._content[pos:pos+content_len]
370
# The first entry in a delta is the decompressed length
371
decomp_len, delta_pos = decode_base128_int(delta_content)
372
result.append(('d', content_len, decomp_len, delta_info))
374
while delta_pos < content_len:
375
c = ord(delta_content[delta_pos])
379
delta_pos) = decode_copy_instruction(delta_content, c,
382
text = self._content[offset:offset+length]
383
delta_info.append(('c', offset, length, text))
385
delta_info.append(('c', offset, length))
386
measured_len += length
389
txt = delta_content[delta_pos:delta_pos+c]
392
delta_info.append(('i', c, txt))
395
if delta_pos != content_len:
396
raise ValueError('Delta consumed a bad number of bytes:'
397
' %d != %d' % (delta_pos, content_len))
398
if measured_len != decomp_len:
399
raise ValueError('Delta claimed fulltext was %d bytes, but'
400
' extraction resulted in %d bytes'
401
% (decomp_len, measured_len))
406
class _LazyGroupCompressFactory(object):
407
"""Yield content from a GroupCompressBlock on demand."""
409
def __init__(self, key, parents, manager, start, end, first):
410
"""Create a _LazyGroupCompressFactory
412
:param key: The key of just this record
413
:param parents: The parents of this key (possibly None)
414
:param gc_block: A GroupCompressBlock object
415
:param start: Offset of the first byte for this record in the
417
:param end: Offset of the byte just after the end of this record
418
(ie, bytes = content[start:end])
419
:param first: Is this the first Factory for the given block?
422
self.parents = parents
424
# Note: This attribute coupled with Manager._factories creates a
425
# reference cycle. Perhaps we would rather use a weakref(), or
426
# find an appropriate time to release the ref. After the first
427
# get_bytes_as call? After Manager.get_record_stream() returns
429
self._manager = manager
431
self.storage_kind = 'groupcompress-block'
433
self.storage_kind = 'groupcompress-block-ref'
439
return '%s(%s, first=%s)' % (self.__class__.__name__,
440
self.key, self._first)
442
def get_bytes_as(self, storage_kind):
443
if storage_kind == self.storage_kind:
445
# wire bytes, something...
446
return self._manager._wire_bytes()
449
if storage_kind in ('fulltext', 'chunked'):
450
if self._bytes is None:
451
# Grab and cache the raw bytes for this entry
452
# and break the ref-cycle with _manager since we don't need it
454
self._manager._prepare_for_extract()
455
block = self._manager._block
456
self._bytes = block.extract(self.key, self._start, self._end)
457
# There are code paths that first extract as fulltext, and then
458
# extract as storage_kind (smart fetch). So we don't break the
459
# refcycle here, but instead in manager.get_record_stream()
460
if storage_kind == 'fulltext':
464
raise errors.UnavailableRepresentation(self.key, storage_kind,
468
class _LazyGroupContentManager(object):
469
"""This manages a group of _LazyGroupCompressFactory objects."""
471
_max_cut_fraction = 0.75 # We allow a block to be trimmed to 75% of
472
# current size, and still be considered
474
_full_block_size = 4*1024*1024
475
_full_mixed_block_size = 2*1024*1024
476
_full_enough_block_size = 3*1024*1024 # size at which we won't repack
477
_full_enough_mixed_block_size = 2*768*1024 # 1.5MB
479
def __init__(self, block):
481
# We need to preserve the ordering
485
def add_factory(self, key, parents, start, end):
486
if not self._factories:
490
# Note that this creates a reference cycle....
491
factory = _LazyGroupCompressFactory(key, parents, self,
492
start, end, first=first)
493
# max() works here, but as a function call, doing a compare seems to be
494
# significantly faster, timeit says 250ms for max() and 100ms for the
496
if end > self._last_byte:
497
self._last_byte = end
498
self._factories.append(factory)
500
def get_record_stream(self):
501
"""Get a record for all keys added so far."""
502
for factory in self._factories:
504
# Break the ref-cycle
505
factory._bytes = None
506
factory._manager = None
507
# TODO: Consider setting self._factories = None after the above loop,
508
# as it will break the reference cycle
510
def _trim_block(self, last_byte):
511
"""Create a new GroupCompressBlock, with just some of the content."""
512
# None of the factories need to be adjusted, because the content is
513
# located in an identical place. Just that some of the unreferenced
514
# trailing bytes are stripped
515
trace.mutter('stripping trailing bytes from groupcompress block'
516
' %d => %d', self._block._content_length, last_byte)
517
new_block = GroupCompressBlock()
518
self._block._ensure_content(last_byte)
519
new_block.set_content(self._block._content[:last_byte])
520
self._block = new_block
522
def _rebuild_block(self):
523
"""Create a new GroupCompressBlock with only the referenced texts."""
524
compressor = GroupCompressor()
526
old_length = self._block._content_length
528
for factory in self._factories:
529
bytes = factory.get_bytes_as('fulltext')
530
(found_sha1, start_point, end_point,
531
type) = compressor.compress(factory.key, bytes, factory.sha1)
532
# Now update this factory with the new offsets, etc
533
factory.sha1 = found_sha1
534
factory._start = start_point
535
factory._end = end_point
536
self._last_byte = end_point
537
new_block = compressor.flush()
538
# TODO: Should we check that new_block really *is* smaller than the old
539
# block? It seems hard to come up with a method that it would
540
# expand, since we do full compression again. Perhaps based on a
541
# request that ends up poorly ordered?
542
delta = time.time() - tstart
543
self._block = new_block
544
trace.mutter('creating new compressed block on-the-fly in %.3fs'
545
' %d bytes => %d bytes', delta, old_length,
546
self._block._content_length)
548
def _prepare_for_extract(self):
549
"""A _LazyGroupCompressFactory is about to extract to fulltext."""
550
# We expect that if one child is going to fulltext, all will be. This
551
# helps prevent all of them from extracting a small amount at a time.
552
# Which in itself isn't terribly expensive, but resizing 2MB 32kB at a
553
# time (self._block._content) is a little expensive.
554
self._block._ensure_content(self._last_byte)
556
def _check_rebuild_action(self):
557
"""Check to see if our block should be repacked."""
560
for factory in self._factories:
561
total_bytes_used += factory._end - factory._start
562
if last_byte_used < factory._end:
563
last_byte_used = factory._end
564
# If we are using more than half of the bytes from the block, we have
565
# nothing else to check
566
if total_bytes_used * 2 >= self._block._content_length:
567
return None, last_byte_used, total_bytes_used
568
# We are using less than 50% of the content. Is the content we are
569
# using at the beginning of the block? If so, we can just trim the
570
# tail, rather than rebuilding from scratch.
571
if total_bytes_used * 2 > last_byte_used:
572
return 'trim', last_byte_used, total_bytes_used
574
# We are using a small amount of the data, and it isn't just packed
575
# nicely at the front, so rebuild the content.
576
# Note: This would be *nicer* as a strip-data-from-group, rather than
577
# building it up again from scratch
578
# It might be reasonable to consider the fulltext sizes for
579
# different bits when deciding this, too. As you may have a small
580
# fulltext, and a trivial delta, and you are just trading around
581
# for another fulltext. If we do a simple 'prune' you may end up
582
# expanding many deltas into fulltexts, as well.
583
# If we build a cheap enough 'strip', then we could try a strip,
584
# if that expands the content, we then rebuild.
585
return 'rebuild', last_byte_used, total_bytes_used
587
def check_is_well_utilized(self):
588
"""Is the current block considered 'well utilized'?
590
This is a bit of a heuristic, but it basically asks if the current
591
block considers itself to be a fully developed group, rather than just
592
a loose collection of data.
594
if len(self._factories) == 1:
595
# A block of length 1 is never considered 'well utilized' :)
597
action, last_byte_used, total_bytes_used = self._check_rebuild_action()
598
block_size = self._block._content_length
599
if total_bytes_used < block_size * self._max_cut_fraction:
600
# This block wants to trim itself small enough that we want to
601
# consider it under-utilized.
603
# TODO: This code is meant to be the twin of _insert_record_stream's
604
# 'start_new_block' logic. It would probably be better to factor
605
# out that logic into a shared location, so that it stays
607
# We currently assume a block is properly utilized whenever it is >75%
608
# of the size of a 'full' block. In normal operation, a block is
609
# considered full when it hits 4MB of same-file content. So any block
610
# >3MB is 'full enough'.
611
# The only time this isn't true is when a given block has large-object
612
# content. (a single file >4MB, etc.)
613
# Under these circumstances, we allow a block to grow to
614
# 2 x largest_content. Which means that if a given block had a large
615
# object, it may actually be under-utilized. However, given that this
616
# is 'pack-on-the-fly' it is probably reasonable to not repack large
617
# contet blobs on-the-fly.
618
if block_size >= self._full_enough_block_size:
620
# If a block is <3MB, it still may be considered 'full' if it contains
621
# mixed content. The current rule is 2MB of mixed content is considered
622
# full. So check to see if this block contains mixed content, and
623
# set the threshold appropriately.
625
for factory in self._factories:
626
prefix = factory.key[:-1]
627
if common_prefix is None:
628
common_prefix = prefix
629
elif prefix != common_prefix:
630
# Mixed content, check the size appropriately
631
if block_size >= self._full_enough_mixed_block_size:
634
# The content failed both the mixed check and the single-content check
635
# so obviously it is not fully utilized
636
# TODO: there is one other constraint that isn't being checked
637
# namely, that the entries in the block are in the appropriate
638
# order. For example, you could insert the entries in exactly
639
# reverse groupcompress order, and we would think that is ok.
640
# (all the right objects are in one group, and it is fully
641
# utilized, etc.) For now, we assume that case is rare,
642
# especially since we should always fetch in 'groupcompress'
646
def _check_rebuild_block(self):
647
action, last_byte_used, total_bytes_used = self._check_rebuild_action()
651
self._trim_block(last_byte_used)
652
elif action == 'rebuild':
653
self._rebuild_block()
655
raise ValueError('unknown rebuild action: %r' % (action,))
657
def _wire_bytes(self):
658
"""Return a byte stream suitable for transmitting over the wire."""
659
self._check_rebuild_block()
660
# The outer block starts with:
661
# 'groupcompress-block\n'
662
# <length of compressed key info>\n
663
# <length of uncompressed info>\n
664
# <length of gc block>\n
667
lines = ['groupcompress-block\n']
668
# The minimal info we need is the key, the start offset, and the
669
# parents. The length and type are encoded in the record itself.
670
# However, passing in the other bits makes it easier. The list of
671
# keys, and the start offset, the length
673
# 1 line with parents, '' for ()
674
# 1 line for start offset
675
# 1 line for end byte
677
for factory in self._factories:
678
key_bytes = '\x00'.join(factory.key)
679
parents = factory.parents
681
parent_bytes = 'None:'
683
parent_bytes = '\t'.join('\x00'.join(key) for key in parents)
684
record_header = '%s\n%s\n%d\n%d\n' % (
685
key_bytes, parent_bytes, factory._start, factory._end)
686
header_lines.append(record_header)
687
# TODO: Can we break the refcycle at this point and set
688
# factory._manager = None?
689
header_bytes = ''.join(header_lines)
691
header_bytes_len = len(header_bytes)
692
z_header_bytes = zlib.compress(header_bytes)
694
z_header_bytes_len = len(z_header_bytes)
695
block_bytes = self._block.to_bytes()
696
lines.append('%d\n%d\n%d\n' % (z_header_bytes_len, header_bytes_len,
698
lines.append(z_header_bytes)
699
lines.append(block_bytes)
700
del z_header_bytes, block_bytes
701
return ''.join(lines)
704
def from_bytes(cls, bytes):
705
# TODO: This does extra string copying, probably better to do it a
707
(storage_kind, z_header_len, header_len,
708
block_len, rest) = bytes.split('\n', 4)
710
if storage_kind != 'groupcompress-block':
711
raise ValueError('Unknown storage kind: %s' % (storage_kind,))
712
z_header_len = int(z_header_len)
713
if len(rest) < z_header_len:
714
raise ValueError('Compressed header len shorter than all bytes')
715
z_header = rest[:z_header_len]
716
header_len = int(header_len)
717
header = zlib.decompress(z_header)
718
if len(header) != header_len:
719
raise ValueError('invalid length for decompressed bytes')
721
block_len = int(block_len)
722
if len(rest) != z_header_len + block_len:
723
raise ValueError('Invalid length for block')
724
block_bytes = rest[z_header_len:]
726
# So now we have a valid GCB, we just need to parse the factories that
728
header_lines = header.split('\n')
730
last = header_lines.pop()
732
raise ValueError('header lines did not end with a trailing'
734
if len(header_lines) % 4 != 0:
735
raise ValueError('The header was not an even multiple of 4 lines')
736
block = GroupCompressBlock.from_bytes(block_bytes)
739
for start in xrange(0, len(header_lines), 4):
741
key = tuple(header_lines[start].split('\x00'))
742
parents_line = header_lines[start+1]
743
if parents_line == 'None:':
746
parents = tuple([tuple(segment.split('\x00'))
747
for segment in parents_line.split('\t')
749
start_offset = int(header_lines[start+2])
750
end_offset = int(header_lines[start+3])
751
result.add_factory(key, parents, start_offset, end_offset)
755
def network_block_to_records(storage_kind, bytes, line_end):
756
if storage_kind != 'groupcompress-block':
757
raise ValueError('Unknown storage kind: %s' % (storage_kind,))
758
manager = _LazyGroupContentManager.from_bytes(bytes)
759
return manager.get_record_stream()
762
class _CommonGroupCompressor(object):
765
"""Create a GroupCompressor."""
770
self.labels_deltas = {}
771
self._delta_index = None # Set by the children
772
self._block = GroupCompressBlock()
774
def compress(self, key, bytes, expected_sha, nostore_sha=None, soft=False):
775
"""Compress lines with label key.
777
:param key: A key tuple. It is stored in the output
778
for identification of the text during decompression. If the last
779
element is 'None' it is replaced with the sha1 of the text -
781
:param bytes: The bytes to be compressed
782
:param expected_sha: If non-None, the sha the lines are believed to
783
have. During compression the sha is calculated; a mismatch will
785
:param nostore_sha: If the computed sha1 sum matches, we will raise
786
ExistingContent rather than adding the text.
787
:param soft: Do a 'soft' compression. This means that we require larger
788
ranges to match to be considered for a copy command.
790
:return: The sha1 of lines, the start and end offsets in the delta, and
791
the type ('fulltext' or 'delta').
793
:seealso VersionedFiles.add_lines:
795
if not bytes: # empty, like a dir entry, etc
796
if nostore_sha == _null_sha1:
797
raise errors.ExistingContent()
798
return _null_sha1, 0, 0, 'fulltext'
799
# we assume someone knew what they were doing when they passed it in
800
if expected_sha is not None:
803
sha1 = osutils.sha_string(bytes)
804
if nostore_sha is not None:
805
if sha1 == nostore_sha:
806
raise errors.ExistingContent()
808
key = key[:-1] + ('sha1:' + sha1,)
810
start, end, type = self._compress(key, bytes, len(bytes) / 2, soft)
811
return sha1, start, end, type
813
def _compress(self, key, bytes, max_delta_size, soft=False):
814
"""Compress lines with label key.
816
:param key: A key tuple. It is stored in the output for identification
817
of the text during decompression.
819
:param bytes: The bytes to be compressed
821
:param max_delta_size: The size above which we issue a fulltext instead
824
:param soft: Do a 'soft' compression. This means that we require larger
825
ranges to match to be considered for a copy command.
827
:return: The sha1 of lines, the start and end offsets in the delta, and
828
the type ('fulltext' or 'delta').
830
raise NotImplementedError(self._compress)
832
def extract(self, key):
833
"""Extract a key previously added to the compressor.
835
:param key: The key to extract.
836
:return: An iterable over bytes and the sha1.
838
(start_byte, start_chunk, end_byte, end_chunk) = self.labels_deltas[key]
839
delta_chunks = self.chunks[start_chunk:end_chunk]
840
stored_bytes = ''.join(delta_chunks)
841
if stored_bytes[0] == 'f':
842
fulltext_len, offset = decode_base128_int(stored_bytes[1:10])
843
data_len = fulltext_len + 1 + offset
844
if data_len != len(stored_bytes):
845
raise ValueError('Index claimed fulltext len, but stored bytes'
847
% (len(stored_bytes), data_len))
848
bytes = stored_bytes[offset + 1:]
850
# XXX: This is inefficient at best
851
source = ''.join(self.chunks[:start_chunk])
852
if stored_bytes[0] != 'd':
853
raise ValueError('Unknown content kind, bytes claim %s'
854
% (stored_bytes[0],))
855
delta_len, offset = decode_base128_int(stored_bytes[1:10])
856
data_len = delta_len + 1 + offset
857
if data_len != len(stored_bytes):
858
raise ValueError('Index claimed delta len, but stored bytes'
860
% (len(stored_bytes), data_len))
861
bytes = apply_delta(source, stored_bytes[offset + 1:])
862
bytes_sha1 = osutils.sha_string(bytes)
863
return bytes, bytes_sha1
866
"""Finish this group, creating a formatted stream.
868
After calling this, the compressor should no longer be used
870
# TODO: this causes us to 'bloat' to 2x the size of content in the
871
# group. This has an impact for 'commit' of large objects.
872
# One possibility is to use self._content_chunks, and be lazy and
873
# only fill out self._content as a full string when we actually
874
# need it. That would at least drop the peak memory consumption
875
# for 'commit' down to ~1x the size of the largest file, at a
876
# cost of increased complexity within this code. 2x is still <<
877
# 3x the size of the largest file, so we are doing ok.
878
self._block.set_chunked_content(self.chunks, self.endpoint)
880
self._delta_index = None
884
"""Call this if you want to 'revoke' the last compression.
886
After this, the data structures will be rolled back, but you cannot do
889
self._delta_index = None
890
del self.chunks[self._last[0]:]
891
self.endpoint = self._last[1]
895
"""Return the overall compression ratio."""
896
return float(self.input_bytes) / float(self.endpoint)
899
class PythonGroupCompressor(_CommonGroupCompressor):
902
"""Create a GroupCompressor.
904
Used only if the pyrex version is not available.
906
super(PythonGroupCompressor, self).__init__()
907
self._delta_index = LinesDeltaIndex([])
908
# The actual content is managed by LinesDeltaIndex
909
self.chunks = self._delta_index.lines
911
def _compress(self, key, bytes, max_delta_size, soft=False):
912
"""see _CommonGroupCompressor._compress"""
913
input_len = len(bytes)
914
new_lines = osutils.split_lines(bytes)
915
out_lines, index_lines = self._delta_index.make_delta(
916
new_lines, bytes_length=input_len, soft=soft)
917
delta_length = sum(map(len, out_lines))
918
if delta_length > max_delta_size:
919
# The delta is longer than the fulltext, insert a fulltext
921
out_lines = ['f', encode_base128_int(input_len)]
922
out_lines.extend(new_lines)
923
index_lines = [False, False]
924
index_lines.extend([True] * len(new_lines))
926
# this is a worthy delta, output it
929
# Update the delta_length to include those two encoded integers
930
out_lines[1] = encode_base128_int(delta_length)
932
start = self.endpoint
933
chunk_start = len(self.chunks)
934
self._last = (chunk_start, self.endpoint)
935
self._delta_index.extend_lines(out_lines, index_lines)
936
self.endpoint = self._delta_index.endpoint
937
self.input_bytes += input_len
938
chunk_end = len(self.chunks)
939
self.labels_deltas[key] = (start, chunk_start,
940
self.endpoint, chunk_end)
941
return start, self.endpoint, type
944
class PyrexGroupCompressor(_CommonGroupCompressor):
945
"""Produce a serialised group of compressed texts.
947
It contains code very similar to SequenceMatcher because of having a similar
948
task. However some key differences apply:
949
- there is no junk, we want a minimal edit not a human readable diff.
950
- we don't filter very common lines (because we don't know where a good
951
range will start, and after the first text we want to be emitting minmal
953
- we chain the left side, not the right side
954
- we incrementally update the adjacency matrix as new lines are provided.
955
- we look for matches in all of the left side, so the routine which does
956
the analagous task of find_longest_match does not need to filter on the
961
super(PyrexGroupCompressor, self).__init__()
962
self._delta_index = DeltaIndex()
964
def _compress(self, key, bytes, max_delta_size, soft=False):
965
"""see _CommonGroupCompressor._compress"""
966
input_len = len(bytes)
967
# By having action/label/sha1/len, we can parse the group if the index
968
# was ever destroyed, we have the key in 'label', we know the final
969
# bytes are valid from sha1, and we know where to find the end of this
970
# record because of 'len'. (the delta record itself will store the
971
# total length for the expanded record)
972
# 'len: %d\n' costs approximately 1% increase in total data
973
# Having the labels at all costs us 9-10% increase, 38% increase for
974
# inventory pages, and 5.8% increase for text pages
975
# new_chunks = ['label:%s\nsha1:%s\n' % (label, sha1)]
976
if self._delta_index._source_offset != self.endpoint:
977
raise AssertionError('_source_offset != endpoint'
978
' somehow the DeltaIndex got out of sync with'
980
delta = self._delta_index.make_delta(bytes, max_delta_size)
983
enc_length = encode_base128_int(len(bytes))
984
len_mini_header = 1 + len(enc_length)
985
self._delta_index.add_source(bytes, len_mini_header)
986
new_chunks = ['f', enc_length, bytes]
989
enc_length = encode_base128_int(len(delta))
990
len_mini_header = 1 + len(enc_length)
991
new_chunks = ['d', enc_length, delta]
992
self._delta_index.add_delta_source(delta, len_mini_header)
994
start = self.endpoint
995
chunk_start = len(self.chunks)
996
# Now output these bytes
997
self._output_chunks(new_chunks)
998
self.input_bytes += input_len
999
chunk_end = len(self.chunks)
1000
self.labels_deltas[key] = (start, chunk_start,
1001
self.endpoint, chunk_end)
1002
if not self._delta_index._source_offset == self.endpoint:
1003
raise AssertionError('the delta index is out of sync'
1004
'with the output lines %s != %s'
1005
% (self._delta_index._source_offset, self.endpoint))
1006
return start, self.endpoint, type
1008
def _output_chunks(self, new_chunks):
1009
"""Output some chunks.
1011
:param new_chunks: The chunks to output.
1013
self._last = (len(self.chunks), self.endpoint)
1014
endpoint = self.endpoint
1015
self.chunks.extend(new_chunks)
1016
endpoint += sum(map(len, new_chunks))
1017
self.endpoint = endpoint
1020
def make_pack_factory(graph, delta, keylength, inconsistency_fatal=True):
1021
"""Create a factory for creating a pack based groupcompress.
1023
This is only functional enough to run interface tests, it doesn't try to
1024
provide a full pack environment.
1026
:param graph: Store a graph.
1027
:param delta: Delta compress contents.
1028
:param keylength: How long should keys be.
1030
def factory(transport):
1035
graph_index = BTreeBuilder(reference_lists=ref_length,
1036
key_elements=keylength)
1037
stream = transport.open_write_stream('newpack')
1038
writer = pack.ContainerWriter(stream.write)
1040
index = _GCGraphIndex(graph_index, lambda:True, parents=parents,
1041
add_callback=graph_index.add_nodes,
1042
inconsistency_fatal=inconsistency_fatal)
1043
access = knit._DirectPackAccess({})
1044
access.set_writer(writer, graph_index, (transport, 'newpack'))
1045
result = GroupCompressVersionedFiles(index, access, delta)
1046
result.stream = stream
1047
result.writer = writer
1052
def cleanup_pack_group(versioned_files):
1053
versioned_files.writer.end()
1054
versioned_files.stream.close()
1057
class _BatchingBlockFetcher(object):
1058
"""Fetch group compress blocks in batches.
1060
:ivar total_bytes: int of expected number of bytes needed to fetch the
1061
currently pending batch.
1064
def __init__(self, gcvf, locations):
1066
self.locations = locations
1068
self.batch_memos = {}
1069
self.memos_to_get = []
1070
self.total_bytes = 0
1071
self.last_read_memo = None
1074
def add_key(self, key):
1075
"""Add another to key to fetch.
1077
:return: The estimated number of bytes needed to fetch the batch so
1080
self.keys.append(key)
1081
index_memo, _, _, _ = self.locations[key]
1082
read_memo = index_memo[0:3]
1083
# Three possibilities for this read_memo:
1084
# - it's already part of this batch; or
1085
# - it's not yet part of this batch, but is already cached; or
1086
# - it's not yet part of this batch and will need to be fetched.
1087
if read_memo in self.batch_memos:
1088
# This read memo is already in this batch.
1089
return self.total_bytes
1091
cached_block = self.gcvf._group_cache[read_memo]
1093
# This read memo is new to this batch, and the data isn't cached
1095
self.batch_memos[read_memo] = None
1096
self.memos_to_get.append(read_memo)
1097
byte_length = read_memo[2]
1098
self.total_bytes += byte_length
1100
# This read memo is new to this batch, but cached.
1101
# Keep a reference to the cached block in batch_memos because it's
1102
# certain that we'll use it when this batch is processed, but
1103
# there's a risk that it would fall out of _group_cache between now
1105
self.batch_memos[read_memo] = cached_block
1106
return self.total_bytes
1108
def _flush_manager(self):
1109
if self.manager is not None:
1110
for factory in self.manager.get_record_stream():
1113
self.last_read_memo = None
1115
def yield_factories(self, full_flush=False):
1116
"""Yield factories for keys added since the last yield. They will be
1117
returned in the order they were added via add_key.
1119
:param full_flush: by default, some results may not be returned in case
1120
they can be part of the next batch. If full_flush is True, then
1121
all results are returned.
1123
if self.manager is None and not self.keys:
1125
# Fetch all memos in this batch.
1126
blocks = self.gcvf._get_blocks(self.memos_to_get)
1127
# Turn blocks into factories and yield them.
1128
memos_to_get_stack = list(self.memos_to_get)
1129
memos_to_get_stack.reverse()
1130
for key in self.keys:
1131
index_memo, _, parents, _ = self.locations[key]
1132
read_memo = index_memo[:3]
1133
if self.last_read_memo != read_memo:
1134
# We are starting a new block. If we have a
1135
# manager, we have found everything that fits for
1136
# now, so yield records
1137
for factory in self._flush_manager():
1139
# Now start a new manager.
1140
if memos_to_get_stack and memos_to_get_stack[-1] == read_memo:
1141
# The next block from _get_blocks will be the block we
1143
block_read_memo, block = blocks.next()
1144
if block_read_memo != read_memo:
1145
raise AssertionError(
1146
"block_read_memo out of sync with read_memo"
1147
"(%r != %r)" % (block_read_memo, read_memo))
1148
self.batch_memos[read_memo] = block
1149
memos_to_get_stack.pop()
1151
block = self.batch_memos[read_memo]
1152
self.manager = _LazyGroupContentManager(block)
1153
self.last_read_memo = read_memo
1154
start, end = index_memo[3:5]
1155
self.manager.add_factory(key, parents, start, end)
1157
for factory in self._flush_manager():
1160
self.batch_memos.clear()
1161
del self.memos_to_get[:]
1162
self.total_bytes = 0
1165
class GroupCompressVersionedFiles(VersionedFiles):
1166
"""A group-compress based VersionedFiles implementation."""
1168
def __init__(self, index, access, delta=True):
1169
"""Create a GroupCompressVersionedFiles object.
1171
:param index: The index object storing access and graph data.
1172
:param access: The access object storing raw data.
1173
:param delta: Whether to delta compress or just entropy compress.
1176
self._access = access
1178
self._unadded_refs = {}
1179
self._group_cache = LRUSizeCache(max_size=50*1024*1024)
1180
self._fallback_vfs = []
1182
def add_lines(self, key, parents, lines, parent_texts=None,
1183
left_matching_blocks=None, nostore_sha=None, random_id=False,
1184
check_content=True):
1185
"""Add a text to the store.
1187
:param key: The key tuple of the text to add.
1188
:param parents: The parents key tuples of the text to add.
1189
:param lines: A list of lines. Each line must be a bytestring. And all
1190
of them except the last must be terminated with \n and contain no
1191
other \n's. The last line may either contain no \n's or a single
1192
terminating \n. If the lines list does meet this constraint the add
1193
routine may error or may succeed - but you will be unable to read
1194
the data back accurately. (Checking the lines have been split
1195
correctly is expensive and extremely unlikely to catch bugs so it
1196
is not done at runtime unless check_content is True.)
1197
:param parent_texts: An optional dictionary containing the opaque
1198
representations of some or all of the parents of version_id to
1199
allow delta optimisations. VERY IMPORTANT: the texts must be those
1200
returned by add_lines or data corruption can be caused.
1201
:param left_matching_blocks: a hint about which areas are common
1202
between the text and its left-hand-parent. The format is
1203
the SequenceMatcher.get_matching_blocks format.
1204
:param nostore_sha: Raise ExistingContent and do not add the lines to
1205
the versioned file if the digest of the lines matches this.
1206
:param random_id: If True a random id has been selected rather than
1207
an id determined by some deterministic process such as a converter
1208
from a foreign VCS. When True the backend may choose not to check
1209
for uniqueness of the resulting key within the versioned file, so
1210
this should only be done when the result is expected to be unique
1212
:param check_content: If True, the lines supplied are verified to be
1213
bytestrings that are correctly formed lines.
1214
:return: The text sha1, the number of bytes in the text, and an opaque
1215
representation of the inserted version which can be provided
1216
back to future add_lines calls in the parent_texts dictionary.
1218
self._index._check_write_ok()
1219
self._check_add(key, lines, random_id, check_content)
1221
# The caller might pass None if there is no graph data, but kndx
1222
# indexes can't directly store that, so we give them
1223
# an empty tuple instead.
1225
# double handling for now. Make it work until then.
1226
length = sum(map(len, lines))
1227
record = ChunkedContentFactory(key, parents, None, lines)
1228
sha1 = list(self._insert_record_stream([record], random_id=random_id,
1229
nostore_sha=nostore_sha))[0]
1230
return sha1, length, None
1232
def _add_text(self, key, parents, text, nostore_sha=None, random_id=False):
1233
"""See VersionedFiles._add_text()."""
1234
self._index._check_write_ok()
1235
self._check_add(key, None, random_id, check_content=False)
1236
if text.__class__ is not str:
1237
raise errors.BzrBadParameterUnicode("text")
1239
# The caller might pass None if there is no graph data, but kndx
1240
# indexes can't directly store that, so we give them
1241
# an empty tuple instead.
1243
# double handling for now. Make it work until then.
1245
record = FulltextContentFactory(key, parents, None, text)
1246
sha1 = list(self._insert_record_stream([record], random_id=random_id,
1247
nostore_sha=nostore_sha))[0]
1248
return sha1, length, None
1250
def add_fallback_versioned_files(self, a_versioned_files):
1251
"""Add a source of texts for texts not present in this knit.
1253
:param a_versioned_files: A VersionedFiles object.
1255
self._fallback_vfs.append(a_versioned_files)
1257
def annotate(self, key):
1258
"""See VersionedFiles.annotate."""
1259
ann = annotate.Annotator(self)
1260
return ann.annotate_flat(key)
1262
def get_annotator(self):
1263
return annotate.Annotator(self)
1265
def check(self, progress_bar=None, keys=None):
1266
"""See VersionedFiles.check()."""
1269
for record in self.get_record_stream(keys, 'unordered', True):
1270
record.get_bytes_as('fulltext')
1272
return self.get_record_stream(keys, 'unordered', True)
1274
def _check_add(self, key, lines, random_id, check_content):
1275
"""check that version_id and lines are safe to add."""
1276
version_id = key[-1]
1277
if version_id is not None:
1278
if osutils.contains_whitespace(version_id):
1279
raise errors.InvalidRevisionId(version_id, self)
1280
self.check_not_reserved_id(version_id)
1281
# TODO: If random_id==False and the key is already present, we should
1282
# probably check that the existing content is identical to what is
1283
# being inserted, and otherwise raise an exception. This would make
1284
# the bundle code simpler.
1286
self._check_lines_not_unicode(lines)
1287
self._check_lines_are_lines(lines)
1289
def get_known_graph_ancestry(self, keys):
1290
"""Get a KnownGraph instance with the ancestry of keys."""
1291
# Note that this is identical to
1292
# KnitVersionedFiles.get_known_graph_ancestry, but they don't share
1294
parent_map, missing_keys = self._index.find_ancestry(keys)
1295
for fallback in self._fallback_vfs:
1296
if not missing_keys:
1298
(f_parent_map, f_missing_keys) = fallback._index.find_ancestry(
1300
parent_map.update(f_parent_map)
1301
missing_keys = f_missing_keys
1302
kg = _mod_graph.KnownGraph(parent_map)
1305
def get_parent_map(self, keys):
1306
"""Get a map of the graph parents of keys.
1308
:param keys: The keys to look up parents for.
1309
:return: A mapping from keys to parents. Absent keys are absent from
1312
return self._get_parent_map_with_sources(keys)[0]
1314
def _get_parent_map_with_sources(self, keys):
1315
"""Get a map of the parents of keys.
1317
:param keys: The keys to look up parents for.
1318
:return: A tuple. The first element is a mapping from keys to parents.
1319
Absent keys are absent from the mapping. The second element is a
1320
list with the locations each key was found in. The first element
1321
is the in-this-knit parents, the second the first fallback source,
1325
sources = [self._index] + self._fallback_vfs
1328
for source in sources:
1331
new_result = source.get_parent_map(missing)
1332
source_results.append(new_result)
1333
result.update(new_result)
1334
missing.difference_update(set(new_result))
1335
return result, source_results
1337
def _get_blocks(self, read_memos):
1338
"""Get GroupCompressBlocks for the given read_memos.
1340
:returns: a series of (read_memo, block) pairs, in the order they were
1344
for read_memo in read_memos:
1346
block = self._group_cache[read_memo]
1350
cached[read_memo] = block
1352
not_cached_seen = set()
1353
for read_memo in read_memos:
1354
if read_memo in cached:
1355
# Don't fetch what we already have
1357
if read_memo in not_cached_seen:
1358
# Don't try to fetch the same data twice
1360
not_cached.append(read_memo)
1361
not_cached_seen.add(read_memo)
1362
raw_records = self._access.get_raw_records(not_cached)
1363
for read_memo in read_memos:
1365
yield read_memo, cached[read_memo]
1367
# Read the block, and cache it.
1368
zdata = raw_records.next()
1369
block = GroupCompressBlock.from_bytes(zdata)
1370
self._group_cache[read_memo] = block
1371
cached[read_memo] = block
1372
yield read_memo, block
1374
def get_missing_compression_parent_keys(self):
1375
"""Return the keys of missing compression parents.
1377
Missing compression parents occur when a record stream was missing
1378
basis texts, or a index was scanned that had missing basis texts.
1380
# GroupCompress cannot currently reference texts that are not in the
1381
# group, so this is valid for now
1384
def get_record_stream(self, keys, ordering, include_delta_closure):
1385
"""Get a stream of records for keys.
1387
:param keys: The keys to include.
1388
:param ordering: Either 'unordered' or 'topological'. A topologically
1389
sorted stream has compression parents strictly before their
1391
:param include_delta_closure: If True then the closure across any
1392
compression parents will be included (in the opaque data).
1393
:return: An iterator of ContentFactory objects, each of which is only
1394
valid until the iterator is advanced.
1396
# keys might be a generator
1397
orig_keys = list(keys)
1401
if (not self._index.has_graph
1402
and ordering in ('topological', 'groupcompress')):
1403
# Cannot topological order when no graph has been stored.
1404
# but we allow 'as-requested' or 'unordered'
1405
ordering = 'unordered'
1407
remaining_keys = keys
1410
keys = set(remaining_keys)
1411
for content_factory in self._get_remaining_record_stream(keys,
1412
orig_keys, ordering, include_delta_closure):
1413
remaining_keys.discard(content_factory.key)
1414
yield content_factory
1416
except errors.RetryWithNewPacks, e:
1417
self._access.reload_or_raise(e)
1419
def _find_from_fallback(self, missing):
1420
"""Find whatever keys you can from the fallbacks.
1422
:param missing: A set of missing keys. This set will be mutated as keys
1423
are found from a fallback_vfs
1424
:return: (parent_map, key_to_source_map, source_results)
1425
parent_map the overall key => parent_keys
1426
key_to_source_map a dict from {key: source}
1427
source_results a list of (source: keys)
1430
key_to_source_map = {}
1432
for source in self._fallback_vfs:
1435
source_parents = source.get_parent_map(missing)
1436
parent_map.update(source_parents)
1437
source_parents = list(source_parents)
1438
source_results.append((source, source_parents))
1439
key_to_source_map.update((key, source) for key in source_parents)
1440
missing.difference_update(source_parents)
1441
return parent_map, key_to_source_map, source_results
1443
def _get_ordered_source_keys(self, ordering, parent_map, key_to_source_map):
1444
"""Get the (source, [keys]) list.
1446
The returned objects should be in the order defined by 'ordering',
1447
which can weave between different sources.
1448
:param ordering: Must be one of 'topological' or 'groupcompress'
1449
:return: List of [(source, [keys])] tuples, such that all keys are in
1450
the defined order, regardless of source.
1452
if ordering == 'topological':
1453
present_keys = topo_sort(parent_map)
1455
# ordering == 'groupcompress'
1456
# XXX: This only optimizes for the target ordering. We may need
1457
# to balance that with the time it takes to extract
1458
# ordering, by somehow grouping based on
1459
# locations[key][0:3]
1460
present_keys = sort_gc_optimal(parent_map)
1461
# Now group by source:
1463
current_source = None
1464
for key in present_keys:
1465
source = key_to_source_map.get(key, self)
1466
if source is not current_source:
1467
source_keys.append((source, []))
1468
current_source = source
1469
source_keys[-1][1].append(key)
1472
def _get_as_requested_source_keys(self, orig_keys, locations, unadded_keys,
1475
current_source = None
1476
for key in orig_keys:
1477
if key in locations or key in unadded_keys:
1479
elif key in key_to_source_map:
1480
source = key_to_source_map[key]
1483
if source is not current_source:
1484
source_keys.append((source, []))
1485
current_source = source
1486
source_keys[-1][1].append(key)
1489
def _get_io_ordered_source_keys(self, locations, unadded_keys,
1492
# This is the group the bytes are stored in, followed by the
1493
# location in the group
1494
return locations[key][0]
1495
present_keys = sorted(locations.iterkeys(), key=get_group)
1496
# We don't have an ordering for keys in the in-memory object, but
1497
# lets process the in-memory ones first.
1498
present_keys = list(unadded_keys) + present_keys
1499
# Now grab all of the ones from other sources
1500
source_keys = [(self, present_keys)]
1501
source_keys.extend(source_result)
1504
def _get_remaining_record_stream(self, keys, orig_keys, ordering,
1505
include_delta_closure):
1506
"""Get a stream of records for keys.
1508
:param keys: The keys to include.
1509
:param ordering: one of 'unordered', 'topological', 'groupcompress' or
1511
:param include_delta_closure: If True then the closure across any
1512
compression parents will be included (in the opaque data).
1513
:return: An iterator of ContentFactory objects, each of which is only
1514
valid until the iterator is advanced.
1517
locations = self._index.get_build_details(keys)
1518
unadded_keys = set(self._unadded_refs).intersection(keys)
1519
missing = keys.difference(locations)
1520
missing.difference_update(unadded_keys)
1521
(fallback_parent_map, key_to_source_map,
1522
source_result) = self._find_from_fallback(missing)
1523
if ordering in ('topological', 'groupcompress'):
1524
# would be better to not globally sort initially but instead
1525
# start with one key, recurse to its oldest parent, then grab
1526
# everything in the same group, etc.
1527
parent_map = dict((key, details[2]) for key, details in
1528
locations.iteritems())
1529
for key in unadded_keys:
1530
parent_map[key] = self._unadded_refs[key]
1531
parent_map.update(fallback_parent_map)
1532
source_keys = self._get_ordered_source_keys(ordering, parent_map,
1534
elif ordering == 'as-requested':
1535
source_keys = self._get_as_requested_source_keys(orig_keys,
1536
locations, unadded_keys, key_to_source_map)
1538
# We want to yield the keys in a semi-optimal (read-wise) ordering.
1539
# Otherwise we thrash the _group_cache and destroy performance
1540
source_keys = self._get_io_ordered_source_keys(locations,
1541
unadded_keys, source_result)
1543
yield AbsentContentFactory(key)
1544
# Batch up as many keys as we can until either:
1545
# - we encounter an unadded ref, or
1546
# - we run out of keys, or
1547
# - the total bytes to retrieve for this batch > BATCH_SIZE
1548
batcher = _BatchingBlockFetcher(self, locations)
1549
for source, keys in source_keys:
1552
if key in self._unadded_refs:
1553
# Flush batch, then yield unadded ref from
1555
for factory in batcher.yield_factories(full_flush=True):
1557
bytes, sha1 = self._compressor.extract(key)
1558
parents = self._unadded_refs[key]
1559
yield FulltextContentFactory(key, parents, sha1, bytes)
1561
if batcher.add_key(key) > BATCH_SIZE:
1562
# Ok, this batch is big enough. Yield some results.
1563
for factory in batcher.yield_factories():
1566
for factory in batcher.yield_factories(full_flush=True):
1568
for record in source.get_record_stream(keys, ordering,
1569
include_delta_closure):
1571
for factory in batcher.yield_factories(full_flush=True):
1574
def get_sha1s(self, keys):
1575
"""See VersionedFiles.get_sha1s()."""
1577
for record in self.get_record_stream(keys, 'unordered', True):
1578
if record.sha1 != None:
1579
result[record.key] = record.sha1
1581
if record.storage_kind != 'absent':
1582
result[record.key] = osutils.sha_string(
1583
record.get_bytes_as('fulltext'))
1586
def insert_record_stream(self, stream):
1587
"""Insert a record stream into this container.
1589
:param stream: A stream of records to insert.
1591
:seealso VersionedFiles.get_record_stream:
1593
# XXX: Setting random_id=True makes
1594
# test_insert_record_stream_existing_keys fail for groupcompress and
1595
# groupcompress-nograph, this needs to be revisited while addressing
1596
# 'bzr branch' performance issues.
1597
for _ in self._insert_record_stream(stream, random_id=False):
1600
def _insert_record_stream(self, stream, random_id=False, nostore_sha=None,
1602
"""Internal core to insert a record stream into this container.
1604
This helper function has a different interface than insert_record_stream
1605
to allow add_lines to be minimal, but still return the needed data.
1607
:param stream: A stream of records to insert.
1608
:param nostore_sha: If the sha1 of a given text matches nostore_sha,
1609
raise ExistingContent, rather than committing the new text.
1610
:param reuse_blocks: If the source is streaming from
1611
groupcompress-blocks, just insert the blocks as-is, rather than
1612
expanding the texts and inserting again.
1613
:return: An iterator over the sha1 of the inserted records.
1614
:seealso insert_record_stream:
1618
def get_adapter(adapter_key):
1620
return adapters[adapter_key]
1622
adapter_factory = adapter_registry.get(adapter_key)
1623
adapter = adapter_factory(self)
1624
adapters[adapter_key] = adapter
1626
# This will go up to fulltexts for gc to gc fetching, which isn't
1628
self._compressor = GroupCompressor()
1629
self._unadded_refs = {}
1632
bytes = self._compressor.flush().to_bytes()
1633
index, start, length = self._access.add_raw_records(
1634
[(None, len(bytes))], bytes)[0]
1636
for key, reads, refs in keys_to_add:
1637
nodes.append((key, "%d %d %s" % (start, length, reads), refs))
1638
self._index.add_records(nodes, random_id=random_id)
1639
self._unadded_refs = {}
1641
self._compressor = GroupCompressor()
1644
max_fulltext_len = 0
1645
max_fulltext_prefix = None
1646
insert_manager = None
1649
# XXX: TODO: remove this, it is just for safety checking for now
1650
inserted_keys = set()
1651
reuse_this_block = reuse_blocks
1652
for record in stream:
1653
# Raise an error when a record is missing.
1654
if record.storage_kind == 'absent':
1655
raise errors.RevisionNotPresent(record.key, self)
1657
if record.key in inserted_keys:
1658
trace.note('Insert claimed random_id=True,'
1659
' but then inserted %r two times', record.key)
1661
inserted_keys.add(record.key)
1663
# If the reuse_blocks flag is set, check to see if we can just
1664
# copy a groupcompress block as-is.
1665
# We only check on the first record (groupcompress-block) not
1666
# on all of the (groupcompress-block-ref) entries.
1667
# The reuse_this_block flag is then kept for as long as
1668
if record.storage_kind == 'groupcompress-block':
1669
# Check to see if we really want to re-use this block
1670
insert_manager = record._manager
1671
reuse_this_block = insert_manager.check_is_well_utilized()
1673
reuse_this_block = False
1674
if reuse_this_block:
1675
# We still want to reuse this block
1676
if record.storage_kind == 'groupcompress-block':
1677
# Insert the raw block into the target repo
1678
insert_manager = record._manager
1679
bytes = record._manager._block.to_bytes()
1680
_, start, length = self._access.add_raw_records(
1681
[(None, len(bytes))], bytes)[0]
1684
block_length = length
1685
if record.storage_kind in ('groupcompress-block',
1686
'groupcompress-block-ref'):
1687
if insert_manager is None:
1688
raise AssertionError('No insert_manager set')
1689
if insert_manager is not record._manager:
1690
raise AssertionError('insert_manager does not match'
1691
' the current record, we cannot be positive'
1692
' that the appropriate content was inserted.'
1694
value = "%d %d %d %d" % (block_start, block_length,
1695
record._start, record._end)
1696
nodes = [(record.key, value, (record.parents,))]
1697
# TODO: Consider buffering up many nodes to be added, not
1698
# sure how much overhead this has, but we're seeing
1699
# ~23s / 120s in add_records calls
1700
self._index.add_records(nodes, random_id=random_id)
1703
bytes = record.get_bytes_as('fulltext')
1704
except errors.UnavailableRepresentation:
1705
adapter_key = record.storage_kind, 'fulltext'
1706
adapter = get_adapter(adapter_key)
1707
bytes = adapter.get_bytes(record)
1708
if len(record.key) > 1:
1709
prefix = record.key[0]
1710
soft = (prefix == last_prefix)
1714
if max_fulltext_len < len(bytes):
1715
max_fulltext_len = len(bytes)
1716
max_fulltext_prefix = prefix
1717
(found_sha1, start_point, end_point,
1718
type) = self._compressor.compress(record.key,
1719
bytes, record.sha1, soft=soft,
1720
nostore_sha=nostore_sha)
1721
# delta_ratio = float(len(bytes)) / (end_point - start_point)
1722
# Check if we want to continue to include that text
1723
if (prefix == max_fulltext_prefix
1724
and end_point < 2 * max_fulltext_len):
1725
# As long as we are on the same file_id, we will fill at least
1726
# 2 * max_fulltext_len
1727
start_new_block = False
1728
elif end_point > 4*1024*1024:
1729
start_new_block = True
1730
elif (prefix is not None and prefix != last_prefix
1731
and end_point > 2*1024*1024):
1732
start_new_block = True
1734
start_new_block = False
1735
last_prefix = prefix
1737
self._compressor.pop_last()
1739
max_fulltext_len = len(bytes)
1740
(found_sha1, start_point, end_point,
1741
type) = self._compressor.compress(record.key, bytes,
1743
if record.key[-1] is None:
1744
key = record.key[:-1] + ('sha1:' + found_sha1,)
1747
self._unadded_refs[key] = record.parents
1749
keys_to_add.append((key, '%d %d' % (start_point, end_point),
1751
if len(keys_to_add):
1753
self._compressor = None
1755
def iter_lines_added_or_present_in_keys(self, keys, pb=None):
1756
"""Iterate over the lines in the versioned files from keys.
1758
This may return lines from other keys. Each item the returned
1759
iterator yields is a tuple of a line and a text version that that line
1760
is present in (not introduced in).
1762
Ordering of results is in whatever order is most suitable for the
1763
underlying storage format.
1765
If a progress bar is supplied, it may be used to indicate progress.
1766
The caller is responsible for cleaning up progress bars (because this
1770
* Lines are normalised by the underlying store: they will all have \n
1772
* Lines are returned in arbitrary order.
1774
:return: An iterator over (line, key).
1778
# we don't care about inclusions, the caller cares.
1779
# but we need to setup a list of records to visit.
1780
# we need key, position, length
1781
for key_idx, record in enumerate(self.get_record_stream(keys,
1782
'unordered', True)):
1783
# XXX: todo - optimise to use less than full texts.
1786
pb.update('Walking content', key_idx, total)
1787
if record.storage_kind == 'absent':
1788
raise errors.RevisionNotPresent(key, self)
1789
lines = osutils.split_lines(record.get_bytes_as('fulltext'))
1793
pb.update('Walking content', total, total)
1796
"""See VersionedFiles.keys."""
1797
if 'evil' in debug.debug_flags:
1798
trace.mutter_callsite(2, "keys scales with size of history")
1799
sources = [self._index] + self._fallback_vfs
1801
for source in sources:
1802
result.update(source.keys())
1806
class _GCGraphIndex(object):
1807
"""Mapper from GroupCompressVersionedFiles needs into GraphIndex storage."""
1809
def __init__(self, graph_index, is_locked, parents=True,
1810
add_callback=None, track_external_parent_refs=False,
1811
inconsistency_fatal=True):
1812
"""Construct a _GCGraphIndex on a graph_index.
1814
:param graph_index: An implementation of bzrlib.index.GraphIndex.
1815
:param is_locked: A callback, returns True if the index is locked and
1817
:param parents: If True, record knits parents, if not do not record
1819
:param add_callback: If not None, allow additions to the index and call
1820
this callback with a list of added GraphIndex nodes:
1821
[(node, value, node_refs), ...]
1822
:param track_external_parent_refs: As keys are added, keep track of the
1823
keys they reference, so that we can query get_missing_parents(),
1825
:param inconsistency_fatal: When asked to add records that are already
1826
present, and the details are inconsistent with the existing
1827
record, raise an exception instead of warning (and skipping the
1830
self._add_callback = add_callback
1831
self._graph_index = graph_index
1832
self._parents = parents
1833
self.has_graph = parents
1834
self._is_locked = is_locked
1835
self._inconsistency_fatal = inconsistency_fatal
1836
if track_external_parent_refs:
1837
self._key_dependencies = knit._KeyRefs()
1839
self._key_dependencies = None
1841
def add_records(self, records, random_id=False):
1842
"""Add multiple records to the index.
1844
This function does not insert data into the Immutable GraphIndex
1845
backing the KnitGraphIndex, instead it prepares data for insertion by
1846
the caller and checks that it is safe to insert then calls
1847
self._add_callback with the prepared GraphIndex nodes.
1849
:param records: a list of tuples:
1850
(key, options, access_memo, parents).
1851
:param random_id: If True the ids being added were randomly generated
1852
and no check for existence will be performed.
1854
if not self._add_callback:
1855
raise errors.ReadOnlyError(self)
1856
# we hope there are no repositories with inconsistent parentage
1861
for (key, value, refs) in records:
1862
if not self._parents:
1866
raise errors.KnitCorrupt(self,
1867
"attempt to add node with parents "
1868
"in parentless index.")
1871
keys[key] = (value, refs)
1874
present_nodes = self._get_entries(keys)
1875
for (index, key, value, node_refs) in present_nodes:
1876
if node_refs != keys[key][1]:
1877
details = '%s %s %s' % (key, (value, node_refs), keys[key])
1878
if self._inconsistency_fatal:
1879
raise errors.KnitCorrupt(self, "inconsistent details"
1880
" in add_records: %s" %
1883
trace.warning("inconsistent details in skipped"
1884
" record: %s", details)
1890
for key, (value, node_refs) in keys.iteritems():
1891
result.append((key, value, node_refs))
1893
for key, (value, node_refs) in keys.iteritems():
1894
result.append((key, value))
1896
key_dependencies = self._key_dependencies
1897
if key_dependencies is not None and self._parents:
1898
for key, value, refs in records:
1900
key_dependencies.add_references(key, parents)
1901
self._add_callback(records)
1903
def _check_read(self):
1904
"""Raise an exception if reads are not permitted."""
1905
if not self._is_locked():
1906
raise errors.ObjectNotLocked(self)
1908
def _check_write_ok(self):
1909
"""Raise an exception if writes are not permitted."""
1910
if not self._is_locked():
1911
raise errors.ObjectNotLocked(self)
1913
def _get_entries(self, keys, check_present=False):
1914
"""Get the entries for keys.
1916
Note: Callers are responsible for checking that the index is locked
1917
before calling this method.
1919
:param keys: An iterable of index key tuples.
1924
for node in self._graph_index.iter_entries(keys):
1926
found_keys.add(node[1])
1928
# adapt parentless index to the rest of the code.
1929
for node in self._graph_index.iter_entries(keys):
1930
yield node[0], node[1], node[2], ()
1931
found_keys.add(node[1])
1933
missing_keys = keys.difference(found_keys)
1935
raise errors.RevisionNotPresent(missing_keys.pop(), self)
1937
def find_ancestry(self, keys):
1938
"""See CombinedGraphIndex.find_ancestry"""
1939
return self._graph_index.find_ancestry(keys, 0)
1941
def get_parent_map(self, keys):
1942
"""Get a map of the parents of keys.
1944
:param keys: The keys to look up parents for.
1945
:return: A mapping from keys to parents. Absent keys are absent from
1949
nodes = self._get_entries(keys)
1953
result[node[1]] = node[3][0]
1956
result[node[1]] = None
1959
def get_missing_parents(self):
1960
"""Return the keys of missing parents."""
1961
# Copied from _KnitGraphIndex.get_missing_parents
1962
# We may have false positives, so filter those out.
1963
self._key_dependencies.add_keys(
1964
self.get_parent_map(self._key_dependencies.get_unsatisfied_refs()))
1965
return frozenset(self._key_dependencies.get_unsatisfied_refs())
1967
def get_build_details(self, keys):
1968
"""Get the various build details for keys.
1970
Ghosts are omitted from the result.
1972
:param keys: An iterable of keys.
1973
:return: A dict of key:
1974
(index_memo, compression_parent, parents, record_details).
1976
opaque structure to pass to read_records to extract the raw
1979
Content that this record is built upon, may be None
1981
Logical parents of this node
1983
extra information about the content which needs to be passed to
1984
Factory.parse_record
1988
entries = self._get_entries(keys)
1989
for entry in entries:
1991
if not self._parents:
1994
parents = entry[3][0]
1996
result[key] = (self._node_to_position(entry),
1997
None, parents, (method, None))
2001
"""Get all the keys in the collection.
2003
The keys are not ordered.
2006
return [node[1] for node in self._graph_index.iter_all_entries()]
2008
def _node_to_position(self, node):
2009
"""Convert an index value to position details."""
2010
bits = node[2].split(' ')
2011
# It would be nice not to read the entire gzip.
2012
start = int(bits[0])
2014
basis_end = int(bits[2])
2015
delta_end = int(bits[3])
2016
return node[0], start, stop, basis_end, delta_end
2018
def scan_unvalidated_index(self, graph_index):
2019
"""Inform this _GCGraphIndex that there is an unvalidated index.
2021
This allows this _GCGraphIndex to keep track of any missing
2022
compression parents we may want to have filled in to make those
2025
:param graph_index: A GraphIndex
2027
if self._key_dependencies is not None:
2028
# Add parent refs from graph_index (and discard parent refs that
2029
# the graph_index has).
2030
add_refs = self._key_dependencies.add_references
2031
for node in graph_index.iter_all_entries():
2032
add_refs(node[1], node[3][0])
2036
from bzrlib._groupcompress_py import (
2038
apply_delta_to_source,
2041
decode_copy_instruction,
2045
from bzrlib._groupcompress_pyx import (
2047
apply_delta_to_source,
2052
GroupCompressor = PyrexGroupCompressor
2054
GroupCompressor = PythonGroupCompressor
added
removed
bzrlib/groupcompress.py
