1
# Copyright (C) 2008-2011 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."""
19
from __future__ import absolute_import
24
from ..lazy_import import lazy_import
25
lazy_import(globals(), """
35
from breezy.bzr import (
41
from breezy.i18n import gettext
47
from .btree_index import BTreeBuilder
48
from ..lru_cache import LRUSizeCache
49
from ..sixish import (
55
from .versionedfile import (
59
ChunkedContentFactory,
60
FulltextContentFactory,
61
VersionedFilesWithFallbacks,
64
# Minimum number of uncompressed bytes to try fetch at once when retrieving
65
# groupcompress blocks.
68
# osutils.sha_string(b'')
69
_null_sha1 = b'da39a3ee5e6b4b0d3255bfef95601890afd80709'
72
def sort_gc_optimal(parent_map):
73
"""Sort and group the keys in parent_map into groupcompress order.
75
groupcompress is defined (currently) as reverse-topological order, grouped
78
:return: A sorted-list of keys
80
# groupcompress ordering is approximately reverse topological,
81
# properly grouped by file-id.
83
for key, value in viewitems(parent_map):
84
if isinstance(key, bytes) or len(key) == 1:
89
per_prefix_map[prefix][key] = value
91
per_prefix_map[prefix] = {key: value}
94
for prefix in sorted(per_prefix_map):
95
present_keys.extend(reversed(tsort.topo_sort(per_prefix_map[prefix])))
99
class DecompressCorruption(errors.BzrError):
101
_fmt = "Corruption while decompressing repository file%(orig_error)s"
103
def __init__(self, orig_error=None):
104
if orig_error is not None:
105
self.orig_error = ", %s" % (orig_error,)
108
errors.BzrError.__init__(self)
111
# The max zlib window size is 32kB, so if we set 'max_size' output of the
112
# decompressor to the requested bytes + 32kB, then we should guarantee
113
# num_bytes coming out.
114
_ZLIB_DECOMP_WINDOW = 32 * 1024
117
class GroupCompressBlock(object):
118
"""An object which maintains the internal structure of the compressed data.
120
This tracks the meta info (start of text, length, type, etc.)
123
# Group Compress Block v1 Zlib
124
GCB_HEADER = b'gcb1z\n'
125
# Group Compress Block v1 Lzma
126
GCB_LZ_HEADER = b'gcb1l\n'
127
GCB_KNOWN_HEADERS = (GCB_HEADER, GCB_LZ_HEADER)
130
# map by key? or just order in file?
131
self._compressor_name = None
132
self._z_content_chunks = None
133
self._z_content_decompressor = None
134
self._z_content_length = None
135
self._content_length = None
137
self._content_chunks = None
140
# This is the maximum number of bytes this object will reference if
141
# everything is decompressed. However, if we decompress less than
142
# everything... (this would cause some problems for LRUSizeCache)
143
return self._content_length + self._z_content_length
145
def _ensure_content(self, num_bytes=None):
146
"""Make sure that content has been expanded enough.
148
:param num_bytes: Ensure that we have extracted at least num_bytes of
149
content. If None, consume everything
151
if self._content_length is None:
152
raise AssertionError('self._content_length should never be None')
153
if num_bytes is None:
154
num_bytes = self._content_length
155
elif (self._content_length is not None
156
and num_bytes > self._content_length):
157
raise AssertionError(
158
'requested num_bytes (%d) > content length (%d)'
159
% (num_bytes, self._content_length))
160
# Expand the content if required
161
if self._content is None:
162
if self._content_chunks is not None:
163
self._content = b''.join(self._content_chunks)
164
self._content_chunks = None
165
if self._content is None:
166
# We join self._z_content_chunks here, because if we are
167
# decompressing, then it is *very* likely that we have a single
169
if self._z_content_chunks is None:
170
raise AssertionError('No content to decompress')
171
z_content = b''.join(self._z_content_chunks)
174
elif self._compressor_name == 'lzma':
175
# We don't do partial lzma decomp yet
177
self._content = pylzma.decompress(z_content)
178
elif self._compressor_name == 'zlib':
179
# Start a zlib decompressor
180
if num_bytes * 4 > self._content_length * 3:
181
# If we are requesting more that 3/4ths of the content,
182
# just extract the whole thing in a single pass
183
num_bytes = self._content_length
184
self._content = zlib.decompress(z_content)
186
self._z_content_decompressor = zlib.decompressobj()
187
# Seed the decompressor with the uncompressed bytes, so
188
# that the rest of the code is simplified
189
self._content = self._z_content_decompressor.decompress(
190
z_content, num_bytes + _ZLIB_DECOMP_WINDOW)
191
if not self._z_content_decompressor.unconsumed_tail:
192
self._z_content_decompressor = None
194
raise AssertionError('Unknown compressor: %r'
195
% self._compressor_name)
196
# Any bytes remaining to be decompressed will be in the decompressors
199
# Do we have enough bytes already?
200
if len(self._content) >= num_bytes:
202
# If we got this far, and don't have a decompressor, something is wrong
203
if self._z_content_decompressor is None:
204
raise AssertionError(
205
'No decompressor to decompress %d bytes' % num_bytes)
206
remaining_decomp = self._z_content_decompressor.unconsumed_tail
207
if not remaining_decomp:
208
raise AssertionError('Nothing left to decompress')
209
needed_bytes = num_bytes - len(self._content)
210
# We always set max_size to 32kB over the minimum needed, so that
211
# zlib will give us as much as we really want.
212
# TODO: If this isn't good enough, we could make a loop here,
213
# that keeps expanding the request until we get enough
214
self._content += self._z_content_decompressor.decompress(
215
remaining_decomp, needed_bytes + _ZLIB_DECOMP_WINDOW)
216
if len(self._content) < num_bytes:
217
raise AssertionError('%d bytes wanted, only %d available'
218
% (num_bytes, len(self._content)))
219
if not self._z_content_decompressor.unconsumed_tail:
220
# The stream is finished
221
self._z_content_decompressor = None
223
def _parse_bytes(self, data, pos):
224
"""Read the various lengths from the header.
226
This also populates the various 'compressed' buffers.
228
:return: The position in bytes just after the last newline
230
# At present, we have 2 integers for the compressed and uncompressed
231
# content. In base10 (ascii) 14 bytes can represent > 1TB, so to avoid
232
# checking too far, cap the search to 14 bytes.
233
pos2 = data.index(b'\n', pos, pos + 14)
234
self._z_content_length = int(data[pos:pos2])
236
pos2 = data.index(b'\n', pos, pos + 14)
237
self._content_length = int(data[pos:pos2])
239
if len(data) != (pos + self._z_content_length):
240
# XXX: Define some GCCorrupt error ?
241
raise AssertionError('Invalid bytes: (%d) != %d + %d' %
242
(len(data), pos, self._z_content_length))
243
self._z_content_chunks = (data[pos:],)
246
def _z_content(self):
247
"""Return z_content_chunks as a simple string.
249
Meant only to be used by the test suite.
251
if self._z_content_chunks is not None:
252
return b''.join(self._z_content_chunks)
256
def from_bytes(cls, bytes):
259
if header not in cls.GCB_KNOWN_HEADERS:
260
raise ValueError('bytes did not start with any of %r'
261
% (cls.GCB_KNOWN_HEADERS,))
262
if header == cls.GCB_HEADER:
263
out._compressor_name = 'zlib'
264
elif header == cls.GCB_LZ_HEADER:
265
out._compressor_name = 'lzma'
267
raise ValueError('unknown compressor: %r' % (header,))
268
out._parse_bytes(bytes, 6)
271
def extract(self, key, start, end, sha1=None):
272
"""Extract the text for a specific key.
274
:param key: The label used for this content
275
:param sha1: TODO (should we validate only when sha1 is supplied?)
276
:return: The bytes for the content
278
if start == end == 0:
280
self._ensure_content(end)
281
# The bytes are 'f' or 'd' for the type, then a variable-length
282
# base128 integer for the content size, then the actual content
283
# We know that the variable-length integer won't be longer than 5
284
# bytes (it takes 5 bytes to encode 2^32)
285
c = self._content[start:start + 1]
290
raise ValueError('Unknown content control code: %s'
293
content_len, len_len = decode_base128_int(
294
self._content[start + 1:start + 6])
295
content_start = start + 1 + len_len
296
if end != content_start + content_len:
297
raise ValueError('end != len according to field header'
298
' %s != %s' % (end, content_start + content_len))
300
return self._content[content_start:end]
301
# Must be type delta as checked above
302
return apply_delta_to_source(self._content, content_start, end)
304
def set_chunked_content(self, content_chunks, length):
305
"""Set the content of this block to the given chunks."""
306
# If we have lots of short lines, it is may be more efficient to join
307
# the content ahead of time. If the content is <10MiB, we don't really
308
# care about the extra memory consumption, so we can just pack it and
309
# be done. However, timing showed 18s => 17.9s for repacking 1k revs of
310
# mysql, which is below the noise margin
311
self._content_length = length
312
self._content_chunks = content_chunks
314
self._z_content_chunks = None
316
def set_content(self, content):
317
"""Set the content of this block."""
318
self._content_length = len(content)
319
self._content = content
320
self._z_content_chunks = None
322
def _create_z_content_from_chunks(self, chunks):
323
compressor = zlib.compressobj(zlib.Z_DEFAULT_COMPRESSION)
324
# Peak in this point is 1 fulltext, 1 compressed text, + zlib overhead
325
# (measured peak is maybe 30MB over the above...)
326
compressed_chunks = list(map(compressor.compress, chunks))
327
compressed_chunks.append(compressor.flush())
328
# Ignore empty chunks
329
self._z_content_chunks = [c for c in compressed_chunks if c]
330
self._z_content_length = sum(map(len, self._z_content_chunks))
332
def _create_z_content(self):
333
if self._z_content_chunks is not None:
335
if self._content_chunks is not None:
336
chunks = self._content_chunks
338
chunks = (self._content,)
339
self._create_z_content_from_chunks(chunks)
342
"""Create the byte stream as a series of 'chunks'"""
343
self._create_z_content()
344
header = self.GCB_HEADER
345
chunks = [b'%s%d\n%d\n'
346
% (header, self._z_content_length, self._content_length),
348
chunks.extend(self._z_content_chunks)
349
total_len = sum(map(len, chunks))
350
return total_len, chunks
353
"""Encode the information into a byte stream."""
354
total_len, chunks = self.to_chunks()
355
return b''.join(chunks)
357
def _dump(self, include_text=False):
358
"""Take this block, and spit out a human-readable structure.
360
:param include_text: Inserts also include text bits, chose whether you
361
want this displayed in the dump or not.
362
:return: A dump of the given block. The layout is something like:
363
[('f', length), ('d', delta_length, text_length, [delta_info])]
364
delta_info := [('i', num_bytes, text), ('c', offset, num_bytes),
367
self._ensure_content()
370
while pos < self._content_length:
371
kind = self._content[pos:pos + 1]
373
if kind not in (b'f', b'd'):
374
raise ValueError('invalid kind character: %r' % (kind,))
375
content_len, len_len = decode_base128_int(
376
self._content[pos:pos + 5])
378
if content_len + pos > self._content_length:
379
raise ValueError('invalid content_len %d for record @ pos %d'
380
% (content_len, pos - len_len - 1))
381
if kind == b'f': # Fulltext
383
text = self._content[pos:pos + content_len]
384
result.append((b'f', content_len, text))
386
result.append((b'f', content_len))
387
elif kind == b'd': # Delta
388
delta_content = self._content[pos:pos + content_len]
390
# The first entry in a delta is the decompressed length
391
decomp_len, delta_pos = decode_base128_int(delta_content)
392
result.append((b'd', content_len, decomp_len, delta_info))
394
while delta_pos < content_len:
395
c = indexbytes(delta_content, delta_pos)
399
delta_pos) = decode_copy_instruction(delta_content, c,
402
text = self._content[offset:offset + length]
403
delta_info.append((b'c', offset, length, text))
405
delta_info.append((b'c', offset, length))
406
measured_len += length
409
txt = delta_content[delta_pos:delta_pos + c]
412
delta_info.append((b'i', c, txt))
415
if delta_pos != content_len:
416
raise ValueError('Delta consumed a bad number of bytes:'
417
' %d != %d' % (delta_pos, content_len))
418
if measured_len != decomp_len:
419
raise ValueError('Delta claimed fulltext was %d bytes, but'
420
' extraction resulted in %d bytes'
421
% (decomp_len, measured_len))
426
class _LazyGroupCompressFactory(object):
427
"""Yield content from a GroupCompressBlock on demand."""
429
def __init__(self, key, parents, manager, start, end, first):
430
"""Create a _LazyGroupCompressFactory
432
:param key: The key of just this record
433
:param parents: The parents of this key (possibly None)
434
:param gc_block: A GroupCompressBlock object
435
:param start: Offset of the first byte for this record in the
437
:param end: Offset of the byte just after the end of this record
438
(ie, bytes = content[start:end])
439
:param first: Is this the first Factory for the given block?
442
self.parents = parents
444
# Note: This attribute coupled with Manager._factories creates a
445
# reference cycle. Perhaps we would rather use a weakref(), or
446
# find an appropriate time to release the ref. After the first
447
# get_bytes_as call? After Manager.get_record_stream() returns
449
self._manager = manager
451
self.storage_kind = 'groupcompress-block'
453
self.storage_kind = 'groupcompress-block-ref'
459
return '%s(%s, first=%s)' % (self.__class__.__name__,
460
self.key, self._first)
462
def get_bytes_as(self, storage_kind):
463
if storage_kind == self.storage_kind:
465
# wire bytes, something...
466
return self._manager._wire_bytes()
469
if storage_kind in ('fulltext', 'chunked'):
470
if self._bytes is None:
471
# Grab and cache the raw bytes for this entry
472
# and break the ref-cycle with _manager since we don't need it
475
self._manager._prepare_for_extract()
476
except zlib.error as value:
477
raise DecompressCorruption("zlib: " + str(value))
478
block = self._manager._block
479
self._bytes = block.extract(self.key, self._start, self._end)
480
# There are code paths that first extract as fulltext, and then
481
# extract as storage_kind (smart fetch). So we don't break the
482
# refcycle here, but instead in manager.get_record_stream()
483
if storage_kind == 'fulltext':
487
raise errors.UnavailableRepresentation(self.key, storage_kind,
491
class _LazyGroupContentManager(object):
492
"""This manages a group of _LazyGroupCompressFactory objects."""
494
_max_cut_fraction = 0.75 # We allow a block to be trimmed to 75% of
495
# current size, and still be considered
497
_full_block_size = 4 * 1024 * 1024
498
_full_mixed_block_size = 2 * 1024 * 1024
499
_full_enough_block_size = 3 * 1024 * 1024 # size at which we won't repack
500
_full_enough_mixed_block_size = 2 * 768 * 1024 # 1.5MB
502
def __init__(self, block, get_compressor_settings=None):
504
# We need to preserve the ordering
507
self._get_settings = get_compressor_settings
508
self._compressor_settings = None
510
def _get_compressor_settings(self):
511
if self._compressor_settings is not None:
512
return self._compressor_settings
514
if self._get_settings is not None:
515
settings = self._get_settings()
517
vf = GroupCompressVersionedFiles
518
settings = vf._DEFAULT_COMPRESSOR_SETTINGS
519
self._compressor_settings = settings
520
return self._compressor_settings
522
def add_factory(self, key, parents, start, end):
523
if not self._factories:
527
# Note that this creates a reference cycle....
528
factory = _LazyGroupCompressFactory(key, parents, self,
529
start, end, first=first)
530
# max() works here, but as a function call, doing a compare seems to be
531
# significantly faster, timeit says 250ms for max() and 100ms for the
533
if end > self._last_byte:
534
self._last_byte = end
535
self._factories.append(factory)
537
def get_record_stream(self):
538
"""Get a record for all keys added so far."""
539
for factory in self._factories:
541
# Break the ref-cycle
542
factory._bytes = None
543
factory._manager = None
544
# TODO: Consider setting self._factories = None after the above loop,
545
# as it will break the reference cycle
547
def _trim_block(self, last_byte):
548
"""Create a new GroupCompressBlock, with just some of the content."""
549
# None of the factories need to be adjusted, because the content is
550
# located in an identical place. Just that some of the unreferenced
551
# trailing bytes are stripped
552
trace.mutter('stripping trailing bytes from groupcompress block'
553
' %d => %d', self._block._content_length, last_byte)
554
new_block = GroupCompressBlock()
555
self._block._ensure_content(last_byte)
556
new_block.set_content(self._block._content[:last_byte])
557
self._block = new_block
559
def _make_group_compressor(self):
560
return GroupCompressor(self._get_compressor_settings())
562
def _rebuild_block(self):
563
"""Create a new GroupCompressBlock with only the referenced texts."""
564
compressor = self._make_group_compressor()
566
old_length = self._block._content_length
568
for factory in self._factories:
569
bytes = factory.get_bytes_as('fulltext')
570
(found_sha1, start_point, end_point,
571
type) = compressor.compress(factory.key, bytes, factory.sha1)
572
# Now update this factory with the new offsets, etc
573
factory.sha1 = found_sha1
574
factory._start = start_point
575
factory._end = end_point
576
self._last_byte = end_point
577
new_block = compressor.flush()
578
# TODO: Should we check that new_block really *is* smaller than the old
579
# block? It seems hard to come up with a method that it would
580
# expand, since we do full compression again. Perhaps based on a
581
# request that ends up poorly ordered?
582
# TODO: If the content would have expanded, then we would want to
583
# handle a case where we need to split the block.
584
# Now that we have a user-tweakable option
585
# (max_bytes_to_index), it is possible that one person set it
586
# to a very low value, causing poor compression.
587
delta = time.time() - tstart
588
self._block = new_block
589
trace.mutter('creating new compressed block on-the-fly in %.3fs'
590
' %d bytes => %d bytes', delta, old_length,
591
self._block._content_length)
593
def _prepare_for_extract(self):
594
"""A _LazyGroupCompressFactory is about to extract to fulltext."""
595
# We expect that if one child is going to fulltext, all will be. This
596
# helps prevent all of them from extracting a small amount at a time.
597
# Which in itself isn't terribly expensive, but resizing 2MB 32kB at a
598
# time (self._block._content) is a little expensive.
599
self._block._ensure_content(self._last_byte)
601
def _check_rebuild_action(self):
602
"""Check to see if our block should be repacked."""
605
for factory in self._factories:
606
total_bytes_used += factory._end - factory._start
607
if last_byte_used < factory._end:
608
last_byte_used = factory._end
609
# If we are using more than half of the bytes from the block, we have
610
# nothing else to check
611
if total_bytes_used * 2 >= self._block._content_length:
612
return None, last_byte_used, total_bytes_used
613
# We are using less than 50% of the content. Is the content we are
614
# using at the beginning of the block? If so, we can just trim the
615
# tail, rather than rebuilding from scratch.
616
if total_bytes_used * 2 > last_byte_used:
617
return 'trim', last_byte_used, total_bytes_used
619
# We are using a small amount of the data, and it isn't just packed
620
# nicely at the front, so rebuild the content.
621
# Note: This would be *nicer* as a strip-data-from-group, rather than
622
# building it up again from scratch
623
# It might be reasonable to consider the fulltext sizes for
624
# different bits when deciding this, too. As you may have a small
625
# fulltext, and a trivial delta, and you are just trading around
626
# for another fulltext. If we do a simple 'prune' you may end up
627
# expanding many deltas into fulltexts, as well.
628
# If we build a cheap enough 'strip', then we could try a strip,
629
# if that expands the content, we then rebuild.
630
return 'rebuild', last_byte_used, total_bytes_used
632
def check_is_well_utilized(self):
633
"""Is the current block considered 'well utilized'?
635
This heuristic asks if the current block considers itself to be a fully
636
developed group, rather than just a loose collection of data.
638
if len(self._factories) == 1:
639
# A block of length 1 could be improved by combining with other
640
# groups - don't look deeper. Even larger than max size groups
641
# could compress well with adjacent versions of the same thing.
643
action, last_byte_used, total_bytes_used = self._check_rebuild_action()
644
block_size = self._block._content_length
645
if total_bytes_used < block_size * self._max_cut_fraction:
646
# This block wants to trim itself small enough that we want to
647
# consider it under-utilized.
649
# TODO: This code is meant to be the twin of _insert_record_stream's
650
# 'start_new_block' logic. It would probably be better to factor
651
# out that logic into a shared location, so that it stays
653
# We currently assume a block is properly utilized whenever it is >75%
654
# of the size of a 'full' block. In normal operation, a block is
655
# considered full when it hits 4MB of same-file content. So any block
656
# >3MB is 'full enough'.
657
# The only time this isn't true is when a given block has large-object
658
# content. (a single file >4MB, etc.)
659
# Under these circumstances, we allow a block to grow to
660
# 2 x largest_content. Which means that if a given block had a large
661
# object, it may actually be under-utilized. However, given that this
662
# is 'pack-on-the-fly' it is probably reasonable to not repack large
663
# content blobs on-the-fly. Note that because we return False for all
664
# 1-item blobs, we will repack them; we may wish to reevaluate our
665
# treatment of large object blobs in the future.
666
if block_size >= self._full_enough_block_size:
668
# If a block is <3MB, it still may be considered 'full' if it contains
669
# mixed content. The current rule is 2MB of mixed content is considered
670
# full. So check to see if this block contains mixed content, and
671
# set the threshold appropriately.
673
for factory in self._factories:
674
prefix = factory.key[:-1]
675
if common_prefix is None:
676
common_prefix = prefix
677
elif prefix != common_prefix:
678
# Mixed content, check the size appropriately
679
if block_size >= self._full_enough_mixed_block_size:
682
# The content failed both the mixed check and the single-content check
683
# so obviously it is not fully utilized
684
# TODO: there is one other constraint that isn't being checked
685
# namely, that the entries in the block are in the appropriate
686
# order. For example, you could insert the entries in exactly
687
# reverse groupcompress order, and we would think that is ok.
688
# (all the right objects are in one group, and it is fully
689
# utilized, etc.) For now, we assume that case is rare,
690
# especially since we should always fetch in 'groupcompress'
694
def _check_rebuild_block(self):
695
action, last_byte_used, total_bytes_used = self._check_rebuild_action()
699
self._trim_block(last_byte_used)
700
elif action == 'rebuild':
701
self._rebuild_block()
703
raise ValueError('unknown rebuild action: %r' % (action,))
705
def _wire_bytes(self):
706
"""Return a byte stream suitable for transmitting over the wire."""
707
self._check_rebuild_block()
708
# The outer block starts with:
709
# 'groupcompress-block\n'
710
# <length of compressed key info>\n
711
# <length of uncompressed info>\n
712
# <length of gc block>\n
715
lines = [b'groupcompress-block\n']
716
# The minimal info we need is the key, the start offset, and the
717
# parents. The length and type are encoded in the record itself.
718
# However, passing in the other bits makes it easier. The list of
719
# keys, and the start offset, the length
721
# 1 line with parents, '' for ()
722
# 1 line for start offset
723
# 1 line for end byte
725
for factory in self._factories:
726
key_bytes = b'\x00'.join(factory.key)
727
parents = factory.parents
729
parent_bytes = b'None:'
731
parent_bytes = b'\t'.join(b'\x00'.join(key) for key in parents)
732
record_header = b'%s\n%s\n%d\n%d\n' % (
733
key_bytes, parent_bytes, factory._start, factory._end)
734
header_lines.append(record_header)
735
# TODO: Can we break the refcycle at this point and set
736
# factory._manager = None?
737
header_bytes = b''.join(header_lines)
739
header_bytes_len = len(header_bytes)
740
z_header_bytes = zlib.compress(header_bytes)
742
z_header_bytes_len = len(z_header_bytes)
743
block_bytes_len, block_chunks = self._block.to_chunks()
744
lines.append(b'%d\n%d\n%d\n' % (
745
z_header_bytes_len, header_bytes_len, block_bytes_len))
746
lines.append(z_header_bytes)
747
lines.extend(block_chunks)
748
del z_header_bytes, block_chunks
749
# TODO: This is a point where we will double the memory consumption. To
750
# avoid this, we probably have to switch to a 'chunked' api
751
return b''.join(lines)
754
def from_bytes(cls, bytes):
755
# TODO: This does extra string copying, probably better to do it a
756
# different way. At a minimum this creates 2 copies of the
758
(storage_kind, z_header_len, header_len,
759
block_len, rest) = bytes.split(b'\n', 4)
761
if storage_kind != b'groupcompress-block':
762
raise ValueError('Unknown storage kind: %s' % (storage_kind,))
763
z_header_len = int(z_header_len)
764
if len(rest) < z_header_len:
765
raise ValueError('Compressed header len shorter than all bytes')
766
z_header = rest[:z_header_len]
767
header_len = int(header_len)
768
header = zlib.decompress(z_header)
769
if len(header) != header_len:
770
raise ValueError('invalid length for decompressed bytes')
772
block_len = int(block_len)
773
if len(rest) != z_header_len + block_len:
774
raise ValueError('Invalid length for block')
775
block_bytes = rest[z_header_len:]
777
# So now we have a valid GCB, we just need to parse the factories that
779
header_lines = header.split(b'\n')
781
last = header_lines.pop()
783
raise ValueError('header lines did not end with a trailing'
785
if len(header_lines) % 4 != 0:
786
raise ValueError('The header was not an even multiple of 4 lines')
787
block = GroupCompressBlock.from_bytes(block_bytes)
790
for start in range(0, len(header_lines), 4):
792
key = tuple(header_lines[start].split(b'\x00'))
793
parents_line = header_lines[start + 1]
794
if parents_line == b'None:':
797
parents = tuple([tuple(segment.split(b'\x00'))
798
for segment in parents_line.split(b'\t')
800
start_offset = int(header_lines[start + 2])
801
end_offset = int(header_lines[start + 3])
802
result.add_factory(key, parents, start_offset, end_offset)
806
def network_block_to_records(storage_kind, bytes, line_end):
807
if storage_kind != 'groupcompress-block':
808
raise ValueError('Unknown storage kind: %s' % (storage_kind,))
809
manager = _LazyGroupContentManager.from_bytes(bytes)
810
return manager.get_record_stream()
813
class _CommonGroupCompressor(object):
815
def __init__(self, settings=None):
816
"""Create a GroupCompressor."""
821
self.labels_deltas = {}
822
self._delta_index = None # Set by the children
823
self._block = GroupCompressBlock()
827
self._settings = settings
829
def compress(self, key, bytes, expected_sha, nostore_sha=None, soft=False):
830
"""Compress lines with label key.
832
:param key: A key tuple. It is stored in the output
833
for identification of the text during decompression. If the last
834
element is b'None' it is replaced with the sha1 of the text -
836
:param bytes: The bytes to be compressed
837
:param expected_sha: If non-None, the sha the lines are believed to
838
have. During compression the sha is calculated; a mismatch will
840
:param nostore_sha: If the computed sha1 sum matches, we will raise
841
ExistingContent rather than adding the text.
842
:param soft: Do a 'soft' compression. This means that we require larger
843
ranges to match to be considered for a copy command.
845
:return: The sha1 of lines, the start and end offsets in the delta, and
846
the type ('fulltext' or 'delta').
848
:seealso VersionedFiles.add_lines:
850
if not bytes: # empty, like a dir entry, etc
851
if nostore_sha == _null_sha1:
852
raise errors.ExistingContent()
853
return _null_sha1, 0, 0, 'fulltext'
854
# we assume someone knew what they were doing when they passed it in
855
if expected_sha is not None:
858
sha1 = osutils.sha_string(bytes)
859
if nostore_sha is not None:
860
if sha1 == nostore_sha:
861
raise errors.ExistingContent()
863
key = key[:-1] + (b'sha1:' + sha1,)
865
start, end, type = self._compress(key, bytes, len(bytes) / 2, soft)
866
return sha1, start, end, type
868
def _compress(self, key, bytes, max_delta_size, soft=False):
869
"""Compress lines with label key.
871
:param key: A key tuple. It is stored in the output for identification
872
of the text during decompression.
874
:param bytes: The bytes to be compressed
876
:param max_delta_size: The size above which we issue a fulltext instead
879
:param soft: Do a 'soft' compression. This means that we require larger
880
ranges to match to be considered for a copy command.
882
:return: The sha1 of lines, the start and end offsets in the delta, and
883
the type ('fulltext' or 'delta').
885
raise NotImplementedError(self._compress)
887
def extract(self, key):
888
"""Extract a key previously added to the compressor.
890
:param key: The key to extract.
891
:return: An iterable over bytes and the sha1.
893
(start_byte, start_chunk, end_byte,
894
end_chunk) = self.labels_deltas[key]
895
delta_chunks = self.chunks[start_chunk:end_chunk]
896
stored_bytes = b''.join(delta_chunks)
897
kind = stored_bytes[:1]
899
fulltext_len, offset = decode_base128_int(stored_bytes[1:10])
900
data_len = fulltext_len + 1 + offset
901
if data_len != len(stored_bytes):
902
raise ValueError('Index claimed fulltext len, but stored bytes'
904
% (len(stored_bytes), data_len))
905
data = stored_bytes[offset + 1:]
908
raise ValueError('Unknown content kind, bytes claim %s' % kind)
909
# XXX: This is inefficient at best
910
source = b''.join(self.chunks[:start_chunk])
911
delta_len, offset = decode_base128_int(stored_bytes[1:10])
912
data_len = delta_len + 1 + offset
913
if data_len != len(stored_bytes):
914
raise ValueError('Index claimed delta len, but stored bytes'
916
% (len(stored_bytes), data_len))
917
data = apply_delta(source, stored_bytes[offset + 1:])
918
data_sha1 = osutils.sha_string(data)
919
return data, data_sha1
922
"""Finish this group, creating a formatted stream.
924
After calling this, the compressor should no longer be used
926
self._block.set_chunked_content(self.chunks, self.endpoint)
928
self._delta_index = None
932
"""Call this if you want to 'revoke' the last compression.
934
After this, the data structures will be rolled back, but you cannot do
937
self._delta_index = None
938
del self.chunks[self._last[0]:]
939
self.endpoint = self._last[1]
943
"""Return the overall compression ratio."""
944
return float(self.input_bytes) / float(self.endpoint)
947
class PythonGroupCompressor(_CommonGroupCompressor):
949
def __init__(self, settings=None):
950
"""Create a GroupCompressor.
952
Used only if the pyrex version is not available.
954
super(PythonGroupCompressor, self).__init__(settings)
955
self._delta_index = LinesDeltaIndex([])
956
# The actual content is managed by LinesDeltaIndex
957
self.chunks = self._delta_index.lines
959
def _compress(self, key, bytes, max_delta_size, soft=False):
960
"""see _CommonGroupCompressor._compress"""
961
input_len = len(bytes)
962
new_lines = osutils.split_lines(bytes)
963
out_lines, index_lines = self._delta_index.make_delta(
964
new_lines, bytes_length=input_len, soft=soft)
965
delta_length = sum(map(len, out_lines))
966
if delta_length > max_delta_size:
967
# The delta is longer than the fulltext, insert a fulltext
969
out_lines = [b'f', encode_base128_int(input_len)]
970
out_lines.extend(new_lines)
971
index_lines = [False, False]
972
index_lines.extend([True] * len(new_lines))
974
# this is a worthy delta, output it
977
# Update the delta_length to include those two encoded integers
978
out_lines[1] = encode_base128_int(delta_length)
980
start = self.endpoint
981
chunk_start = len(self.chunks)
982
self._last = (chunk_start, self.endpoint)
983
self._delta_index.extend_lines(out_lines, index_lines)
984
self.endpoint = self._delta_index.endpoint
985
self.input_bytes += input_len
986
chunk_end = len(self.chunks)
987
self.labels_deltas[key] = (start, chunk_start,
988
self.endpoint, chunk_end)
989
return start, self.endpoint, type
992
class PyrexGroupCompressor(_CommonGroupCompressor):
993
"""Produce a serialised group of compressed texts.
995
It contains code very similar to SequenceMatcher because of having a similar
996
task. However some key differences apply:
998
* there is no junk, we want a minimal edit not a human readable diff.
999
* we don't filter very common lines (because we don't know where a good
1000
range will start, and after the first text we want to be emitting minmal
1002
* we chain the left side, not the right side
1003
* we incrementally update the adjacency matrix as new lines are provided.
1004
* we look for matches in all of the left side, so the routine which does
1005
the analagous task of find_longest_match does not need to filter on the
1009
def __init__(self, settings=None):
1010
super(PyrexGroupCompressor, self).__init__(settings)
1011
max_bytes_to_index = self._settings.get('max_bytes_to_index', 0)
1012
self._delta_index = DeltaIndex(max_bytes_to_index=max_bytes_to_index)
1014
def _compress(self, key, bytes, max_delta_size, soft=False):
1015
"""see _CommonGroupCompressor._compress"""
1016
input_len = len(bytes)
1017
# By having action/label/sha1/len, we can parse the group if the index
1018
# was ever destroyed, we have the key in 'label', we know the final
1019
# bytes are valid from sha1, and we know where to find the end of this
1020
# record because of 'len'. (the delta record itself will store the
1021
# total length for the expanded record)
1022
# 'len: %d\n' costs approximately 1% increase in total data
1023
# Having the labels at all costs us 9-10% increase, 38% increase for
1024
# inventory pages, and 5.8% increase for text pages
1025
# new_chunks = ['label:%s\nsha1:%s\n' % (label, sha1)]
1026
if self._delta_index._source_offset != self.endpoint:
1027
raise AssertionError('_source_offset != endpoint'
1028
' somehow the DeltaIndex got out of sync with'
1029
' the output lines')
1030
delta = self._delta_index.make_delta(bytes, max_delta_size)
1033
enc_length = encode_base128_int(len(bytes))
1034
len_mini_header = 1 + len(enc_length)
1035
self._delta_index.add_source(bytes, len_mini_header)
1036
new_chunks = [b'f', enc_length, bytes]
1039
enc_length = encode_base128_int(len(delta))
1040
len_mini_header = 1 + len(enc_length)
1041
new_chunks = [b'd', enc_length, delta]
1042
self._delta_index.add_delta_source(delta, len_mini_header)
1044
start = self.endpoint
1045
chunk_start = len(self.chunks)
1046
# Now output these bytes
1047
self._output_chunks(new_chunks)
1048
self.input_bytes += input_len
1049
chunk_end = len(self.chunks)
1050
self.labels_deltas[key] = (start, chunk_start,
1051
self.endpoint, chunk_end)
1052
if not self._delta_index._source_offset == self.endpoint:
1053
raise AssertionError('the delta index is out of sync'
1054
'with the output lines %s != %s'
1055
% (self._delta_index._source_offset, self.endpoint))
1056
return start, self.endpoint, type
1058
def _output_chunks(self, new_chunks):
1059
"""Output some chunks.
1061
:param new_chunks: The chunks to output.
1063
self._last = (len(self.chunks), self.endpoint)
1064
endpoint = self.endpoint
1065
self.chunks.extend(new_chunks)
1066
endpoint += sum(map(len, new_chunks))
1067
self.endpoint = endpoint
1070
def make_pack_factory(graph, delta, keylength, inconsistency_fatal=True):
1071
"""Create a factory for creating a pack based groupcompress.
1073
This is only functional enough to run interface tests, it doesn't try to
1074
provide a full pack environment.
1076
:param graph: Store a graph.
1077
:param delta: Delta compress contents.
1078
:param keylength: How long should keys be.
1080
def factory(transport):
1085
graph_index = BTreeBuilder(reference_lists=ref_length,
1086
key_elements=keylength)
1087
stream = transport.open_write_stream('newpack')
1088
writer = pack.ContainerWriter(stream.write)
1090
index = _GCGraphIndex(graph_index, lambda: True, parents=parents,
1091
add_callback=graph_index.add_nodes,
1092
inconsistency_fatal=inconsistency_fatal)
1093
access = pack_repo._DirectPackAccess({})
1094
access.set_writer(writer, graph_index, (transport, 'newpack'))
1095
result = GroupCompressVersionedFiles(index, access, delta)
1096
result.stream = stream
1097
result.writer = writer
1102
def cleanup_pack_group(versioned_files):
1103
versioned_files.writer.end()
1104
versioned_files.stream.close()
1107
class _BatchingBlockFetcher(object):
1108
"""Fetch group compress blocks in batches.
1110
:ivar total_bytes: int of expected number of bytes needed to fetch the
1111
currently pending batch.
1114
def __init__(self, gcvf, locations, get_compressor_settings=None):
1116
self.locations = locations
1118
self.batch_memos = {}
1119
self.memos_to_get = []
1120
self.total_bytes = 0
1121
self.last_read_memo = None
1123
self._get_compressor_settings = get_compressor_settings
1125
def add_key(self, key):
1126
"""Add another to key to fetch.
1128
:return: The estimated number of bytes needed to fetch the batch so
1131
self.keys.append(key)
1132
index_memo, _, _, _ = self.locations[key]
1133
read_memo = index_memo[0:3]
1134
# Three possibilities for this read_memo:
1135
# - it's already part of this batch; or
1136
# - it's not yet part of this batch, but is already cached; or
1137
# - it's not yet part of this batch and will need to be fetched.
1138
if read_memo in self.batch_memos:
1139
# This read memo is already in this batch.
1140
return self.total_bytes
1142
cached_block = self.gcvf._group_cache[read_memo]
1144
# This read memo is new to this batch, and the data isn't cached
1146
self.batch_memos[read_memo] = None
1147
self.memos_to_get.append(read_memo)
1148
byte_length = read_memo[2]
1149
self.total_bytes += byte_length
1151
# This read memo is new to this batch, but cached.
1152
# Keep a reference to the cached block in batch_memos because it's
1153
# certain that we'll use it when this batch is processed, but
1154
# there's a risk that it would fall out of _group_cache between now
1156
self.batch_memos[read_memo] = cached_block
1157
return self.total_bytes
1159
def _flush_manager(self):
1160
if self.manager is not None:
1161
for factory in self.manager.get_record_stream():
1164
self.last_read_memo = None
1166
def yield_factories(self, full_flush=False):
1167
"""Yield factories for keys added since the last yield. They will be
1168
returned in the order they were added via add_key.
1170
:param full_flush: by default, some results may not be returned in case
1171
they can be part of the next batch. If full_flush is True, then
1172
all results are returned.
1174
if self.manager is None and not self.keys:
1176
# Fetch all memos in this batch.
1177
blocks = self.gcvf._get_blocks(self.memos_to_get)
1178
# Turn blocks into factories and yield them.
1179
memos_to_get_stack = list(self.memos_to_get)
1180
memos_to_get_stack.reverse()
1181
for key in self.keys:
1182
index_memo, _, parents, _ = self.locations[key]
1183
read_memo = index_memo[:3]
1184
if self.last_read_memo != read_memo:
1185
# We are starting a new block. If we have a
1186
# manager, we have found everything that fits for
1187
# now, so yield records
1188
for factory in self._flush_manager():
1190
# Now start a new manager.
1191
if memos_to_get_stack and memos_to_get_stack[-1] == read_memo:
1192
# The next block from _get_blocks will be the block we
1194
block_read_memo, block = next(blocks)
1195
if block_read_memo != read_memo:
1196
raise AssertionError(
1197
"block_read_memo out of sync with read_memo"
1198
"(%r != %r)" % (block_read_memo, read_memo))
1199
self.batch_memos[read_memo] = block
1200
memos_to_get_stack.pop()
1202
block = self.batch_memos[read_memo]
1203
self.manager = _LazyGroupContentManager(block,
1204
get_compressor_settings=self._get_compressor_settings)
1205
self.last_read_memo = read_memo
1206
start, end = index_memo[3:5]
1207
self.manager.add_factory(key, parents, start, end)
1209
for factory in self._flush_manager():
1212
self.batch_memos.clear()
1213
del self.memos_to_get[:]
1214
self.total_bytes = 0
1217
class GroupCompressVersionedFiles(VersionedFilesWithFallbacks):
1218
"""A group-compress based VersionedFiles implementation."""
1220
# This controls how the GroupCompress DeltaIndex works. Basically, we
1221
# compute hash pointers into the source blocks (so hash(text) => text).
1222
# However each of these references costs some memory in trade against a
1223
# more accurate match result. For very large files, they either are
1224
# pre-compressed and change in bulk whenever they change, or change in just
1225
# local blocks. Either way, 'improved resolution' is not very helpful,
1226
# versus running out of memory trying to track everything. The default max
1227
# gives 100% sampling of a 1MB file.
1228
_DEFAULT_MAX_BYTES_TO_INDEX = 1024 * 1024
1229
_DEFAULT_COMPRESSOR_SETTINGS = {'max_bytes_to_index':
1230
_DEFAULT_MAX_BYTES_TO_INDEX}
1232
def __init__(self, index, access, delta=True, _unadded_refs=None,
1234
"""Create a GroupCompressVersionedFiles object.
1236
:param index: The index object storing access and graph data.
1237
:param access: The access object storing raw data.
1238
:param delta: Whether to delta compress or just entropy compress.
1239
:param _unadded_refs: private parameter, don't use.
1240
:param _group_cache: private parameter, don't use.
1243
self._access = access
1245
if _unadded_refs is None:
1247
self._unadded_refs = _unadded_refs
1248
if _group_cache is None:
1249
_group_cache = LRUSizeCache(max_size=50 * 1024 * 1024)
1250
self._group_cache = _group_cache
1251
self._immediate_fallback_vfs = []
1252
self._max_bytes_to_index = None
1254
def without_fallbacks(self):
1255
"""Return a clone of this object without any fallbacks configured."""
1256
return GroupCompressVersionedFiles(self._index, self._access,
1257
self._delta, _unadded_refs=dict(
1258
self._unadded_refs),
1259
_group_cache=self._group_cache)
1261
def add_lines(self, key, parents, lines, parent_texts=None,
1262
left_matching_blocks=None, nostore_sha=None, random_id=False,
1263
check_content=True):
1264
"""Add a text to the store.
1266
:param key: The key tuple of the text to add.
1267
:param parents: The parents key tuples of the text to add.
1268
:param lines: A list of lines. Each line must be a bytestring. And all
1269
of them except the last must be terminated with \\n and contain no
1270
other \\n's. The last line may either contain no \\n's or a single
1271
terminating \\n. If the lines list does meet this constraint the
1272
add routine may error or may succeed - but you will be unable to
1273
read the data back accurately. (Checking the lines have been split
1274
correctly is expensive and extremely unlikely to catch bugs so it
1275
is not done at runtime unless check_content is True.)
1276
:param parent_texts: An optional dictionary containing the opaque
1277
representations of some or all of the parents of version_id to
1278
allow delta optimisations. VERY IMPORTANT: the texts must be those
1279
returned by add_lines or data corruption can be caused.
1280
:param left_matching_blocks: a hint about which areas are common
1281
between the text and its left-hand-parent. The format is
1282
the SequenceMatcher.get_matching_blocks format.
1283
:param nostore_sha: Raise ExistingContent and do not add the lines to
1284
the versioned file if the digest of the lines matches this.
1285
:param random_id: If True a random id has been selected rather than
1286
an id determined by some deterministic process such as a converter
1287
from a foreign VCS. When True the backend may choose not to check
1288
for uniqueness of the resulting key within the versioned file, so
1289
this should only be done when the result is expected to be unique
1291
:param check_content: If True, the lines supplied are verified to be
1292
bytestrings that are correctly formed lines.
1293
:return: The text sha1, the number of bytes in the text, and an opaque
1294
representation of the inserted version which can be provided
1295
back to future add_lines calls in the parent_texts dictionary.
1297
self._index._check_write_ok()
1298
self._check_add(key, lines, random_id, check_content)
1300
# The caller might pass None if there is no graph data, but kndx
1301
# indexes can't directly store that, so we give them
1302
# an empty tuple instead.
1304
# double handling for now. Make it work until then.
1305
length = sum(map(len, lines))
1306
record = ChunkedContentFactory(key, parents, None, lines)
1307
sha1 = list(self._insert_record_stream([record], random_id=random_id,
1308
nostore_sha=nostore_sha))[0]
1309
return sha1, length, None
1311
def add_fallback_versioned_files(self, a_versioned_files):
1312
"""Add a source of texts for texts not present in this knit.
1314
:param a_versioned_files: A VersionedFiles object.
1316
self._immediate_fallback_vfs.append(a_versioned_files)
1318
def annotate(self, key):
1319
"""See VersionedFiles.annotate."""
1320
ann = annotate.Annotator(self)
1321
return ann.annotate_flat(key)
1323
def get_annotator(self):
1324
return annotate.Annotator(self)
1326
def check(self, progress_bar=None, keys=None):
1327
"""See VersionedFiles.check()."""
1330
for record in self.get_record_stream(keys, 'unordered', True):
1331
record.get_bytes_as('fulltext')
1333
return self.get_record_stream(keys, 'unordered', True)
1335
def clear_cache(self):
1336
"""See VersionedFiles.clear_cache()"""
1337
self._group_cache.clear()
1338
self._index._graph_index.clear_cache()
1339
self._index._int_cache.clear()
1341
def _check_add(self, key, lines, random_id, check_content):
1342
"""check that version_id and lines are safe to add."""
1343
version_id = key[-1]
1344
if version_id is not None:
1345
if osutils.contains_whitespace(version_id):
1346
raise errors.InvalidRevisionId(version_id, self)
1347
self.check_not_reserved_id(version_id)
1348
# TODO: If random_id==False and the key is already present, we should
1349
# probably check that the existing content is identical to what is
1350
# being inserted, and otherwise raise an exception. This would make
1351
# the bundle code simpler.
1353
self._check_lines_not_unicode(lines)
1354
self._check_lines_are_lines(lines)
1356
def get_parent_map(self, keys):
1357
"""Get a map of the graph parents of keys.
1359
:param keys: The keys to look up parents for.
1360
:return: A mapping from keys to parents. Absent keys are absent from
1363
return self._get_parent_map_with_sources(keys)[0]
1365
def _get_parent_map_with_sources(self, keys):
1366
"""Get a map of the parents of keys.
1368
:param keys: The keys to look up parents for.
1369
:return: A tuple. The first element is a mapping from keys to parents.
1370
Absent keys are absent from the mapping. The second element is a
1371
list with the locations each key was found in. The first element
1372
is the in-this-knit parents, the second the first fallback source,
1376
sources = [self._index] + self._immediate_fallback_vfs
1379
for source in sources:
1382
new_result = source.get_parent_map(missing)
1383
source_results.append(new_result)
1384
result.update(new_result)
1385
missing.difference_update(set(new_result))
1386
return result, source_results
1388
def _get_blocks(self, read_memos):
1389
"""Get GroupCompressBlocks for the given read_memos.
1391
:returns: a series of (read_memo, block) pairs, in the order they were
1395
for read_memo in read_memos:
1397
block = self._group_cache[read_memo]
1401
cached[read_memo] = block
1403
not_cached_seen = set()
1404
for read_memo in read_memos:
1405
if read_memo in cached:
1406
# Don't fetch what we already have
1408
if read_memo in not_cached_seen:
1409
# Don't try to fetch the same data twice
1411
not_cached.append(read_memo)
1412
not_cached_seen.add(read_memo)
1413
raw_records = self._access.get_raw_records(not_cached)
1414
for read_memo in read_memos:
1416
yield read_memo, cached[read_memo]
1418
# Read the block, and cache it.
1419
zdata = next(raw_records)
1420
block = GroupCompressBlock.from_bytes(zdata)
1421
self._group_cache[read_memo] = block
1422
cached[read_memo] = block
1423
yield read_memo, block
1425
def get_missing_compression_parent_keys(self):
1426
"""Return the keys of missing compression parents.
1428
Missing compression parents occur when a record stream was missing
1429
basis texts, or a index was scanned that had missing basis texts.
1431
# GroupCompress cannot currently reference texts that are not in the
1432
# group, so this is valid for now
1435
def get_record_stream(self, keys, ordering, include_delta_closure):
1436
"""Get a stream of records for keys.
1438
:param keys: The keys to include.
1439
:param ordering: Either 'unordered' or 'topological'. A topologically
1440
sorted stream has compression parents strictly before their
1442
:param include_delta_closure: If True then the closure across any
1443
compression parents will be included (in the opaque data).
1444
:return: An iterator of ContentFactory objects, each of which is only
1445
valid until the iterator is advanced.
1447
# keys might be a generator
1448
orig_keys = list(keys)
1452
if (not self._index.has_graph
1453
and ordering in ('topological', 'groupcompress')):
1454
# Cannot topological order when no graph has been stored.
1455
# but we allow 'as-requested' or 'unordered'
1456
ordering = 'unordered'
1458
remaining_keys = keys
1461
keys = set(remaining_keys)
1462
for content_factory in self._get_remaining_record_stream(keys,
1463
orig_keys, ordering, include_delta_closure):
1464
remaining_keys.discard(content_factory.key)
1465
yield content_factory
1467
except errors.RetryWithNewPacks as e:
1468
self._access.reload_or_raise(e)
1470
def _find_from_fallback(self, missing):
1471
"""Find whatever keys you can from the fallbacks.
1473
:param missing: A set of missing keys. This set will be mutated as keys
1474
are found from a fallback_vfs
1475
:return: (parent_map, key_to_source_map, source_results)
1476
parent_map the overall key => parent_keys
1477
key_to_source_map a dict from {key: source}
1478
source_results a list of (source: keys)
1481
key_to_source_map = {}
1483
for source in self._immediate_fallback_vfs:
1486
source_parents = source.get_parent_map(missing)
1487
parent_map.update(source_parents)
1488
source_parents = list(source_parents)
1489
source_results.append((source, source_parents))
1490
key_to_source_map.update((key, source) for key in source_parents)
1491
missing.difference_update(source_parents)
1492
return parent_map, key_to_source_map, source_results
1494
def _get_ordered_source_keys(self, ordering, parent_map, key_to_source_map):
1495
"""Get the (source, [keys]) list.
1497
The returned objects should be in the order defined by 'ordering',
1498
which can weave between different sources.
1500
:param ordering: Must be one of 'topological' or 'groupcompress'
1501
:return: List of [(source, [keys])] tuples, such that all keys are in
1502
the defined order, regardless of source.
1504
if ordering == 'topological':
1505
present_keys = tsort.topo_sort(parent_map)
1507
# ordering == 'groupcompress'
1508
# XXX: This only optimizes for the target ordering. We may need
1509
# to balance that with the time it takes to extract
1510
# ordering, by somehow grouping based on
1511
# locations[key][0:3]
1512
present_keys = sort_gc_optimal(parent_map)
1513
# Now group by source:
1515
current_source = None
1516
for key in present_keys:
1517
source = key_to_source_map.get(key, self)
1518
if source is not current_source:
1519
source_keys.append((source, []))
1520
current_source = source
1521
source_keys[-1][1].append(key)
1524
def _get_as_requested_source_keys(self, orig_keys, locations, unadded_keys,
1527
current_source = None
1528
for key in orig_keys:
1529
if key in locations or key in unadded_keys:
1531
elif key in key_to_source_map:
1532
source = key_to_source_map[key]
1535
if source is not current_source:
1536
source_keys.append((source, []))
1537
current_source = source
1538
source_keys[-1][1].append(key)
1541
def _get_io_ordered_source_keys(self, locations, unadded_keys,
1544
# This is the group the bytes are stored in, followed by the
1545
# location in the group
1546
return locations[key][0]
1547
# We don't have an ordering for keys in the in-memory object, but
1548
# lets process the in-memory ones first.
1549
present_keys = list(unadded_keys)
1550
present_keys.extend(sorted(locations, key=get_group))
1551
# Now grab all of the ones from other sources
1552
source_keys = [(self, present_keys)]
1553
source_keys.extend(source_result)
1556
def _get_remaining_record_stream(self, keys, orig_keys, ordering,
1557
include_delta_closure):
1558
"""Get a stream of records for keys.
1560
:param keys: The keys to include.
1561
:param ordering: one of 'unordered', 'topological', 'groupcompress' or
1563
:param include_delta_closure: If True then the closure across any
1564
compression parents will be included (in the opaque data).
1565
:return: An iterator of ContentFactory objects, each of which is only
1566
valid until the iterator is advanced.
1569
locations = self._index.get_build_details(keys)
1570
unadded_keys = set(self._unadded_refs).intersection(keys)
1571
missing = keys.difference(locations)
1572
missing.difference_update(unadded_keys)
1573
(fallback_parent_map, key_to_source_map,
1574
source_result) = self._find_from_fallback(missing)
1575
if ordering in ('topological', 'groupcompress'):
1576
# would be better to not globally sort initially but instead
1577
# start with one key, recurse to its oldest parent, then grab
1578
# everything in the same group, etc.
1579
parent_map = dict((key, details[2]) for key, details in
1580
viewitems(locations))
1581
for key in unadded_keys:
1582
parent_map[key] = self._unadded_refs[key]
1583
parent_map.update(fallback_parent_map)
1584
source_keys = self._get_ordered_source_keys(ordering, parent_map,
1586
elif ordering == 'as-requested':
1587
source_keys = self._get_as_requested_source_keys(orig_keys,
1588
locations, unadded_keys, key_to_source_map)
1590
# We want to yield the keys in a semi-optimal (read-wise) ordering.
1591
# Otherwise we thrash the _group_cache and destroy performance
1592
source_keys = self._get_io_ordered_source_keys(locations,
1593
unadded_keys, source_result)
1595
yield AbsentContentFactory(key)
1596
# Batch up as many keys as we can until either:
1597
# - we encounter an unadded ref, or
1598
# - we run out of keys, or
1599
# - the total bytes to retrieve for this batch > BATCH_SIZE
1600
batcher = _BatchingBlockFetcher(self, locations,
1601
get_compressor_settings=self._get_compressor_settings)
1602
for source, keys in source_keys:
1605
if key in self._unadded_refs:
1606
# Flush batch, then yield unadded ref from
1608
for factory in batcher.yield_factories(full_flush=True):
1610
bytes, sha1 = self._compressor.extract(key)
1611
parents = self._unadded_refs[key]
1612
yield FulltextContentFactory(key, parents, sha1, bytes)
1614
if batcher.add_key(key) > BATCH_SIZE:
1615
# Ok, this batch is big enough. Yield some results.
1616
for factory in batcher.yield_factories():
1619
for factory in batcher.yield_factories(full_flush=True):
1621
for record in source.get_record_stream(keys, ordering,
1622
include_delta_closure):
1624
for factory in batcher.yield_factories(full_flush=True):
1627
def get_sha1s(self, keys):
1628
"""See VersionedFiles.get_sha1s()."""
1630
for record in self.get_record_stream(keys, 'unordered', True):
1631
if record.sha1 is not None:
1632
result[record.key] = record.sha1
1634
if record.storage_kind != 'absent':
1635
result[record.key] = osutils.sha_string(
1636
record.get_bytes_as('fulltext'))
1639
def insert_record_stream(self, stream):
1640
"""Insert a record stream into this container.
1642
:param stream: A stream of records to insert.
1644
:seealso VersionedFiles.get_record_stream:
1646
# XXX: Setting random_id=True makes
1647
# test_insert_record_stream_existing_keys fail for groupcompress and
1648
# groupcompress-nograph, this needs to be revisited while addressing
1649
# 'bzr branch' performance issues.
1650
for _ in self._insert_record_stream(stream, random_id=False):
1653
def _get_compressor_settings(self):
1654
if self._max_bytes_to_index is None:
1655
# TODO: VersionedFiles don't know about their containing
1656
# repository, so they don't have much of an idea about their
1657
# location. So for now, this is only a global option.
1658
c = config.GlobalConfig()
1659
val = c.get_user_option('bzr.groupcompress.max_bytes_to_index')
1663
except ValueError as e:
1664
trace.warning('Value for '
1665
'"bzr.groupcompress.max_bytes_to_index"'
1666
' %r is not an integer'
1670
val = self._DEFAULT_MAX_BYTES_TO_INDEX
1671
self._max_bytes_to_index = val
1672
return {'max_bytes_to_index': self._max_bytes_to_index}
1674
def _make_group_compressor(self):
1675
return GroupCompressor(self._get_compressor_settings())
1677
def _insert_record_stream(self, stream, random_id=False, nostore_sha=None,
1679
"""Internal core to insert a record stream into this container.
1681
This helper function has a different interface than insert_record_stream
1682
to allow add_lines to be minimal, but still return the needed data.
1684
:param stream: A stream of records to insert.
1685
:param nostore_sha: If the sha1 of a given text matches nostore_sha,
1686
raise ExistingContent, rather than committing the new text.
1687
:param reuse_blocks: If the source is streaming from
1688
groupcompress-blocks, just insert the blocks as-is, rather than
1689
expanding the texts and inserting again.
1690
:return: An iterator over the sha1 of the inserted records.
1691
:seealso insert_record_stream:
1696
def get_adapter(adapter_key):
1698
return adapters[adapter_key]
1700
adapter_factory = adapter_registry.get(adapter_key)
1701
adapter = adapter_factory(self)
1702
adapters[adapter_key] = adapter
1704
# This will go up to fulltexts for gc to gc fetching, which isn't
1706
self._compressor = self._make_group_compressor()
1707
self._unadded_refs = {}
1711
bytes_len, chunks = self._compressor.flush().to_chunks()
1712
self._compressor = self._make_group_compressor()
1713
# Note: At this point we still have 1 copy of the fulltext (in
1714
# record and the var 'bytes'), and this generates 2 copies of
1715
# the compressed text (one for bytes, one in chunks)
1716
# TODO: Push 'chunks' down into the _access api, so that we don't
1717
# have to double compressed memory here
1718
# TODO: Figure out how to indicate that we would be happy to free
1719
# the fulltext content at this point. Note that sometimes we
1720
# will want it later (streaming CHK pages), but most of the
1721
# time we won't (everything else)
1722
data = b''.join(chunks)
1724
index, start, length = self._access.add_raw_records(
1725
[(None, len(data))], data)[0]
1727
for key, reads, refs in keys_to_add:
1728
nodes.append((key, b"%d %d %s" % (start, length, reads), refs))
1729
self._index.add_records(nodes, random_id=random_id)
1730
self._unadded_refs = {}
1734
max_fulltext_len = 0
1735
max_fulltext_prefix = None
1736
insert_manager = None
1739
# XXX: TODO: remove this, it is just for safety checking for now
1740
inserted_keys = set()
1741
reuse_this_block = reuse_blocks
1742
for record in stream:
1743
# Raise an error when a record is missing.
1744
if record.storage_kind == 'absent':
1745
raise errors.RevisionNotPresent(record.key, self)
1747
if record.key in inserted_keys:
1748
trace.note(gettext('Insert claimed random_id=True,'
1749
' but then inserted %r two times'), record.key)
1751
inserted_keys.add(record.key)
1753
# If the reuse_blocks flag is set, check to see if we can just
1754
# copy a groupcompress block as-is.
1755
# We only check on the first record (groupcompress-block) not
1756
# on all of the (groupcompress-block-ref) entries.
1757
# The reuse_this_block flag is then kept for as long as
1758
if record.storage_kind == 'groupcompress-block':
1759
# Check to see if we really want to re-use this block
1760
insert_manager = record._manager
1761
reuse_this_block = insert_manager.check_is_well_utilized()
1763
reuse_this_block = False
1764
if reuse_this_block:
1765
# We still want to reuse this block
1766
if record.storage_kind == 'groupcompress-block':
1767
# Insert the raw block into the target repo
1768
insert_manager = record._manager
1769
bytes = record._manager._block.to_bytes()
1770
_, start, length = self._access.add_raw_records(
1771
[(None, len(bytes))], bytes)[0]
1774
block_length = length
1775
if record.storage_kind in ('groupcompress-block',
1776
'groupcompress-block-ref'):
1777
if insert_manager is None:
1778
raise AssertionError('No insert_manager set')
1779
if insert_manager is not record._manager:
1780
raise AssertionError('insert_manager does not match'
1781
' the current record, we cannot be positive'
1782
' that the appropriate content was inserted.'
1784
value = b"%d %d %d %d" % (block_start, block_length,
1785
record._start, record._end)
1786
nodes = [(record.key, value, (record.parents,))]
1787
# TODO: Consider buffering up many nodes to be added, not
1788
# sure how much overhead this has, but we're seeing
1789
# ~23s / 120s in add_records calls
1790
self._index.add_records(nodes, random_id=random_id)
1793
bytes = record.get_bytes_as('fulltext')
1794
except errors.UnavailableRepresentation:
1795
adapter_key = record.storage_kind, 'fulltext'
1796
adapter = get_adapter(adapter_key)
1797
bytes = adapter.get_bytes(record)
1798
if len(record.key) > 1:
1799
prefix = record.key[0]
1800
soft = (prefix == last_prefix)
1804
if max_fulltext_len < len(bytes):
1805
max_fulltext_len = len(bytes)
1806
max_fulltext_prefix = prefix
1807
(found_sha1, start_point, end_point,
1808
type) = self._compressor.compress(record.key,
1809
bytes, record.sha1, soft=soft,
1810
nostore_sha=nostore_sha)
1811
# delta_ratio = float(len(bytes)) / (end_point - start_point)
1812
# Check if we want to continue to include that text
1813
if (prefix == max_fulltext_prefix
1814
and end_point < 2 * max_fulltext_len):
1815
# As long as we are on the same file_id, we will fill at least
1816
# 2 * max_fulltext_len
1817
start_new_block = False
1818
elif end_point > 4 * 1024 * 1024:
1819
start_new_block = True
1820
elif (prefix is not None and prefix != last_prefix
1821
and end_point > 2 * 1024 * 1024):
1822
start_new_block = True
1824
start_new_block = False
1825
last_prefix = prefix
1827
self._compressor.pop_last()
1829
max_fulltext_len = len(bytes)
1830
(found_sha1, start_point, end_point,
1831
type) = self._compressor.compress(record.key, bytes,
1833
if record.key[-1] is None:
1834
key = record.key[:-1] + (b'sha1:' + found_sha1,)
1837
self._unadded_refs[key] = record.parents
1839
as_st = static_tuple.StaticTuple.from_sequence
1840
if record.parents is not None:
1841
parents = as_st([as_st(p) for p in record.parents])
1844
refs = static_tuple.StaticTuple(parents)
1846
(key, b'%d %d' % (start_point, end_point), refs))
1847
if len(keys_to_add):
1849
self._compressor = None
1851
def iter_lines_added_or_present_in_keys(self, keys, pb=None):
1852
"""Iterate over the lines in the versioned files from keys.
1854
This may return lines from other keys. Each item the returned
1855
iterator yields is a tuple of a line and a text version that that line
1856
is present in (not introduced in).
1858
Ordering of results is in whatever order is most suitable for the
1859
underlying storage format.
1861
If a progress bar is supplied, it may be used to indicate progress.
1862
The caller is responsible for cleaning up progress bars (because this
1866
* Lines are normalised by the underlying store: they will all have \n
1868
* Lines are returned in arbitrary order.
1870
:return: An iterator over (line, key).
1874
# we don't care about inclusions, the caller cares.
1875
# but we need to setup a list of records to visit.
1876
# we need key, position, length
1877
for key_idx, record in enumerate(self.get_record_stream(keys,
1878
'unordered', True)):
1879
# XXX: todo - optimise to use less than full texts.
1882
pb.update('Walking content', key_idx, total)
1883
if record.storage_kind == 'absent':
1884
raise errors.RevisionNotPresent(key, self)
1885
lines = osutils.split_lines(record.get_bytes_as('fulltext'))
1889
pb.update('Walking content', total, total)
1892
"""See VersionedFiles.keys."""
1893
if 'evil' in debug.debug_flags:
1894
trace.mutter_callsite(2, "keys scales with size of history")
1895
sources = [self._index] + self._immediate_fallback_vfs
1897
for source in sources:
1898
result.update(source.keys())
1902
class _GCBuildDetails(object):
1903
"""A blob of data about the build details.
1905
This stores the minimal data, which then allows compatibility with the old
1906
api, without taking as much memory.
1909
__slots__ = ('_index', '_group_start', '_group_end', '_basis_end',
1910
'_delta_end', '_parents')
1913
compression_parent = None
1915
def __init__(self, parents, position_info):
1916
self._parents = parents
1917
(self._index, self._group_start, self._group_end, self._basis_end,
1918
self._delta_end) = position_info
1921
return '%s(%s, %s)' % (self.__class__.__name__,
1922
self.index_memo, self._parents)
1925
def index_memo(self):
1926
return (self._index, self._group_start, self._group_end,
1927
self._basis_end, self._delta_end)
1930
def record_details(self):
1931
return static_tuple.StaticTuple(self.method, None)
1933
def __getitem__(self, offset):
1934
"""Compatibility thunk to act like a tuple."""
1936
return self.index_memo
1938
return self.compression_parent # Always None
1940
return self._parents
1942
return self.record_details
1944
raise IndexError('offset out of range')
1950
class _GCGraphIndex(object):
1951
"""Mapper from GroupCompressVersionedFiles needs into GraphIndex storage."""
1953
def __init__(self, graph_index, is_locked, parents=True,
1954
add_callback=None, track_external_parent_refs=False,
1955
inconsistency_fatal=True, track_new_keys=False):
1956
"""Construct a _GCGraphIndex on a graph_index.
1958
:param graph_index: An implementation of breezy.index.GraphIndex.
1959
:param is_locked: A callback, returns True if the index is locked and
1961
:param parents: If True, record knits parents, if not do not record
1963
:param add_callback: If not None, allow additions to the index and call
1964
this callback with a list of added GraphIndex nodes:
1965
[(node, value, node_refs), ...]
1966
:param track_external_parent_refs: As keys are added, keep track of the
1967
keys they reference, so that we can query get_missing_parents(),
1969
:param inconsistency_fatal: When asked to add records that are already
1970
present, and the details are inconsistent with the existing
1971
record, raise an exception instead of warning (and skipping the
1974
self._add_callback = add_callback
1975
self._graph_index = graph_index
1976
self._parents = parents
1977
self.has_graph = parents
1978
self._is_locked = is_locked
1979
self._inconsistency_fatal = inconsistency_fatal
1980
# GroupCompress records tend to have the same 'group' start + offset
1981
# repeated over and over, this creates a surplus of ints
1982
self._int_cache = {}
1983
if track_external_parent_refs:
1984
self._key_dependencies = _KeyRefs(
1985
track_new_keys=track_new_keys)
1987
self._key_dependencies = None
1989
def add_records(self, records, random_id=False):
1990
"""Add multiple records to the index.
1992
This function does not insert data into the Immutable GraphIndex
1993
backing the KnitGraphIndex, instead it prepares data for insertion by
1994
the caller and checks that it is safe to insert then calls
1995
self._add_callback with the prepared GraphIndex nodes.
1997
:param records: a list of tuples:
1998
(key, options, access_memo, parents).
1999
:param random_id: If True the ids being added were randomly generated
2000
and no check for existence will be performed.
2002
if not self._add_callback:
2003
raise errors.ReadOnlyError(self)
2004
# we hope there are no repositories with inconsistent parentage
2009
for (key, value, refs) in records:
2010
if not self._parents:
2014
raise knit.KnitCorrupt(self,
2015
"attempt to add node with parents "
2016
"in parentless index.")
2019
keys[key] = (value, refs)
2022
present_nodes = self._get_entries(keys)
2023
for (index, key, value, node_refs) in present_nodes:
2024
# Sometimes these are passed as a list rather than a tuple
2025
node_refs = static_tuple.as_tuples(node_refs)
2026
passed = static_tuple.as_tuples(keys[key])
2027
if node_refs != passed[1]:
2028
details = '%s %s %s' % (key, (value, node_refs), passed)
2029
if self._inconsistency_fatal:
2030
raise knit.KnitCorrupt(self, "inconsistent details"
2031
" in add_records: %s" %
2034
trace.warning("inconsistent details in skipped"
2035
" record: %s", details)
2041
for key, (value, node_refs) in viewitems(keys):
2042
result.append((key, value, node_refs))
2044
for key, (value, node_refs) in viewitems(keys):
2045
result.append((key, value))
2047
key_dependencies = self._key_dependencies
2048
if key_dependencies is not None:
2050
for key, value, refs in records:
2052
key_dependencies.add_references(key, parents)
2054
for key, value, refs in records:
2055
new_keys.add_key(key)
2056
self._add_callback(records)
2058
def _check_read(self):
2059
"""Raise an exception if reads are not permitted."""
2060
if not self._is_locked():
2061
raise errors.ObjectNotLocked(self)
2063
def _check_write_ok(self):
2064
"""Raise an exception if writes are not permitted."""
2065
if not self._is_locked():
2066
raise errors.ObjectNotLocked(self)
2068
def _get_entries(self, keys, check_present=False):
2069
"""Get the entries for keys.
2071
Note: Callers are responsible for checking that the index is locked
2072
before calling this method.
2074
:param keys: An iterable of index key tuples.
2079
for node in self._graph_index.iter_entries(keys):
2081
found_keys.add(node[1])
2083
# adapt parentless index to the rest of the code.
2084
for node in self._graph_index.iter_entries(keys):
2085
yield node[0], node[1], node[2], ()
2086
found_keys.add(node[1])
2088
missing_keys = keys.difference(found_keys)
2090
raise errors.RevisionNotPresent(missing_keys.pop(), self)
2092
def find_ancestry(self, keys):
2093
"""See CombinedGraphIndex.find_ancestry"""
2094
return self._graph_index.find_ancestry(keys, 0)
2096
def get_parent_map(self, keys):
2097
"""Get a map of the parents of keys.
2099
:param keys: The keys to look up parents for.
2100
:return: A mapping from keys to parents. Absent keys are absent from
2104
nodes = self._get_entries(keys)
2108
result[node[1]] = node[3][0]
2111
result[node[1]] = None
2114
def get_missing_parents(self):
2115
"""Return the keys of missing parents."""
2116
# Copied from _KnitGraphIndex.get_missing_parents
2117
# We may have false positives, so filter those out.
2118
self._key_dependencies.satisfy_refs_for_keys(
2119
self.get_parent_map(self._key_dependencies.get_unsatisfied_refs()))
2120
return frozenset(self._key_dependencies.get_unsatisfied_refs())
2122
def get_build_details(self, keys):
2123
"""Get the various build details for keys.
2125
Ghosts are omitted from the result.
2127
:param keys: An iterable of keys.
2128
:return: A dict of key:
2129
(index_memo, compression_parent, parents, record_details).
2131
* index_memo: opaque structure to pass to read_records to extract
2133
* compression_parent: Content that this record is built upon, may
2135
* parents: Logical parents of this node
2136
* record_details: extra information about the content which needs
2137
to be passed to Factory.parse_record
2141
entries = self._get_entries(keys)
2142
for entry in entries:
2144
if not self._parents:
2147
parents = entry[3][0]
2148
details = _GCBuildDetails(parents, self._node_to_position(entry))
2149
result[key] = details
2153
"""Get all the keys in the collection.
2155
The keys are not ordered.
2158
return [node[1] for node in self._graph_index.iter_all_entries()]
2160
def _node_to_position(self, node):
2161
"""Convert an index value to position details."""
2162
bits = node[2].split(b' ')
2163
# It would be nice not to read the entire gzip.
2164
# start and stop are put into _int_cache because they are very common.
2165
# They define the 'group' that an entry is in, and many groups can have
2166
# thousands of objects.
2167
# Branching Launchpad, for example, saves ~600k integers, at 12 bytes
2168
# each, or about 7MB. Note that it might be even more when you consider
2169
# how PyInt is allocated in separate slabs. And you can't return a slab
2170
# to the OS if even 1 int on it is in use. Note though that Python uses
2171
# a LIFO when re-using PyInt slots, which might cause more
2173
start = int(bits[0])
2174
start = self._int_cache.setdefault(start, start)
2176
stop = self._int_cache.setdefault(stop, stop)
2177
basis_end = int(bits[2])
2178
delta_end = int(bits[3])
2179
# We can't use StaticTuple here, because node[0] is a BTreeGraphIndex
2181
return (node[0], start, stop, basis_end, delta_end)
2183
def scan_unvalidated_index(self, graph_index):
2184
"""Inform this _GCGraphIndex that there is an unvalidated index.
2186
This allows this _GCGraphIndex to keep track of any missing
2187
compression parents we may want to have filled in to make those
2188
indices valid. It also allows _GCGraphIndex to track any new keys.
2190
:param graph_index: A GraphIndex
2192
key_dependencies = self._key_dependencies
2193
if key_dependencies is None:
2195
for node in graph_index.iter_all_entries():
2196
# Add parent refs from graph_index (and discard parent refs
2197
# that the graph_index has).
2198
key_dependencies.add_references(node[1], node[3][0])
2201
from ._groupcompress_py import (
2203
apply_delta_to_source,
2206
decode_copy_instruction,
2210
from ._groupcompress_pyx import (
2212
apply_delta_to_source,
2217
GroupCompressor = PyrexGroupCompressor
2218
except ImportError as e:
2219
osutils.failed_to_load_extension(e)
2220
GroupCompressor = PythonGroupCompressor
added
removed
breezy/bzr/groupcompress.py
