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.graph import Graph
37
from bzrlib.btree_index import BTreeBuilder
38
from bzrlib.lru_cache import LRUSizeCache
39
from bzrlib.tsort import topo_sort
40
from bzrlib.versionedfile import (
43
ChunkedContentFactory,
44
FulltextContentFactory,
48
_USE_LZMA = False and (pylzma is not None)
50
# osutils.sha_string('')
51
_null_sha1 = 'da39a3ee5e6b4b0d3255bfef95601890afd80709'
54
def sort_gc_optimal(parent_map):
55
"""Sort and group the keys in parent_map into groupcompress order.
57
groupcompress is defined (currently) as reverse-topological order, grouped
60
:return: A sorted-list of keys
62
# groupcompress ordering is approximately reverse topological,
63
# properly grouped by file-id.
65
for item in parent_map.iteritems():
67
if isinstance(key, str) or len(key) == 1:
72
per_prefix_map[prefix].append(item)
74
per_prefix_map[prefix] = [item]
77
for prefix in sorted(per_prefix_map):
78
present_keys.extend(reversed(topo_sort(per_prefix_map[prefix])))
82
# The max zlib window size is 32kB, so if we set 'max_size' output of the
83
# decompressor to the requested bytes + 32kB, then we should guarantee
84
# num_bytes coming out.
85
_ZLIB_DECOMP_WINDOW = 32*1024
87
class GroupCompressBlock(object):
88
"""An object which maintains the internal structure of the compressed data.
90
This tracks the meta info (start of text, length, type, etc.)
93
# Group Compress Block v1 Zlib
94
GCB_HEADER = 'gcb1z\n'
95
# Group Compress Block v1 Lzma
96
GCB_LZ_HEADER = 'gcb1l\n'
97
GCB_KNOWN_HEADERS = (GCB_HEADER, GCB_LZ_HEADER)
100
# map by key? or just order in file?
101
self._compressor_name = None
102
self._z_content = None
103
self._z_content_decompressor = None
104
self._z_content_length = None
105
self._content_length = None
107
self._content_chunks = None
110
# This is the maximum number of bytes this object will reference if
111
# everything is decompressed. However, if we decompress less than
112
# everything... (this would cause some problems for LRUSizeCache)
113
return self._content_length + self._z_content_length
115
def _ensure_content(self, num_bytes=None):
116
"""Make sure that content has been expanded enough.
118
:param num_bytes: Ensure that we have extracted at least num_bytes of
119
content. If None, consume everything
121
# TODO: If we re-use the same content block at different times during
122
# get_record_stream(), it is possible that the first pass will
123
# get inserted, triggering an extract/_ensure_content() which
124
# will get rid of _z_content. And then the next use of the block
125
# will try to access _z_content (to send it over the wire), and
126
# fail because it is already extracted. Consider never releasing
127
# _z_content because of this.
128
if num_bytes is None:
129
num_bytes = self._content_length
130
elif (self._content_length is not None
131
and num_bytes > self._content_length):
132
raise AssertionError(
133
'requested num_bytes (%d) > content length (%d)'
134
% (num_bytes, self._content_length))
135
# Expand the content if required
136
if self._content is None:
137
if self._content_chunks is not None:
138
self._content = ''.join(self._content_chunks)
139
self._content_chunks = None
140
if self._content is None:
141
if self._z_content is None:
142
raise AssertionError('No content to decompress')
143
if self._z_content == '':
145
elif self._compressor_name == 'lzma':
146
# We don't do partial lzma decomp yet
147
self._content = pylzma.decompress(self._z_content)
148
elif self._compressor_name == 'zlib':
149
# Start a zlib decompressor
150
if num_bytes is None:
151
self._content = zlib.decompress(self._z_content)
153
self._z_content_decompressor = zlib.decompressobj()
154
# Seed the decompressor with the uncompressed bytes, so
155
# that the rest of the code is simplified
156
self._content = self._z_content_decompressor.decompress(
157
self._z_content, num_bytes + _ZLIB_DECOMP_WINDOW)
159
raise AssertionError('Unknown compressor: %r'
160
% self._compressor_name)
161
# Any bytes remaining to be decompressed will be in the decompressors
164
# 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
170
# If we got this far, and don't have a decompressor, something is wrong
171
if self._z_content_decompressor is None:
172
raise AssertionError(
173
'No decompressor to decompress %d bytes' % num_bytes)
174
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
205
def _parse_bytes(self, bytes, pos):
206
"""Read the various lengths from the header.
208
This also populates the various 'compressed' buffers.
210
:return: The position in bytes just after the last newline
212
# At present, we have 2 integers for the compressed and uncompressed
213
# content. In base10 (ascii) 14 bytes can represent > 1TB, so to avoid
214
# checking too far, cap the search to 14 bytes.
215
pos2 = bytes.index('\n', pos, pos + 14)
216
self._z_content_length = int(bytes[pos:pos2])
218
pos2 = bytes.index('\n', pos, pos + 14)
219
self._content_length = int(bytes[pos:pos2])
221
if len(bytes) != (pos + self._z_content_length):
222
# XXX: Define some GCCorrupt error ?
223
raise AssertionError('Invalid bytes: (%d) != %d + %d' %
224
(len(bytes), pos, self._z_content_length))
225
self._z_content = bytes[pos:]
228
def from_bytes(cls, bytes):
230
if bytes[:6] not in cls.GCB_KNOWN_HEADERS:
231
raise ValueError('bytes did not start with any of %r'
232
% (cls.GCB_KNOWN_HEADERS,))
233
# XXX: why not testing the whole header ?
235
out._compressor_name = 'zlib'
236
elif bytes[4] == 'l':
237
out._compressor_name = 'lzma'
239
raise ValueError('unknown compressor: %r' % (bytes,))
240
out._parse_bytes(bytes, 6)
243
def extract(self, key, start, end, sha1=None):
244
"""Extract the text for a specific key.
246
:param key: The label used for this content
247
:param sha1: TODO (should we validate only when sha1 is supplied?)
248
:return: The bytes for the content
250
if start == end == 0:
252
self._ensure_content(end)
253
# The bytes are 'f' or 'd' for the type, then a variable-length
254
# base128 integer for the content size, then the actual content
255
# We know that the variable-length integer won't be longer than 5
256
# bytes (it takes 5 bytes to encode 2^32)
257
c = self._content[start]
262
raise ValueError('Unknown content control code: %s'
265
content_len, len_len = decode_base128_int(
266
self._content[start + 1:start + 6])
267
content_start = start + 1 + len_len
268
if end != content_start + content_len:
269
raise ValueError('end != len according to field header'
270
' %s != %s' % (end, content_start + content_len))
272
bytes = self._content[content_start:end]
274
bytes = apply_delta_to_source(self._content, content_start, end)
277
def set_chunked_content(self, content_chunks, length):
278
"""Set the content of this block to the given chunks."""
279
# If we have lots of short lines, it is may be more efficient to join
280
# the content ahead of time. If the content is <10MiB, we don't really
281
# care about the extra memory consumption, so we can just pack it and
282
# be done. However, timing showed 18s => 17.9s for repacking 1k revs of
283
# mysql, which is below the noise margin
284
self._content_length = length
285
self._content_chunks = content_chunks
287
self._z_content = None
289
def set_content(self, content):
290
"""Set the content of this block."""
291
self._content_length = len(content)
292
self._content = content
293
self._z_content = None
295
def _create_z_content_using_lzma(self):
296
if self._content_chunks is not None:
297
self._content = ''.join(self._content_chunks)
298
self._content_chunks = None
299
if self._content is None:
300
raise AssertionError('Nothing to compress')
301
self._z_content = pylzma.compress(self._content)
302
self._z_content_length = len(self._z_content)
304
def _create_z_content_from_chunks(self):
305
compressor = zlib.compressobj(zlib.Z_DEFAULT_COMPRESSION)
306
compressed_chunks = map(compressor.compress, self._content_chunks)
307
compressed_chunks.append(compressor.flush())
308
self._z_content = ''.join(compressed_chunks)
309
self._z_content_length = len(self._z_content)
311
def _create_z_content(self):
312
if self._z_content is not None:
315
self._create_z_content_using_lzma()
317
if self._content_chunks is not None:
318
self._create_z_content_from_chunks()
320
self._z_content = zlib.compress(self._content)
321
self._z_content_length = len(self._z_content)
324
"""Encode the information into a byte stream."""
325
self._create_z_content()
327
header = self.GCB_LZ_HEADER
329
header = self.GCB_HEADER
331
'%d\n%d\n' % (self._z_content_length, self._content_length),
334
return ''.join(chunks)
336
def _dump(self, include_text=False):
337
"""Take this block, and spit out a human-readable structure.
339
:param include_text: Inserts also include text bits, chose whether you
340
want this displayed in the dump or not.
341
:return: A dump of the given block. The layout is something like:
342
[('f', length), ('d', delta_length, text_length, [delta_info])]
343
delta_info := [('i', num_bytes, text), ('c', offset, num_bytes),
346
self._ensure_content()
349
while pos < self._content_length:
350
kind = self._content[pos]
352
if kind not in ('f', 'd'):
353
raise ValueError('invalid kind character: %r' % (kind,))
354
content_len, len_len = decode_base128_int(
355
self._content[pos:pos + 5])
357
if content_len + pos > self._content_length:
358
raise ValueError('invalid content_len %d for record @ pos %d'
359
% (content_len, pos - len_len - 1))
360
if kind == 'f': # Fulltext
362
text = self._content[pos:pos+content_len]
363
result.append(('f', content_len, text))
365
result.append(('f', content_len))
366
elif kind == 'd': # Delta
367
delta_content = self._content[pos:pos+content_len]
369
# The first entry in a delta is the decompressed length
370
decomp_len, delta_pos = decode_base128_int(delta_content)
371
result.append(('d', content_len, decomp_len, delta_info))
373
while delta_pos < content_len:
374
c = ord(delta_content[delta_pos])
378
delta_pos) = decode_copy_instruction(delta_content, c,
381
text = self._content[offset:offset+length]
382
delta_info.append(('c', offset, length, text))
384
delta_info.append(('c', offset, length))
385
measured_len += length
388
txt = delta_content[delta_pos:delta_pos+c]
391
delta_info.append(('i', c, txt))
394
if delta_pos != content_len:
395
raise ValueError('Delta consumed a bad number of bytes:'
396
' %d != %d' % (delta_pos, content_len))
397
if measured_len != decomp_len:
398
raise ValueError('Delta claimed fulltext was %d bytes, but'
399
' extraction resulted in %d bytes'
400
% (decomp_len, measured_len))
405
class _LazyGroupCompressFactory(object):
406
"""Yield content from a GroupCompressBlock on demand."""
408
def __init__(self, key, parents, manager, start, end, first):
409
"""Create a _LazyGroupCompressFactory
411
:param key: The key of just this record
412
:param parents: The parents of this key (possibly None)
413
:param gc_block: A GroupCompressBlock object
414
:param start: Offset of the first byte for this record in the
416
:param end: Offset of the byte just after the end of this record
417
(ie, bytes = content[start:end])
418
:param first: Is this the first Factory for the given block?
421
self.parents = parents
423
# Note: This attribute coupled with Manager._factories creates a
424
# reference cycle. Perhaps we would rather use a weakref(), or
425
# find an appropriate time to release the ref. After the first
426
# get_bytes_as call? After Manager.get_record_stream() returns
428
self._manager = manager
430
self.storage_kind = 'groupcompress-block'
432
self.storage_kind = 'groupcompress-block-ref'
438
return '%s(%s, first=%s)' % (self.__class__.__name__,
439
self.key, self._first)
441
def get_bytes_as(self, storage_kind):
442
if storage_kind == self.storage_kind:
444
# wire bytes, something...
445
return self._manager._wire_bytes()
448
if storage_kind in ('fulltext', 'chunked'):
449
if self._bytes is None:
450
# Grab and cache the raw bytes for this entry
451
# and break the ref-cycle with _manager since we don't need it
453
self._manager._prepare_for_extract()
454
block = self._manager._block
455
self._bytes = block.extract(self.key, self._start, self._end)
456
# There are code paths that first extract as fulltext, and then
457
# extract as storage_kind (smart fetch). So we don't break the
458
# refcycle here, but instead in manager.get_record_stream()
459
# self._manager = None
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
def __init__(self, block):
473
# We need to preserve the ordering
477
def add_factory(self, key, parents, start, end):
478
if not self._factories:
482
# Note that this creates a reference cycle....
483
factory = _LazyGroupCompressFactory(key, parents, self,
484
start, end, first=first)
485
# max() works here, but as a function call, doing a compare seems to be
486
# significantly faster, timeit says 250ms for max() and 100ms for the
488
if end > self._last_byte:
489
self._last_byte = end
490
self._factories.append(factory)
492
def get_record_stream(self):
493
"""Get a record for all keys added so far."""
494
for factory in self._factories:
496
# Break the ref-cycle
497
factory._bytes = None
498
factory._manager = None
499
# TODO: Consider setting self._factories = None after the above loop,
500
# as it will break the reference cycle
502
def _trim_block(self, last_byte):
503
"""Create a new GroupCompressBlock, with just some of the content."""
504
# None of the factories need to be adjusted, because the content is
505
# located in an identical place. Just that some of the unreferenced
506
# trailing bytes are stripped
507
trace.mutter('stripping trailing bytes from groupcompress block'
508
' %d => %d', self._block._content_length, last_byte)
509
new_block = GroupCompressBlock()
510
self._block._ensure_content(last_byte)
511
new_block.set_content(self._block._content[:last_byte])
512
self._block = new_block
514
def _rebuild_block(self):
515
"""Create a new GroupCompressBlock with only the referenced texts."""
516
compressor = GroupCompressor()
518
old_length = self._block._content_length
520
for factory in self._factories:
521
bytes = factory.get_bytes_as('fulltext')
522
(found_sha1, start_point, end_point,
523
type) = compressor.compress(factory.key, bytes, factory.sha1)
524
# Now update this factory with the new offsets, etc
525
factory.sha1 = found_sha1
526
factory._start = start_point
527
factory._end = end_point
528
self._last_byte = end_point
529
new_block = compressor.flush()
530
# TODO: Should we check that new_block really *is* smaller than the old
531
# block? It seems hard to come up with a method that it would
532
# expand, since we do full compression again. Perhaps based on a
533
# request that ends up poorly ordered?
534
delta = time.time() - tstart
535
self._block = new_block
536
trace.mutter('creating new compressed block on-the-fly in %.3fs'
537
' %d bytes => %d bytes', delta, old_length,
538
self._block._content_length)
540
def _prepare_for_extract(self):
541
"""A _LazyGroupCompressFactory is about to extract to fulltext."""
542
# We expect that if one child is going to fulltext, all will be. This
543
# helps prevent all of them from extracting a small amount at a time.
544
# Which in itself isn't terribly expensive, but resizing 2MB 32kB at a
545
# time (self._block._content) is a little expensive.
546
self._block._ensure_content(self._last_byte)
548
def _check_rebuild_block(self):
549
"""Check to see if our block should be repacked."""
552
for factory in self._factories:
553
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)
557
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
561
if total_bytes_used * 2 > last_byte_used:
562
self._trim_block(last_byte_used)
565
# We are using a small amount of the data, and it isn't just packed
566
# nicely at the front, so rebuild the content.
567
# Note: This would be *nicer* as a strip-data-from-group, rather than
568
# building it up again from scratch
569
# It might be reasonable to consider the fulltext sizes for
570
# different bits when deciding this, too. As you may have a small
571
# fulltext, and a trivial delta, and you are just trading around
572
# for another fulltext. If we do a simple 'prune' you may end up
573
# expanding many deltas into fulltexts, as well.
574
# If we build a cheap enough 'strip', then we could try a strip,
575
# if that expands the content, we then rebuild.
576
self._rebuild_block()
578
def _wire_bytes(self):
579
"""Return a byte stream suitable for transmitting over the wire."""
580
self._check_rebuild_block()
581
# The outer block starts with:
582
# 'groupcompress-block\n'
583
# <length of compressed key info>\n
584
# <length of uncompressed info>\n
585
# <length of gc block>\n
588
lines = ['groupcompress-block\n']
589
# The minimal info we need is the key, the start offset, and the
590
# parents. The length and type are encoded in the record itself.
591
# However, passing in the other bits makes it easier. The list of
592
# keys, and the start offset, the length
594
# 1 line with parents, '' for ()
595
# 1 line for start offset
596
# 1 line for end byte
598
for factory in self._factories:
599
key_bytes = '\x00'.join(factory.key)
600
parents = factory.parents
602
parent_bytes = 'None:'
604
parent_bytes = '\t'.join('\x00'.join(key) for key in parents)
605
record_header = '%s\n%s\n%d\n%d\n' % (
606
key_bytes, parent_bytes, factory._start, factory._end)
607
header_lines.append(record_header)
608
# TODO: Can we break the refcycle at this point and set
609
# factory._manager = None?
610
header_bytes = ''.join(header_lines)
612
header_bytes_len = len(header_bytes)
613
z_header_bytes = zlib.compress(header_bytes)
615
z_header_bytes_len = len(z_header_bytes)
616
block_bytes = self._block.to_bytes()
617
lines.append('%d\n%d\n%d\n' % (z_header_bytes_len, header_bytes_len,
619
lines.append(z_header_bytes)
620
lines.append(block_bytes)
621
del z_header_bytes, block_bytes
622
return ''.join(lines)
625
def from_bytes(cls, bytes):
626
# TODO: This does extra string copying, probably better to do it a
628
(storage_kind, z_header_len, header_len,
629
block_len, rest) = bytes.split('\n', 4)
631
if storage_kind != 'groupcompress-block':
632
raise ValueError('Unknown storage kind: %s' % (storage_kind,))
633
z_header_len = int(z_header_len)
634
if len(rest) < z_header_len:
635
raise ValueError('Compressed header len shorter than all bytes')
636
z_header = rest[:z_header_len]
637
header_len = int(header_len)
638
header = zlib.decompress(z_header)
639
if len(header) != header_len:
640
raise ValueError('invalid length for decompressed bytes')
642
block_len = int(block_len)
643
if len(rest) != z_header_len + block_len:
644
raise ValueError('Invalid length for block')
645
block_bytes = rest[z_header_len:]
647
# So now we have a valid GCB, we just need to parse the factories that
649
header_lines = header.split('\n')
651
last = header_lines.pop()
653
raise ValueError('header lines did not end with a trailing'
655
if len(header_lines) % 4 != 0:
656
raise ValueError('The header was not an even multiple of 4 lines')
657
block = GroupCompressBlock.from_bytes(block_bytes)
660
for start in xrange(0, len(header_lines), 4):
662
key = tuple(header_lines[start].split('\x00'))
663
parents_line = header_lines[start+1]
664
if parents_line == 'None:':
667
parents = tuple([tuple(segment.split('\x00'))
668
for segment in parents_line.split('\t')
670
start_offset = int(header_lines[start+2])
671
end_offset = int(header_lines[start+3])
672
result.add_factory(key, parents, start_offset, end_offset)
676
def network_block_to_records(storage_kind, bytes, line_end):
677
if storage_kind != 'groupcompress-block':
678
raise ValueError('Unknown storage kind: %s' % (storage_kind,))
679
manager = _LazyGroupContentManager.from_bytes(bytes)
680
return manager.get_record_stream()
683
class _CommonGroupCompressor(object):
686
"""Create a GroupCompressor."""
691
self.labels_deltas = {}
692
self._delta_index = None # Set by the children
693
self._block = GroupCompressBlock()
695
def compress(self, key, bytes, expected_sha, nostore_sha=None, soft=False):
696
"""Compress lines with label key.
698
:param key: A key tuple. It is stored in the output
699
for identification of the text during decompression. If the last
700
element is 'None' it is replaced with the sha1 of the text -
702
:param bytes: The bytes to be compressed
703
:param expected_sha: If non-None, the sha the lines are believed to
704
have. During compression the sha is calculated; a mismatch will
706
:param nostore_sha: If the computed sha1 sum matches, we will raise
707
ExistingContent rather than adding the text.
708
:param soft: Do a 'soft' compression. This means that we require larger
709
ranges to match to be considered for a copy command.
711
:return: The sha1 of lines, the start and end offsets in the delta, and
712
the type ('fulltext' or 'delta').
714
:seealso VersionedFiles.add_lines:
716
if not bytes: # empty, like a dir entry, etc
717
if nostore_sha == _null_sha1:
718
raise errors.ExistingContent()
719
return _null_sha1, 0, 0, 'fulltext'
720
# we assume someone knew what they were doing when they passed it in
721
if expected_sha is not None:
724
sha1 = osutils.sha_string(bytes)
725
if nostore_sha is not None:
726
if sha1 == nostore_sha:
727
raise errors.ExistingContent()
729
key = key[:-1] + ('sha1:' + sha1,)
731
start, end, type = self._compress(key, bytes, len(bytes) / 2, soft)
732
return sha1, start, end, type
734
def _compress(self, key, bytes, max_delta_size, soft=False):
735
"""Compress lines with label key.
737
:param key: A key tuple. It is stored in the output for identification
738
of the text during decompression.
740
:param bytes: The bytes to be compressed
742
:param max_delta_size: The size above which we issue a fulltext instead
745
:param soft: Do a 'soft' compression. This means that we require larger
746
ranges to match to be considered for a copy command.
748
:return: The sha1 of lines, the start and end offsets in the delta, and
749
the type ('fulltext' or 'delta').
751
raise NotImplementedError(self._compress)
753
def extract(self, key):
754
"""Extract a key previously added to the compressor.
756
:param key: The key to extract.
757
:return: An iterable over bytes and the sha1.
759
(start_byte, start_chunk, end_byte, end_chunk) = self.labels_deltas[key]
760
delta_chunks = self.chunks[start_chunk:end_chunk]
761
stored_bytes = ''.join(delta_chunks)
762
if stored_bytes[0] == 'f':
763
fulltext_len, offset = decode_base128_int(stored_bytes[1:10])
764
data_len = fulltext_len + 1 + offset
765
if data_len != len(stored_bytes):
766
raise ValueError('Index claimed fulltext len, but stored bytes'
768
% (len(stored_bytes), data_len))
769
bytes = stored_bytes[offset + 1:]
771
# XXX: This is inefficient at best
772
source = ''.join(self.chunks[:start_chunk])
773
if stored_bytes[0] != 'd':
774
raise ValueError('Unknown content kind, bytes claim %s'
775
% (stored_bytes[0],))
776
delta_len, offset = decode_base128_int(stored_bytes[1:10])
777
data_len = delta_len + 1 + offset
778
if data_len != len(stored_bytes):
779
raise ValueError('Index claimed delta len, but stored bytes'
781
% (len(stored_bytes), data_len))
782
bytes = apply_delta(source, stored_bytes[offset + 1:])
783
bytes_sha1 = osutils.sha_string(bytes)
784
return bytes, bytes_sha1
787
"""Finish this group, creating a formatted stream.
789
After calling this, the compressor should no longer be used
791
# TODO: this causes us to 'bloat' to 2x the size of content in the
792
# group. This has an impact for 'commit' of large objects.
793
# One possibility is to use self._content_chunks, and be lazy and
794
# only fill out self._content as a full string when we actually
795
# need it. That would at least drop the peak memory consumption
796
# for 'commit' down to ~1x the size of the largest file, at a
797
# cost of increased complexity within this code. 2x is still <<
798
# 3x the size of the largest file, so we are doing ok.
799
self._block.set_chunked_content(self.chunks, self.endpoint)
801
self._delta_index = None
805
"""Call this if you want to 'revoke' the last compression.
807
After this, the data structures will be rolled back, but you cannot do
810
self._delta_index = None
811
del self.chunks[self._last[0]:]
812
self.endpoint = self._last[1]
816
"""Return the overall compression ratio."""
817
return float(self.input_bytes) / float(self.endpoint)
820
class PythonGroupCompressor(_CommonGroupCompressor):
823
"""Create a GroupCompressor.
825
Used only if the pyrex version is not available.
827
super(PythonGroupCompressor, self).__init__()
828
self._delta_index = LinesDeltaIndex([])
829
# The actual content is managed by LinesDeltaIndex
830
self.chunks = self._delta_index.lines
832
def _compress(self, key, bytes, max_delta_size, soft=False):
833
"""see _CommonGroupCompressor._compress"""
834
input_len = len(bytes)
835
new_lines = osutils.split_lines(bytes)
836
out_lines, index_lines = self._delta_index.make_delta(
837
new_lines, bytes_length=input_len, soft=soft)
838
delta_length = sum(map(len, out_lines))
839
if delta_length > max_delta_size:
840
# The delta is longer than the fulltext, insert a fulltext
842
out_lines = ['f', encode_base128_int(input_len)]
843
out_lines.extend(new_lines)
844
index_lines = [False, False]
845
index_lines.extend([True] * len(new_lines))
847
# this is a worthy delta, output it
850
# Update the delta_length to include those two encoded integers
851
out_lines[1] = encode_base128_int(delta_length)
853
start = self.endpoint
854
chunk_start = len(self.chunks)
855
self._last = (chunk_start, self.endpoint)
856
self._delta_index.extend_lines(out_lines, index_lines)
857
self.endpoint = self._delta_index.endpoint
858
self.input_bytes += input_len
859
chunk_end = len(self.chunks)
860
self.labels_deltas[key] = (start, chunk_start,
861
self.endpoint, chunk_end)
862
return start, self.endpoint, type
865
class PyrexGroupCompressor(_CommonGroupCompressor):
866
"""Produce a serialised group of compressed texts.
868
It contains code very similar to SequenceMatcher because of having a similar
869
task. However some key differences apply:
870
- there is no junk, we want a minimal edit not a human readable diff.
871
- we don't filter very common lines (because we don't know where a good
872
range will start, and after the first text we want to be emitting minmal
874
- we chain the left side, not the right side
875
- we incrementally update the adjacency matrix as new lines are provided.
876
- we look for matches in all of the left side, so the routine which does
877
the analagous task of find_longest_match does not need to filter on the
882
super(PyrexGroupCompressor, self).__init__()
883
self._delta_index = DeltaIndex()
885
def _compress(self, key, bytes, max_delta_size, soft=False):
886
"""see _CommonGroupCompressor._compress"""
887
input_len = len(bytes)
888
# By having action/label/sha1/len, we can parse the group if the index
889
# was ever destroyed, we have the key in 'label', we know the final
890
# bytes are valid from sha1, and we know where to find the end of this
891
# record because of 'len'. (the delta record itself will store the
892
# total length for the expanded record)
893
# 'len: %d\n' costs approximately 1% increase in total data
894
# Having the labels at all costs us 9-10% increase, 38% increase for
895
# inventory pages, and 5.8% increase for text pages
896
# new_chunks = ['label:%s\nsha1:%s\n' % (label, sha1)]
897
if self._delta_index._source_offset != self.endpoint:
898
raise AssertionError('_source_offset != endpoint'
899
' somehow the DeltaIndex got out of sync with'
901
delta = self._delta_index.make_delta(bytes, max_delta_size)
904
enc_length = encode_base128_int(len(bytes))
905
len_mini_header = 1 + len(enc_length)
906
self._delta_index.add_source(bytes, len_mini_header)
907
new_chunks = ['f', enc_length, bytes]
910
enc_length = encode_base128_int(len(delta))
911
len_mini_header = 1 + len(enc_length)
912
new_chunks = ['d', enc_length, delta]
913
self._delta_index.add_delta_source(delta, len_mini_header)
915
start = self.endpoint
916
chunk_start = len(self.chunks)
917
# Now output these bytes
918
self._output_chunks(new_chunks)
919
self.input_bytes += input_len
920
chunk_end = len(self.chunks)
921
self.labels_deltas[key] = (start, chunk_start,
922
self.endpoint, chunk_end)
923
if not self._delta_index._source_offset == self.endpoint:
924
raise AssertionError('the delta index is out of sync'
925
'with the output lines %s != %s'
926
% (self._delta_index._source_offset, self.endpoint))
927
return start, self.endpoint, type
929
def _output_chunks(self, new_chunks):
930
"""Output some chunks.
932
:param new_chunks: The chunks to output.
934
self._last = (len(self.chunks), self.endpoint)
935
endpoint = self.endpoint
936
self.chunks.extend(new_chunks)
937
endpoint += sum(map(len, new_chunks))
938
self.endpoint = endpoint
941
def make_pack_factory(graph, delta, keylength, inconsistency_fatal=True):
942
"""Create a factory for creating a pack based groupcompress.
944
This is only functional enough to run interface tests, it doesn't try to
945
provide a full pack environment.
947
:param graph: Store a graph.
948
:param delta: Delta compress contents.
949
:param keylength: How long should keys be.
951
def factory(transport):
956
graph_index = BTreeBuilder(reference_lists=ref_length,
957
key_elements=keylength)
958
stream = transport.open_write_stream('newpack')
959
writer = pack.ContainerWriter(stream.write)
961
index = _GCGraphIndex(graph_index, lambda:True, parents=parents,
962
add_callback=graph_index.add_nodes,
963
inconsistency_fatal=inconsistency_fatal)
964
access = knit._DirectPackAccess({})
965
access.set_writer(writer, graph_index, (transport, 'newpack'))
966
result = GroupCompressVersionedFiles(index, access, delta)
967
result.stream = stream
968
result.writer = writer
973
def cleanup_pack_group(versioned_files):
974
versioned_files.writer.end()
975
versioned_files.stream.close()
978
class GroupCompressVersionedFiles(VersionedFiles):
979
"""A group-compress based VersionedFiles implementation."""
981
def __init__(self, index, access, delta=True):
982
"""Create a GroupCompressVersionedFiles object.
984
:param index: The index object storing access and graph data.
985
:param access: The access object storing raw data.
986
:param delta: Whether to delta compress or just entropy compress.
989
self._access = access
991
self._unadded_refs = {}
992
self._group_cache = LRUSizeCache(max_size=50*1024*1024)
993
self._fallback_vfs = []
995
def add_lines(self, key, parents, lines, parent_texts=None,
996
left_matching_blocks=None, nostore_sha=None, random_id=False,
998
"""Add a text to the store.
1000
:param key: The key tuple of the text to add.
1001
:param parents: The parents key tuples of the text to add.
1002
:param lines: A list of lines. Each line must be a bytestring. And all
1003
of them except the last must be terminated with \n and contain no
1004
other \n's. The last line may either contain no \n's or a single
1005
terminating \n. If the lines list does meet this constraint the add
1006
routine may error or may succeed - but you will be unable to read
1007
the data back accurately. (Checking the lines have been split
1008
correctly is expensive and extremely unlikely to catch bugs so it
1009
is not done at runtime unless check_content is True.)
1010
:param parent_texts: An optional dictionary containing the opaque
1011
representations of some or all of the parents of version_id to
1012
allow delta optimisations. VERY IMPORTANT: the texts must be those
1013
returned by add_lines or data corruption can be caused.
1014
:param left_matching_blocks: a hint about which areas are common
1015
between the text and its left-hand-parent. The format is
1016
the SequenceMatcher.get_matching_blocks format.
1017
:param nostore_sha: Raise ExistingContent and do not add the lines to
1018
the versioned file if the digest of the lines matches this.
1019
:param random_id: If True a random id has been selected rather than
1020
an id determined by some deterministic process such as a converter
1021
from a foreign VCS. When True the backend may choose not to check
1022
for uniqueness of the resulting key within the versioned file, so
1023
this should only be done when the result is expected to be unique
1025
:param check_content: If True, the lines supplied are verified to be
1026
bytestrings that are correctly formed lines.
1027
:return: The text sha1, the number of bytes in the text, and an opaque
1028
representation of the inserted version which can be provided
1029
back to future add_lines calls in the parent_texts dictionary.
1031
self._index._check_write_ok()
1032
self._check_add(key, lines, random_id, check_content)
1034
# The caller might pass None if there is no graph data, but kndx
1035
# indexes can't directly store that, so we give them
1036
# an empty tuple instead.
1038
# double handling for now. Make it work until then.
1039
length = sum(map(len, lines))
1040
record = ChunkedContentFactory(key, parents, None, lines)
1041
sha1 = list(self._insert_record_stream([record], random_id=random_id,
1042
nostore_sha=nostore_sha))[0]
1043
return sha1, length, None
1045
def _add_text(self, key, parents, text, nostore_sha=None, random_id=False):
1046
"""See VersionedFiles._add_text()."""
1047
self._index._check_write_ok()
1048
self._check_add(key, None, random_id, check_content=False)
1049
if text.__class__ is not str:
1050
raise errors.BzrBadParameterUnicode("text")
1052
# The caller might pass None if there is no graph data, but kndx
1053
# indexes can't directly store that, so we give them
1054
# an empty tuple instead.
1056
# double handling for now. Make it work until then.
1058
record = FulltextContentFactory(key, parents, None, text)
1059
sha1 = list(self._insert_record_stream([record], random_id=random_id,
1060
nostore_sha=nostore_sha))[0]
1061
return sha1, length, None
1063
def add_fallback_versioned_files(self, a_versioned_files):
1064
"""Add a source of texts for texts not present in this knit.
1066
:param a_versioned_files: A VersionedFiles object.
1068
self._fallback_vfs.append(a_versioned_files)
1070
def annotate(self, key):
1071
"""See VersionedFiles.annotate."""
1072
ann = annotate.Annotator(self)
1073
return ann.annotate_flat(key)
1075
def get_annotator(self):
1076
return annotate.Annotator(self)
1078
def check(self, progress_bar=None, keys=None):
1079
"""See VersionedFiles.check()."""
1082
for record in self.get_record_stream(keys, 'unordered', True):
1083
record.get_bytes_as('fulltext')
1085
return self.get_record_stream(keys, 'unordered', True)
1087
def _check_add(self, key, lines, random_id, check_content):
1088
"""check that version_id and lines are safe to add."""
1089
version_id = key[-1]
1090
if version_id is not None:
1091
if osutils.contains_whitespace(version_id):
1092
raise errors.InvalidRevisionId(version_id, self)
1093
self.check_not_reserved_id(version_id)
1094
# TODO: If random_id==False and the key is already present, we should
1095
# probably check that the existing content is identical to what is
1096
# being inserted, and otherwise raise an exception. This would make
1097
# the bundle code simpler.
1099
self._check_lines_not_unicode(lines)
1100
self._check_lines_are_lines(lines)
1102
def get_parent_map(self, keys):
1103
"""Get a map of the graph parents of keys.
1105
:param keys: The keys to look up parents for.
1106
:return: A mapping from keys to parents. Absent keys are absent from
1109
return self._get_parent_map_with_sources(keys)[0]
1111
def _get_parent_map_with_sources(self, keys):
1112
"""Get a map of the parents of keys.
1114
:param keys: The keys to look up parents for.
1115
:return: A tuple. The first element is a mapping from keys to parents.
1116
Absent keys are absent from the mapping. The second element is a
1117
list with the locations each key was found in. The first element
1118
is the in-this-knit parents, the second the first fallback source,
1122
sources = [self._index] + self._fallback_vfs
1125
for source in sources:
1128
new_result = source.get_parent_map(missing)
1129
source_results.append(new_result)
1130
result.update(new_result)
1131
missing.difference_update(set(new_result))
1132
return result, source_results
1134
def _get_block(self, index_memo):
1135
read_memo = index_memo[0:3]
1138
block = self._group_cache[read_memo]
1141
zdata = self._access.get_raw_records([read_memo]).next()
1142
# decompress - whole thing - this is not a bug, as it
1143
# permits caching. We might want to store the partially
1144
# decompresed group and decompress object, so that recent
1145
# texts are not penalised by big groups.
1146
block = GroupCompressBlock.from_bytes(zdata)
1147
self._group_cache[read_memo] = block
1149
# print len(zdata), len(plain)
1150
# parse - requires split_lines, better to have byte offsets
1151
# here (but not by much - we only split the region for the
1152
# recipe, and we often want to end up with lines anyway.
1155
def get_missing_compression_parent_keys(self):
1156
"""Return the keys of missing compression parents.
1158
Missing compression parents occur when a record stream was missing
1159
basis texts, or a index was scanned that had missing basis texts.
1161
# GroupCompress cannot currently reference texts that are not in the
1162
# group, so this is valid for now
1165
def get_record_stream(self, keys, ordering, include_delta_closure):
1166
"""Get a stream of records for keys.
1168
:param keys: The keys to include.
1169
:param ordering: Either 'unordered' or 'topological'. A topologically
1170
sorted stream has compression parents strictly before their
1172
:param include_delta_closure: If True then the closure across any
1173
compression parents will be included (in the opaque data).
1174
:return: An iterator of ContentFactory objects, each of which is only
1175
valid until the iterator is advanced.
1177
# keys might be a generator
1178
orig_keys = list(keys)
1182
if (not self._index.has_graph
1183
and ordering in ('topological', 'groupcompress')):
1184
# Cannot topological order when no graph has been stored.
1185
# but we allow 'as-requested' or 'unordered'
1186
ordering = 'unordered'
1188
remaining_keys = keys
1191
keys = set(remaining_keys)
1192
for content_factory in self._get_remaining_record_stream(keys,
1193
orig_keys, ordering, include_delta_closure):
1194
remaining_keys.discard(content_factory.key)
1195
yield content_factory
1197
except errors.RetryWithNewPacks, e:
1198
self._access.reload_or_raise(e)
1200
def _find_from_fallback(self, missing):
1201
"""Find whatever keys you can from the fallbacks.
1203
:param missing: A set of missing keys. This set will be mutated as keys
1204
are found from a fallback_vfs
1205
:return: (parent_map, key_to_source_map, source_results)
1206
parent_map the overall key => parent_keys
1207
key_to_source_map a dict from {key: source}
1208
source_results a list of (source: keys)
1211
key_to_source_map = {}
1213
for source in self._fallback_vfs:
1216
source_parents = source.get_parent_map(missing)
1217
parent_map.update(source_parents)
1218
source_parents = list(source_parents)
1219
source_results.append((source, source_parents))
1220
key_to_source_map.update((key, source) for key in source_parents)
1221
missing.difference_update(source_parents)
1222
return parent_map, key_to_source_map, source_results
1224
def _get_ordered_source_keys(self, ordering, parent_map, key_to_source_map):
1225
"""Get the (source, [keys]) list.
1227
The returned objects should be in the order defined by 'ordering',
1228
which can weave between different sources.
1229
:param ordering: Must be one of 'topological' or 'groupcompress'
1230
:return: List of [(source, [keys])] tuples, such that all keys are in
1231
the defined order, regardless of source.
1233
if ordering == 'topological':
1234
present_keys = topo_sort(parent_map)
1236
# ordering == 'groupcompress'
1237
# XXX: This only optimizes for the target ordering. We may need
1238
# to balance that with the time it takes to extract
1239
# ordering, by somehow grouping based on
1240
# locations[key][0:3]
1241
present_keys = sort_gc_optimal(parent_map)
1242
# Now group by source:
1244
current_source = None
1245
for key in present_keys:
1246
source = key_to_source_map.get(key, self)
1247
if source is not current_source:
1248
source_keys.append((source, []))
1249
current_source = source
1250
source_keys[-1][1].append(key)
1253
def _get_as_requested_source_keys(self, orig_keys, locations, unadded_keys,
1256
current_source = None
1257
for key in orig_keys:
1258
if key in locations or key in unadded_keys:
1260
elif key in key_to_source_map:
1261
source = key_to_source_map[key]
1264
if source is not current_source:
1265
source_keys.append((source, []))
1266
current_source = source
1267
source_keys[-1][1].append(key)
1270
def _get_io_ordered_source_keys(self, locations, unadded_keys,
1273
# This is the group the bytes are stored in, followed by the
1274
# location in the group
1275
return locations[key][0]
1276
present_keys = sorted(locations.iterkeys(), key=get_group)
1277
# We don't have an ordering for keys in the in-memory object, but
1278
# lets process the in-memory ones first.
1279
present_keys = list(unadded_keys) + present_keys
1280
# Now grab all of the ones from other sources
1281
source_keys = [(self, present_keys)]
1282
source_keys.extend(source_result)
1285
def _get_remaining_record_stream(self, keys, orig_keys, ordering,
1286
include_delta_closure):
1287
"""Get a stream of records for keys.
1289
:param keys: The keys to include.
1290
:param ordering: one of 'unordered', 'topological', 'groupcompress' or
1292
:param include_delta_closure: If True then the closure across any
1293
compression parents will be included (in the opaque data).
1294
:return: An iterator of ContentFactory objects, each of which is only
1295
valid until the iterator is advanced.
1298
locations = self._index.get_build_details(keys)
1299
unadded_keys = set(self._unadded_refs).intersection(keys)
1300
missing = keys.difference(locations)
1301
missing.difference_update(unadded_keys)
1302
(fallback_parent_map, key_to_source_map,
1303
source_result) = self._find_from_fallback(missing)
1304
if ordering in ('topological', 'groupcompress'):
1305
# would be better to not globally sort initially but instead
1306
# start with one key, recurse to its oldest parent, then grab
1307
# everything in the same group, etc.
1308
parent_map = dict((key, details[2]) for key, details in
1309
locations.iteritems())
1310
for key in unadded_keys:
1311
parent_map[key] = self._unadded_refs[key]
1312
parent_map.update(fallback_parent_map)
1313
source_keys = self._get_ordered_source_keys(ordering, parent_map,
1315
elif ordering == 'as-requested':
1316
source_keys = self._get_as_requested_source_keys(orig_keys,
1317
locations, unadded_keys, key_to_source_map)
1319
# We want to yield the keys in a semi-optimal (read-wise) ordering.
1320
# Otherwise we thrash the _group_cache and destroy performance
1321
source_keys = self._get_io_ordered_source_keys(locations,
1322
unadded_keys, source_result)
1324
yield AbsentContentFactory(key)
1326
last_read_memo = None
1327
# TODO: This works fairly well at batching up existing groups into a
1328
# streamable format, and possibly allowing for taking one big
1329
# group and splitting it when it isn't fully utilized.
1330
# However, it doesn't allow us to find under-utilized groups and
1331
# combine them into a bigger group on the fly.
1332
# (Consider the issue with how chk_map inserts texts
1333
# one-at-a-time.) This could be done at insert_record_stream()
1334
# time, but it probably would decrease the number of
1335
# bytes-on-the-wire for fetch.
1336
for source, keys in source_keys:
1339
if key in self._unadded_refs:
1340
if manager is not None:
1341
for factory in manager.get_record_stream():
1343
last_read_memo = manager = None
1344
bytes, sha1 = self._compressor.extract(key)
1345
parents = self._unadded_refs[key]
1346
yield FulltextContentFactory(key, parents, sha1, bytes)
1348
index_memo, _, parents, (method, _) = locations[key]
1349
read_memo = index_memo[0:3]
1350
if last_read_memo != read_memo:
1351
# We are starting a new block. If we have a
1352
# manager, we have found everything that fits for
1353
# now, so yield records
1354
if manager is not None:
1355
for factory in manager.get_record_stream():
1357
# Now start a new manager
1358
block = self._get_block(index_memo)
1359
manager = _LazyGroupContentManager(block)
1360
last_read_memo = read_memo
1361
start, end = index_memo[3:5]
1362
manager.add_factory(key, parents, start, end)
1364
if manager is not None:
1365
for factory in manager.get_record_stream():
1367
last_read_memo = manager = None
1368
for record in source.get_record_stream(keys, ordering,
1369
include_delta_closure):
1371
if manager is not None:
1372
for factory in manager.get_record_stream():
1375
def get_sha1s(self, keys):
1376
"""See VersionedFiles.get_sha1s()."""
1378
for record in self.get_record_stream(keys, 'unordered', True):
1379
if record.sha1 != None:
1380
result[record.key] = record.sha1
1382
if record.storage_kind != 'absent':
1383
result[record.key] = osutils.sha_string(
1384
record.get_bytes_as('fulltext'))
1387
def insert_record_stream(self, stream):
1388
"""Insert a record stream into this container.
1390
:param stream: A stream of records to insert.
1392
:seealso VersionedFiles.get_record_stream:
1394
# XXX: Setting random_id=True makes
1395
# test_insert_record_stream_existing_keys fail for groupcompress and
1396
# groupcompress-nograph, this needs to be revisited while addressing
1397
# 'bzr branch' performance issues.
1398
for _ in self._insert_record_stream(stream, random_id=False):
1401
def _insert_record_stream(self, stream, random_id=False, nostore_sha=None,
1403
"""Internal core to insert a record stream into this container.
1405
This helper function has a different interface than insert_record_stream
1406
to allow add_lines to be minimal, but still return the needed data.
1408
:param stream: A stream of records to insert.
1409
:param nostore_sha: If the sha1 of a given text matches nostore_sha,
1410
raise ExistingContent, rather than committing the new text.
1411
:param reuse_blocks: If the source is streaming from
1412
groupcompress-blocks, just insert the blocks as-is, rather than
1413
expanding the texts and inserting again.
1414
:return: An iterator over the sha1 of the inserted records.
1415
:seealso insert_record_stream:
1419
def get_adapter(adapter_key):
1421
return adapters[adapter_key]
1423
adapter_factory = adapter_registry.get(adapter_key)
1424
adapter = adapter_factory(self)
1425
adapters[adapter_key] = adapter
1427
# This will go up to fulltexts for gc to gc fetching, which isn't
1429
self._compressor = GroupCompressor()
1430
self._unadded_refs = {}
1433
bytes = self._compressor.flush().to_bytes()
1434
index, start, length = self._access.add_raw_records(
1435
[(None, len(bytes))], bytes)[0]
1437
for key, reads, refs in keys_to_add:
1438
nodes.append((key, "%d %d %s" % (start, length, reads), refs))
1439
self._index.add_records(nodes, random_id=random_id)
1440
self._unadded_refs = {}
1442
self._compressor = GroupCompressor()
1445
max_fulltext_len = 0
1446
max_fulltext_prefix = None
1447
insert_manager = None
1450
# XXX: TODO: remove this, it is just for safety checking for now
1451
inserted_keys = set()
1452
for record in stream:
1453
# Raise an error when a record is missing.
1454
if record.storage_kind == 'absent':
1455
raise errors.RevisionNotPresent(record.key, self)
1457
if record.key in inserted_keys:
1458
trace.note('Insert claimed random_id=True,'
1459
' but then inserted %r two times', record.key)
1461
inserted_keys.add(record.key)
1463
# If the reuse_blocks flag is set, check to see if we can just
1464
# copy a groupcompress block as-is.
1465
if record.storage_kind == 'groupcompress-block':
1466
# Insert the raw block into the target repo
1467
insert_manager = record._manager
1468
insert_manager._check_rebuild_block()
1469
bytes = record._manager._block.to_bytes()
1470
_, start, length = self._access.add_raw_records(
1471
[(None, len(bytes))], bytes)[0]
1474
block_length = length
1475
if record.storage_kind in ('groupcompress-block',
1476
'groupcompress-block-ref'):
1477
if insert_manager is None:
1478
raise AssertionError('No insert_manager set')
1479
value = "%d %d %d %d" % (block_start, block_length,
1480
record._start, record._end)
1481
nodes = [(record.key, value, (record.parents,))]
1482
# TODO: Consider buffering up many nodes to be added, not
1483
# sure how much overhead this has, but we're seeing
1484
# ~23s / 120s in add_records calls
1485
self._index.add_records(nodes, random_id=random_id)
1488
bytes = record.get_bytes_as('fulltext')
1489
except errors.UnavailableRepresentation:
1490
adapter_key = record.storage_kind, 'fulltext'
1491
adapter = get_adapter(adapter_key)
1492
bytes = adapter.get_bytes(record)
1493
if len(record.key) > 1:
1494
prefix = record.key[0]
1495
soft = (prefix == last_prefix)
1499
if max_fulltext_len < len(bytes):
1500
max_fulltext_len = len(bytes)
1501
max_fulltext_prefix = prefix
1502
(found_sha1, start_point, end_point,
1503
type) = self._compressor.compress(record.key,
1504
bytes, record.sha1, soft=soft,
1505
nostore_sha=nostore_sha)
1506
# delta_ratio = float(len(bytes)) / (end_point - start_point)
1507
# Check if we want to continue to include that text
1508
if (prefix == max_fulltext_prefix
1509
and end_point < 2 * max_fulltext_len):
1510
# As long as we are on the same file_id, we will fill at least
1511
# 2 * max_fulltext_len
1512
start_new_block = False
1513
elif end_point > 4*1024*1024:
1514
start_new_block = True
1515
elif (prefix is not None and prefix != last_prefix
1516
and end_point > 2*1024*1024):
1517
start_new_block = True
1519
start_new_block = False
1520
last_prefix = prefix
1522
self._compressor.pop_last()
1524
max_fulltext_len = len(bytes)
1525
(found_sha1, start_point, end_point,
1526
type) = self._compressor.compress(record.key, bytes,
1528
if record.key[-1] is None:
1529
key = record.key[:-1] + ('sha1:' + found_sha1,)
1532
self._unadded_refs[key] = record.parents
1534
keys_to_add.append((key, '%d %d' % (start_point, end_point),
1536
if len(keys_to_add):
1538
self._compressor = None
1540
def iter_lines_added_or_present_in_keys(self, keys, pb=None):
1541
"""Iterate over the lines in the versioned files from keys.
1543
This may return lines from other keys. Each item the returned
1544
iterator yields is a tuple of a line and a text version that that line
1545
is present in (not introduced in).
1547
Ordering of results is in whatever order is most suitable for the
1548
underlying storage format.
1550
If a progress bar is supplied, it may be used to indicate progress.
1551
The caller is responsible for cleaning up progress bars (because this
1555
* Lines are normalised by the underlying store: they will all have \n
1557
* Lines are returned in arbitrary order.
1559
:return: An iterator over (line, key).
1563
# we don't care about inclusions, the caller cares.
1564
# but we need to setup a list of records to visit.
1565
# we need key, position, length
1566
for key_idx, record in enumerate(self.get_record_stream(keys,
1567
'unordered', True)):
1568
# XXX: todo - optimise to use less than full texts.
1571
pb.update('Walking content', key_idx, total)
1572
if record.storage_kind == 'absent':
1573
raise errors.RevisionNotPresent(key, self)
1574
lines = osutils.split_lines(record.get_bytes_as('fulltext'))
1578
pb.update('Walking content', total, total)
1581
"""See VersionedFiles.keys."""
1582
if 'evil' in debug.debug_flags:
1583
trace.mutter_callsite(2, "keys scales with size of history")
1584
sources = [self._index] + self._fallback_vfs
1586
for source in sources:
1587
result.update(source.keys())
1591
class _GCGraphIndex(object):
1592
"""Mapper from GroupCompressVersionedFiles needs into GraphIndex storage."""
1594
def __init__(self, graph_index, is_locked, parents=True,
1595
add_callback=None, track_external_parent_refs=False,
1596
inconsistency_fatal=True):
1597
"""Construct a _GCGraphIndex on a graph_index.
1599
:param graph_index: An implementation of bzrlib.index.GraphIndex.
1600
:param is_locked: A callback, returns True if the index is locked and
1602
:param parents: If True, record knits parents, if not do not record
1604
:param add_callback: If not None, allow additions to the index and call
1605
this callback with a list of added GraphIndex nodes:
1606
[(node, value, node_refs), ...]
1607
:param track_external_parent_refs: As keys are added, keep track of the
1608
keys they reference, so that we can query get_missing_parents(),
1610
:param inconsistency_fatal: When asked to add records that are already
1611
present, and the details are inconsistent with the existing
1612
record, raise an exception instead of warning (and skipping the
1615
self._add_callback = add_callback
1616
self._graph_index = graph_index
1617
self._parents = parents
1618
self.has_graph = parents
1619
self._is_locked = is_locked
1620
self._inconsistency_fatal = inconsistency_fatal
1621
if track_external_parent_refs:
1622
self._key_dependencies = knit._KeyRefs()
1624
self._key_dependencies = None
1626
def add_records(self, records, random_id=False):
1627
"""Add multiple records to the index.
1629
This function does not insert data into the Immutable GraphIndex
1630
backing the KnitGraphIndex, instead it prepares data for insertion by
1631
the caller and checks that it is safe to insert then calls
1632
self._add_callback with the prepared GraphIndex nodes.
1634
:param records: a list of tuples:
1635
(key, options, access_memo, parents).
1636
:param random_id: If True the ids being added were randomly generated
1637
and no check for existence will be performed.
1639
if not self._add_callback:
1640
raise errors.ReadOnlyError(self)
1641
# we hope there are no repositories with inconsistent parentage
1646
for (key, value, refs) in records:
1647
if not self._parents:
1651
raise errors.KnitCorrupt(self,
1652
"attempt to add node with parents "
1653
"in parentless index.")
1656
keys[key] = (value, refs)
1659
present_nodes = self._get_entries(keys)
1660
for (index, key, value, node_refs) in present_nodes:
1661
if node_refs != keys[key][1]:
1662
details = '%s %s %s' % (key, (value, node_refs), keys[key])
1663
if self._inconsistency_fatal:
1664
raise errors.KnitCorrupt(self, "inconsistent details"
1665
" in add_records: %s" %
1668
trace.warning("inconsistent details in skipped"
1669
" record: %s", details)
1675
for key, (value, node_refs) in keys.iteritems():
1676
result.append((key, value, node_refs))
1678
for key, (value, node_refs) in keys.iteritems():
1679
result.append((key, value))
1681
key_dependencies = self._key_dependencies
1682
if key_dependencies is not None and self._parents:
1683
for key, value, refs in records:
1685
key_dependencies.add_references(key, parents)
1686
self._add_callback(records)
1688
def _check_read(self):
1689
"""Raise an exception if reads are not permitted."""
1690
if not self._is_locked():
1691
raise errors.ObjectNotLocked(self)
1693
def _check_write_ok(self):
1694
"""Raise an exception if writes are not permitted."""
1695
if not self._is_locked():
1696
raise errors.ObjectNotLocked(self)
1698
def _get_entries(self, keys, check_present=False):
1699
"""Get the entries for keys.
1701
Note: Callers are responsible for checking that the index is locked
1702
before calling this method.
1704
:param keys: An iterable of index key tuples.
1709
for node in self._graph_index.iter_entries(keys):
1711
found_keys.add(node[1])
1713
# adapt parentless index to the rest of the code.
1714
for node in self._graph_index.iter_entries(keys):
1715
yield node[0], node[1], node[2], ()
1716
found_keys.add(node[1])
1718
missing_keys = keys.difference(found_keys)
1720
raise errors.RevisionNotPresent(missing_keys.pop(), self)
1722
def get_parent_map(self, keys):
1723
"""Get a map of the parents of keys.
1725
:param keys: The keys to look up parents for.
1726
:return: A mapping from keys to parents. Absent keys are absent from
1730
nodes = self._get_entries(keys)
1734
result[node[1]] = node[3][0]
1737
result[node[1]] = None
1740
def get_missing_parents(self):
1741
"""Return the keys of missing parents."""
1742
# Copied from _KnitGraphIndex.get_missing_parents
1743
# We may have false positives, so filter those out.
1744
self._key_dependencies.add_keys(
1745
self.get_parent_map(self._key_dependencies.get_unsatisfied_refs()))
1746
return frozenset(self._key_dependencies.get_unsatisfied_refs())
1748
def get_build_details(self, keys):
1749
"""Get the various build details for keys.
1751
Ghosts are omitted from the result.
1753
:param keys: An iterable of keys.
1754
:return: A dict of key:
1755
(index_memo, compression_parent, parents, record_details).
1757
opaque structure to pass to read_records to extract the raw
1760
Content that this record is built upon, may be None
1762
Logical parents of this node
1764
extra information about the content which needs to be passed to
1765
Factory.parse_record
1769
entries = self._get_entries(keys)
1770
for entry in entries:
1772
if not self._parents:
1775
parents = entry[3][0]
1777
result[key] = (self._node_to_position(entry),
1778
None, parents, (method, None))
1782
"""Get all the keys in the collection.
1784
The keys are not ordered.
1787
return [node[1] for node in self._graph_index.iter_all_entries()]
1789
def _node_to_position(self, node):
1790
"""Convert an index value to position details."""
1791
bits = node[2].split(' ')
1792
# It would be nice not to read the entire gzip.
1793
start = int(bits[0])
1795
basis_end = int(bits[2])
1796
delta_end = int(bits[3])
1797
return node[0], start, stop, basis_end, delta_end
1799
def scan_unvalidated_index(self, graph_index):
1800
"""Inform this _GCGraphIndex that there is an unvalidated index.
1802
This allows this _GCGraphIndex to keep track of any missing
1803
compression parents we may want to have filled in to make those
1806
:param graph_index: A GraphIndex
1808
if self._key_dependencies is not None:
1809
# Add parent refs from graph_index (and discard parent refs that
1810
# the graph_index has).
1811
add_refs = self._key_dependencies.add_references
1812
for node in graph_index.iter_all_entries():
1813
add_refs(node[1], node[3][0])
1817
from bzrlib._groupcompress_py import (
1819
apply_delta_to_source,
1822
decode_copy_instruction,
1826
from bzrlib._groupcompress_pyx import (
1828
apply_delta_to_source,
1833
GroupCompressor = PyrexGroupCompressor
1835
GroupCompressor = PythonGroupCompressor
added
removed
bzrlib/groupcompress.py
