1
# Copyright (C) 2005, 2006 by Canonical Ltd
2
# Written by Martin Pool.
3
# Modified by Johan Rydberg <jrydberg@gnu.org>
4
# Modified by Robert Collins <robert.collins@canonical.com>
6
# This program is free software; you can redistribute it and/or modify
7
# it under the terms of the GNU General Public License as published by
8
# the Free Software Foundation; either version 2 of the License, or
9
# (at your option) any later version.
11
# This program is distributed in the hope that it will be useful,
12
# but WITHOUT ANY WARRANTY; without even the implied warranty of
13
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14
# GNU General Public License for more details.
16
# You should have received a copy of the GNU General Public License
17
# along with this program; if not, write to the Free Software
18
# Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20
"""Knit versionedfile implementation.
22
A knit is a versioned file implementation that supports efficient append only
26
lifeless: the data file is made up of "delta records". each delta record has a delta header
27
that contains; (1) a version id, (2) the size of the delta (in lines), and (3) the digest of
28
the -expanded data- (ie, the delta applied to the parent). the delta also ends with a
29
end-marker; simply "end VERSION"
31
delta can be line or full contents.a
32
... the 8's there are the index number of the annotation.
33
version robertc@robertcollins.net-20051003014215-ee2990904cc4c7ad 7 c7d23b2a5bd6ca00e8e266cec0ec228158ee9f9e
37
8 e.set('executable', 'yes')
39
8 if elt.get('executable') == 'yes':
40
8 ie.executable = True
41
end robertc@robertcollins.net-20051003014215-ee2990904cc4c7ad
45
09:33 < jrydberg> lifeless: each index is made up of a tuple of; version id, options, position, size, parents
46
09:33 < jrydberg> lifeless: the parents are currently dictionary compressed
47
09:33 < jrydberg> lifeless: (meaning it currently does not support ghosts)
48
09:33 < lifeless> right
49
09:33 < jrydberg> lifeless: the position and size is the range in the data file
52
so the index sequence is the dictionary compressed sequence number used
53
in the deltas to provide line annotation
58
# 10:16 < lifeless> make partial index writes safe
59
# 10:16 < lifeless> implement 'knit.check()' like weave.check()
60
# 10:17 < lifeless> record known ghosts so we can detect when they are filled in rather than the current 'reweave
62
# move sha1 out of the content so that join is faster at verifying parents
63
# record content length ?
67
from cStringIO import StringIO
69
from itertools import izip, chain
74
import bzrlib.errors as errors
75
from bzrlib.errors import FileExists, NoSuchFile, KnitError, \
76
InvalidRevisionId, KnitCorrupt, KnitHeaderError, \
77
RevisionNotPresent, RevisionAlreadyPresent
78
from bzrlib.tuned_gzip import *
79
from bzrlib.trace import mutter
80
from bzrlib.osutils import contains_whitespace, contains_linebreaks, \
82
from bzrlib.versionedfile import VersionedFile, InterVersionedFile
83
from bzrlib.tsort import topo_sort
86
# TODO: Split out code specific to this format into an associated object.
88
# TODO: Can we put in some kind of value to check that the index and data
89
# files belong together?
91
# TODO: accomodate binaries, perhaps by storing a byte count
93
# TODO: function to check whole file
95
# TODO: atomically append data, then measure backwards from the cursor
96
# position after writing to work out where it was located. we may need to
97
# bypass python file buffering.
100
INDEX_SUFFIX = '.kndx'
103
class KnitContent(object):
104
"""Content of a knit version to which deltas can be applied."""
106
def __init__(self, lines):
109
def annotate_iter(self):
110
"""Yield tuples of (origin, text) for each content line."""
111
for origin, text in self._lines:
115
"""Return a list of (origin, text) tuples."""
116
return list(self.annotate_iter())
118
def line_delta_iter(self, new_lines):
119
"""Generate line-based delta from this content to new_lines."""
120
new_texts = [text for origin, text in new_lines._lines]
121
old_texts = [text for origin, text in self._lines]
122
s = SequenceMatcher(None, old_texts, new_texts)
123
for op in s.get_opcodes():
126
# ofrom oto length data
127
yield (op[1], op[2], op[4]-op[3], new_lines._lines[op[3]:op[4]])
129
def line_delta(self, new_lines):
130
return list(self.line_delta_iter(new_lines))
133
return [text for origin, text in self._lines]
136
class _KnitFactory(object):
137
"""Base factory for creating content objects."""
139
def make(self, lines, version):
140
num_lines = len(lines)
141
return KnitContent(zip([version] * num_lines, lines))
144
class KnitAnnotateFactory(_KnitFactory):
145
"""Factory for creating annotated Content objects."""
149
def parse_fulltext(self, content, version):
150
"""Convert fulltext to internal representation
152
fulltext content is of the format
153
revid(utf8) plaintext\n
154
internal representation is of the format:
159
origin, text = line.split(' ', 1)
160
lines.append((origin.decode('utf-8'), text))
161
return KnitContent(lines)
163
def parse_line_delta_iter(self, lines):
164
for result_item in self.parse_line_delta[lines]:
167
def parse_line_delta(self, lines, version):
168
"""Convert a line based delta into internal representation.
170
line delta is in the form of:
171
intstart intend intcount
173
revid(utf8) newline\n
174
internal represnetation is
175
(start, end, count, [1..count tuples (revid, newline)])
180
# walk through the lines parsing.
182
start, end, count = [int(n) for n in header.split(',')]
186
origin, text = next().split(' ', 1)
188
contents.append((origin.decode('utf-8'), text))
189
result.append((start, end, count, contents))
192
def lower_fulltext(self, content):
193
"""convert a fulltext content record into a serializable form.
195
see parse_fulltext which this inverts.
197
return ['%s %s' % (o.encode('utf-8'), t) for o, t in content._lines]
199
def lower_line_delta(self, delta):
200
"""convert a delta into a serializable form.
202
See parse_line_delta which this inverts.
205
for start, end, c, lines in delta:
206
out.append('%d,%d,%d\n' % (start, end, c))
207
for origin, text in lines:
208
out.append('%s %s' % (origin.encode('utf-8'), text))
212
class KnitPlainFactory(_KnitFactory):
213
"""Factory for creating plain Content objects."""
217
def parse_fulltext(self, content, version):
218
"""This parses an unannotated fulltext.
220
Note that this is not a noop - the internal representation
221
has (versionid, line) - its just a constant versionid.
223
return self.make(content, version)
225
def parse_line_delta_iter(self, lines, version):
227
header = lines.pop(0)
228
start, end, c = [int(n) for n in header.split(',')]
229
yield start, end, c, zip([version] * c, lines[:c])
232
def parse_line_delta(self, lines, version):
233
return list(self.parse_line_delta_iter(lines, version))
235
def lower_fulltext(self, content):
236
return content.text()
238
def lower_line_delta(self, delta):
240
for start, end, c, lines in delta:
241
out.append('%d,%d,%d\n' % (start, end, c))
242
out.extend([text for origin, text in lines])
246
def make_empty_knit(transport, relpath):
247
"""Construct a empty knit at the specified location."""
248
k = KnitVersionedFile(transport, relpath, 'w', KnitPlainFactory)
252
class KnitVersionedFile(VersionedFile):
253
"""Weave-like structure with faster random access.
255
A knit stores a number of texts and a summary of the relationships
256
between them. Texts are identified by a string version-id. Texts
257
are normally stored and retrieved as a series of lines, but can
258
also be passed as single strings.
260
Lines are stored with the trailing newline (if any) included, to
261
avoid special cases for files with no final newline. Lines are
262
composed of 8-bit characters, not unicode. The combination of
263
these approaches should mean any 'binary' file can be safely
264
stored and retrieved.
267
def __init__(self, relpath, transport, file_mode=None, access_mode=None, factory=None,
268
basis_knit=None, delta=True, create=False):
269
"""Construct a knit at location specified by relpath.
271
:param create: If not True, only open an existing knit.
273
if access_mode is None:
275
super(KnitVersionedFile, self).__init__(access_mode)
276
assert access_mode in ('r', 'w'), "invalid mode specified %r" % access_mode
277
assert not basis_knit or isinstance(basis_knit, KnitVersionedFile), \
280
self.transport = transport
281
self.filename = relpath
282
self.basis_knit = basis_knit
283
self.factory = factory or KnitAnnotateFactory()
284
self.writable = (access_mode == 'w')
287
self._index = _KnitIndex(transport, relpath + INDEX_SUFFIX,
288
access_mode, create=create, file_mode=file_mode)
289
self._data = _KnitData(transport, relpath + DATA_SUFFIX,
290
access_mode, create=create and not len(self), file_mode=file_mode)
292
def _add_delta(self, version_id, parents, delta_parent, sha1, noeol, delta):
293
"""See VersionedFile._add_delta()."""
294
self._check_add(version_id, []) # should we check the lines ?
295
self._check_versions_present(parents)
299
for parent in parents:
300
if not self.has_version(parent):
301
ghosts.append(parent)
303
present_parents.append(parent)
305
if delta_parent is None:
306
# reconstitute as full text.
307
assert len(delta) == 1 or len(delta) == 0
309
assert delta[0][0] == 0
310
assert delta[0][1] == 0, delta[0][1]
311
return super(KnitVersionedFile, self)._add_delta(version_id,
322
options.append('no-eol')
324
if delta_parent is not None:
325
# determine the current delta chain length.
326
# To speed the extract of texts the delta chain is limited
327
# to a fixed number of deltas. This should minimize both
328
# I/O and the time spend applying deltas.
330
delta_parents = [delta_parent]
332
parent = delta_parents[0]
333
method = self._index.get_method(parent)
334
if method == 'fulltext':
336
delta_parents = self._index.get_parents(parent)
338
if method == 'line-delta':
339
# did not find a fulltext in the delta limit.
340
# just do a normal insertion.
341
return super(KnitVersionedFile, self)._add_delta(version_id,
348
options.append('line-delta')
349
store_lines = self.factory.lower_line_delta(delta)
351
where, size = self._data.add_record(version_id, digest, store_lines)
352
self._index.add_version(version_id, options, where, size, parents)
354
def clear_cache(self):
355
"""Clear the data cache only."""
356
self._data.clear_cache()
358
def copy_to(self, name, transport):
359
"""See VersionedFile.copy_to()."""
360
# copy the current index to a temp index to avoid racing with local
362
transport.put(name + INDEX_SUFFIX + '.tmp', self.transport.get(self._index._filename),)
364
transport.put(name + DATA_SUFFIX, self._data._open_file())
365
# rename the copied index into place
366
transport.rename(name + INDEX_SUFFIX + '.tmp', name + INDEX_SUFFIX)
368
def create_empty(self, name, transport, mode=None):
369
return KnitVersionedFile(name, transport, factory=self.factory, delta=self.delta, create=True)
371
def _fix_parents(self, version, new_parents):
372
"""Fix the parents list for version.
374
This is done by appending a new version to the index
375
with identical data except for the parents list.
376
the parents list must be a superset of the current
379
current_values = self._index._cache[version]
380
assert set(current_values[4]).difference(set(new_parents)) == set()
381
self._index.add_version(version,
387
def get_delta(self, version_id):
388
"""Get a delta for constructing version from some other version."""
389
if not self.has_version(version_id):
390
raise RevisionNotPresent(version_id, self.filename)
392
parents = self.get_parents(version_id)
397
data_pos, data_size = self._index.get_position(version_id)
398
data, sha1 = self._data.read_records(((version_id, data_pos, data_size),))[version_id]
399
version_idx = self._index.lookup(version_id)
400
noeol = 'no-eol' in self._index.get_options(version_id)
401
if 'fulltext' == self._index.get_method(version_id):
402
new_content = self.factory.parse_fulltext(data, version_idx)
403
if parent is not None:
404
reference_content = self._get_content(parent)
405
old_texts = reference_content.text()
408
new_texts = new_content.text()
409
delta_seq = SequenceMatcher(None, old_texts, new_texts)
410
return parent, sha1, noeol, self._make_line_delta(delta_seq, new_content)
412
delta = self.factory.parse_line_delta(data, version_idx)
413
return parent, sha1, noeol, delta
415
def get_graph_with_ghosts(self):
416
"""See VersionedFile.get_graph_with_ghosts()."""
417
graph_items = self._index.get_graph()
418
return dict(graph_items)
420
def get_sha1(self, version_id):
421
"""See VersionedFile.get_sha1()."""
422
components = self._get_components(version_id)
423
return components[-1][-1][-1]
427
"""See VersionedFile.get_suffixes()."""
428
return [DATA_SUFFIX, INDEX_SUFFIX]
430
def has_ghost(self, version_id):
431
"""True if there is a ghost reference in the file to version_id."""
433
if self.has_version(version_id):
435
# optimisable if needed by memoising the _ghosts set.
436
items = self._index.get_graph()
437
for node, parents in items:
438
for parent in parents:
439
if parent not in self._index._cache:
440
if parent == version_id:
445
"""See VersionedFile.versions."""
446
return self._index.get_versions()
448
def has_version(self, version_id):
449
"""See VersionedFile.has_version."""
450
return self._index.has_version(version_id)
452
__contains__ = has_version
454
def _merge_annotations(self, content, parents, parent_texts={},
455
delta=None, annotated=None):
456
"""Merge annotations for content. This is done by comparing
457
the annotations based on changed to the text.
461
for parent_id in parents:
462
merge_content = self._get_content(parent_id, parent_texts)
463
seq = SequenceMatcher(None, merge_content.text(), content.text())
464
if delta_seq is None:
465
# setup a delta seq to reuse.
467
for i, j, n in seq.get_matching_blocks():
470
# this appears to copy (origin, text) pairs across to the new
471
# content for any line that matches the last-checked parent.
472
# FIXME: save the sequence control data for delta compression
473
# against the most relevant parent rather than rediffing.
474
content._lines[j:j+n] = merge_content._lines[i:i+n]
477
reference_content = self._get_content(parents[0], parent_texts)
478
new_texts = content.text()
479
old_texts = reference_content.text()
480
delta_seq = SequenceMatcher(None, old_texts, new_texts)
481
return self._make_line_delta(delta_seq, content)
483
def _make_line_delta(self, delta_seq, new_content):
484
"""Generate a line delta from delta_seq and new_content."""
486
for op in delta_seq.get_opcodes():
489
diff_hunks.append((op[1], op[2], op[4]-op[3], new_content._lines[op[3]:op[4]]))
492
def _get_components(self, version_id):
493
"""Return a list of (version_id, method, data) tuples that
494
makes up version specified by version_id of the knit.
496
The components should be applied in the order of the returned
499
The basis knit will be used to the largest extent possible
500
since it is assumed that accesses to it is faster.
503
# 4168 calls in 14912, 2289 internal
504
# 4168 in 9711 to read_records
505
# 52554 in 1250 to get_parents
506
# 170166 in 865 to list.append
508
# needed_revisions holds a list of (method, version_id) of
509
# versions that is needed to be fetched to construct the final
510
# version of the file.
512
# basis_revisions is a list of versions that needs to be
513
# fetched but exists in the basis knit.
515
basis = self.basis_knit
522
if basis and basis._index.has_version(cursor):
524
basis_versions.append(cursor)
525
method = picked_knit._index.get_method(cursor)
526
needed_versions.append((method, cursor))
527
if method == 'fulltext':
529
cursor = picked_knit.get_parents(cursor)[0]
534
for comp_id in basis_versions:
535
data_pos, data_size = basis._index.get_data_position(comp_id)
536
records.append((piece_id, data_pos, data_size))
537
components.update(basis._data.read_records(records))
540
for comp_id in [vid for method, vid in needed_versions
541
if vid not in basis_versions]:
542
data_pos, data_size = self._index.get_position(comp_id)
543
records.append((comp_id, data_pos, data_size))
544
components.update(self._data.read_records(records))
546
# get_data_records returns a mapping with the version id as
547
# index and the value as data. The order the components need
548
# to be applied is held by needed_versions (reversed).
550
for method, comp_id in reversed(needed_versions):
551
out.append((comp_id, method, components[comp_id]))
555
def _get_content(self, version_id, parent_texts={}):
556
"""Returns a content object that makes up the specified
558
if not self.has_version(version_id):
559
raise RevisionNotPresent(version_id, self.filename)
561
cached_version = parent_texts.get(version_id, None)
562
if cached_version is not None:
563
return cached_version
565
if self.basis_knit and version_id in self.basis_knit:
566
return self.basis_knit._get_content(version_id)
569
components = self._get_components(version_id)
570
for component_id, method, (data, digest) in components:
571
version_idx = self._index.lookup(component_id)
572
if method == 'fulltext':
573
assert content is None
574
content = self.factory.parse_fulltext(data, version_idx)
575
elif method == 'line-delta':
576
delta = self.factory.parse_line_delta(data, version_idx)
577
content._lines = self._apply_delta(content._lines, delta)
579
if 'no-eol' in self._index.get_options(version_id):
580
line = content._lines[-1][1].rstrip('\n')
581
content._lines[-1] = (content._lines[-1][0], line)
583
# digest here is the digest from the last applied component.
584
if sha_strings(content.text()) != digest:
585
import pdb;pdb.set_trace()
586
raise KnitCorrupt(self.filename, 'sha-1 does not match %s' % version_id)
590
def _check_versions_present(self, version_ids):
591
"""Check that all specified versions are present."""
592
version_ids = set(version_ids)
593
for r in list(version_ids):
594
if self._index.has_version(r):
595
version_ids.remove(r)
597
raise RevisionNotPresent(list(version_ids)[0], self.filename)
599
def _add_lines_with_ghosts(self, version_id, parents, lines, parent_texts):
600
"""See VersionedFile.add_lines_with_ghosts()."""
601
self._check_add(version_id, lines)
602
return self._add(version_id, lines[:], parents, self.delta, parent_texts)
604
def _add_lines(self, version_id, parents, lines, parent_texts):
605
"""See VersionedFile.add_lines."""
606
self._check_add(version_id, lines)
607
self._check_versions_present(parents)
608
return self._add(version_id, lines[:], parents, self.delta, parent_texts)
610
def _check_add(self, version_id, lines):
611
"""check that version_id and lines are safe to add."""
612
assert self.writable, "knit is not opened for write"
613
### FIXME escape. RBC 20060228
614
if contains_whitespace(version_id):
615
raise InvalidRevisionId(version_id)
616
if self.has_version(version_id):
617
raise RevisionAlreadyPresent(version_id, self.filename)
618
self._check_lines_not_unicode(lines)
619
self._check_lines_are_lines(lines)
621
def _add(self, version_id, lines, parents, delta, parent_texts):
622
"""Add a set of lines on top of version specified by parents.
624
If delta is true, compress the text as a line-delta against
627
Any versions not present will be converted into ghosts.
629
# 461 0 6546.0390 43.9100 bzrlib.knit:489(_add)
630
# +400 0 889.4890 418.9790 +bzrlib.knit:192(lower_fulltext)
631
# +461 0 1364.8070 108.8030 +bzrlib.knit:996(add_record)
632
# +461 0 193.3940 41.5720 +bzrlib.knit:898(add_version)
633
# +461 0 134.0590 18.3810 +bzrlib.osutils:361(sha_strings)
634
# +461 0 36.3420 15.4540 +bzrlib.knit:146(make)
635
# +1383 0 8.0370 8.0370 +<len>
636
# +61 0 13.5770 7.9190 +bzrlib.knit:199(lower_line_delta)
637
# +61 0 963.3470 7.8740 +bzrlib.knit:427(_get_content)
638
# +61 0 973.9950 5.2950 +bzrlib.knit:136(line_delta)
639
# +61 0 1918.1800 5.2640 +bzrlib.knit:359(_merge_annotations)
643
if parent_texts is None:
645
for parent in parents:
646
if not self.has_version(parent):
647
ghosts.append(parent)
649
present_parents.append(parent)
651
if delta and not len(present_parents):
654
digest = sha_strings(lines)
657
if lines[-1][-1] != '\n':
658
options.append('no-eol')
659
lines[-1] = lines[-1] + '\n'
661
if len(present_parents) and delta:
662
# To speed the extract of texts the delta chain is limited
663
# to a fixed number of deltas. This should minimize both
664
# I/O and the time spend applying deltas.
666
delta_parents = present_parents
668
parent = delta_parents[0]
669
method = self._index.get_method(parent)
670
if method == 'fulltext':
672
delta_parents = self._index.get_parents(parent)
674
if method == 'line-delta':
677
lines = self.factory.make(lines, version_id)
678
if delta or (self.factory.annotated and len(present_parents) > 0):
679
# Merge annotations from parent texts if so is needed.
680
delta_hunks = self._merge_annotations(lines, present_parents, parent_texts,
681
delta, self.factory.annotated)
684
options.append('line-delta')
685
store_lines = self.factory.lower_line_delta(delta_hunks)
687
options.append('fulltext')
688
store_lines = self.factory.lower_fulltext(lines)
690
where, size = self._data.add_record(version_id, digest, store_lines)
691
self._index.add_version(version_id, options, where, size, parents)
694
def check(self, progress_bar=None):
695
"""See VersionedFile.check()."""
697
def _clone_text(self, new_version_id, old_version_id, parents):
698
"""See VersionedFile.clone_text()."""
699
# FIXME RBC 20060228 make fast by only inserting an index with null delta.
700
self.add_lines(new_version_id, parents, self.get_lines(old_version_id))
702
def get_lines(self, version_id):
703
"""See VersionedFile.get_lines()."""
704
return self._get_content(version_id).text()
706
def iter_lines_added_or_present_in_versions(self, version_ids=None):
707
"""See VersionedFile.iter_lines_added_or_present_in_versions()."""
708
if version_ids is None:
709
version_ids = self.versions()
710
# we dont care about inclusions, the caller cares.
711
# but we need to setup a list of records to visit.
712
# we need version_id, position, length
713
version_id_records = []
714
requested_versions = list(version_ids)
715
# filter for available versions
716
for version_id in requested_versions:
717
if not self.has_version(version_id):
718
raise RevisionNotPresent(version_id, self.filename)
719
# get a in-component-order queue:
721
for version_id in self.versions():
722
if version_id in requested_versions:
723
version_ids.append(version_id)
724
data_pos, length = self._index.get_position(version_id)
725
version_id_records.append((version_id, data_pos, length))
727
pb = bzrlib.ui.ui_factory.nested_progress_bar()
729
total = len(version_id_records)
731
pb.update('Walking content.', count, total)
732
for version_id, data, sha_value in \
733
self._data.read_records_iter(version_id_records):
734
pb.update('Walking content.', count, total)
735
method = self._index.get_method(version_id)
736
version_idx = self._index.lookup(version_id)
737
assert method in ('fulltext', 'line-delta')
738
if method == 'fulltext':
739
content = self.factory.parse_fulltext(data, version_idx)
740
for line in content.text():
743
delta = self.factory.parse_line_delta(data, version_idx)
744
for start, end, count, lines in delta:
745
for origin, line in lines:
748
pb.update('Walking content.', total, total)
751
pb.update('Walking content.', total, total)
755
def num_versions(self):
756
"""See VersionedFile.num_versions()."""
757
return self._index.num_versions()
759
__len__ = num_versions
761
def annotate_iter(self, version_id):
762
"""See VersionedFile.annotate_iter."""
763
content = self._get_content(version_id)
764
for origin, text in content.annotate_iter():
767
def get_parents(self, version_id):
768
"""See VersionedFile.get_parents."""
771
# 52554 calls in 1264 872 internal down from 3674
773
return self._index.get_parents(version_id)
775
raise RevisionNotPresent(version_id, self.filename)
777
def get_parents_with_ghosts(self, version_id):
778
"""See VersionedFile.get_parents."""
780
return self._index.get_parents_with_ghosts(version_id)
782
raise RevisionNotPresent(version_id, self.filename)
784
def get_ancestry(self, versions):
785
"""See VersionedFile.get_ancestry."""
786
if isinstance(versions, basestring):
787
versions = [versions]
790
self._check_versions_present(versions)
791
return self._index.get_ancestry(versions)
793
def get_ancestry_with_ghosts(self, versions):
794
"""See VersionedFile.get_ancestry_with_ghosts."""
795
if isinstance(versions, basestring):
796
versions = [versions]
799
self._check_versions_present(versions)
800
return self._index.get_ancestry_with_ghosts(versions)
802
#@deprecated_method(zero_eight)
803
def walk(self, version_ids):
804
"""See VersionedFile.walk."""
805
# We take the short path here, and extract all relevant texts
806
# and put them in a weave and let that do all the work. Far
807
# from optimal, but is much simpler.
808
# FIXME RB 20060228 this really is inefficient!
809
from bzrlib.weave import Weave
811
w = Weave(self.filename)
812
ancestry = self.get_ancestry(version_ids)
813
sorted_graph = topo_sort(self._index.get_graph())
814
version_list = [vid for vid in sorted_graph if vid in ancestry]
816
for version_id in version_list:
817
lines = self.get_lines(version_id)
818
w.add_lines(version_id, self.get_parents(version_id), lines)
820
for lineno, insert_id, dset, line in w.walk(version_ids):
821
yield lineno, insert_id, dset, line
824
class _KnitComponentFile(object):
825
"""One of the files used to implement a knit database"""
827
def __init__(self, transport, filename, mode, file_mode=None):
828
self._transport = transport
829
self._filename = filename
831
self._file_mode=file_mode
833
def write_header(self):
834
if self._transport.append(self._filename, StringIO(self.HEADER),
835
mode=self._file_mode):
836
raise KnitCorrupt(self._filename, 'misaligned after writing header')
838
def check_header(self, fp):
840
if line != self.HEADER:
841
raise KnitHeaderError(badline=line)
844
"""Commit is a nop."""
847
return '%s(%s)' % (self.__class__.__name__, self._filename)
850
class _KnitIndex(_KnitComponentFile):
851
"""Manages knit index file.
853
The index is already kept in memory and read on startup, to enable
854
fast lookups of revision information. The cursor of the index
855
file is always pointing to the end, making it easy to append
858
_cache is a cache for fast mapping from version id to a Index
861
_history is a cache for fast mapping from indexes to version ids.
863
The index data format is dictionary compressed when it comes to
864
parent references; a index entry may only have parents that with a
865
lover index number. As a result, the index is topological sorted.
867
Duplicate entries may be written to the index for a single version id
868
if this is done then the latter one completely replaces the former:
869
this allows updates to correct version and parent information.
870
Note that the two entries may share the delta, and that successive
871
annotations and references MUST point to the first entry.
873
The index file on disc contains a header, followed by one line per knit
874
record. The same revision can be present in an index file more than once.
875
The first occurence gets assigned a sequence number starting from 0.
877
The format of a single line is
878
REVISION_ID FLAGS BYTE_OFFSET LENGTH( PARENT_ID|PARENT_SEQUENCE_ID)* :\n
879
REVISION_ID is a utf8-encoded revision id
880
FLAGS is a comma separated list of flags about the record. Values include
881
no-eol, line-delta, fulltext.
882
BYTE_OFFSET is the ascii representation of the byte offset in the data file
883
that the the compressed data starts at.
884
LENGTH is the ascii representation of the length of the data file.
885
PARENT_ID a utf-8 revision id prefixed by a '.' that is a parent of
887
PARENT_SEQUENCE_ID the ascii representation of the sequence number of a
888
revision id already in the knit that is a parent of REVISION_ID.
889
The ' :' marker is the end of record marker.
892
when a write is interrupted to the index file, it will result in a line that
893
does not end in ' :'. If the ' :' is not present at the end of a line, or at
894
the end of the file, then the record that is missing it will be ignored by
897
When writing new records to the index file, the data is preceeded by '\n'
898
to ensure that records always start on new lines even if the last write was
899
interrupted. As a result its normal for the last line in the index to be
900
missing a trailing newline. One can be added with no harmful effects.
903
HEADER = "# bzr knit index 8\n"
905
# speed of knit parsing went from 280 ms to 280 ms with slots addition.
906
# __slots__ = ['_cache', '_history', '_transport', '_filename']
908
def _cache_version(self, version_id, options, pos, size, parents):
909
"""Cache a version record in the history array and index cache.
911
This is inlined into __init__ for performance. KEEP IN SYNC.
912
(It saves 60ms, 25% of the __init__ overhead on local 4000 record
915
# only want the _history index to reference the 1st index entry
917
if version_id not in self._cache:
918
index = len(self._history)
919
self._history.append(version_id)
921
index = self._cache[version_id][5]
922
self._cache[version_id] = (version_id,
929
def __init__(self, transport, filename, mode, create=False, file_mode=None):
930
_KnitComponentFile.__init__(self, transport, filename, mode, file_mode)
932
# position in _history is the 'official' index for a revision
933
# but the values may have come from a newer entry.
934
# so - wc -l of a knit index is != the number of uniqe names
937
pb = bzrlib.ui.ui_factory.nested_progress_bar()
942
pb.update('read knit index', count, total)
943
fp = self._transport.get(self._filename)
944
self.check_header(fp)
945
# readlines reads the whole file at once:
946
# bad for transports like http, good for local disk
947
# we save 60 ms doing this one change (
948
# from calling readline each time to calling
950
# probably what we want for nice behaviour on
951
# http is a incremental readlines that yields, or
952
# a check for local vs non local indexes,
953
for l in fp.readlines():
955
if len(rec) < 5 or rec[-1] != ':':
957
# FIXME: in the future we should determine if its a
958
# short write - and ignore it
959
# or a different failure, and raise. RBC 20060407
963
#pb.update('read knit index', count, total)
964
# See self._parse_parents
966
for value in rec[4:-1]:
968
# uncompressed reference
969
parents.append(value[1:])
971
# this is 15/4000ms faster than isinstance,
973
# this function is called thousands of times a
974
# second so small variations add up.
975
assert value.__class__ is str
976
parents.append(self._history[int(value)])
977
# end self._parse_parents
978
# self._cache_version(rec[0],
983
# --- self._cache_version
984
# only want the _history index to reference the 1st
985
# index entry for version_id
987
if version_id not in self._cache:
988
index = len(self._history)
989
self._history.append(version_id)
991
index = self._cache[version_id][5]
992
self._cache[version_id] = (version_id,
998
# --- self._cache_version
999
except NoSuchFile, e:
1000
if mode != 'w' or not create:
1004
pb.update('read knit index', total, total)
1007
def _parse_parents(self, compressed_parents):
1008
"""convert a list of string parent values into version ids.
1010
ints are looked up in the index.
1011
.FOO values are ghosts and converted in to FOO.
1013
NOTE: the function is retained here for clarity, and for possible
1014
use in partial index reads. However bulk processing now has
1015
it inlined in __init__ for inner-loop optimisation.
1018
for value in compressed_parents:
1019
if value[-1] == '.':
1020
# uncompressed reference
1021
result.append(value[1:])
1023
# this is 15/4000ms faster than isinstance,
1024
# this function is called thousands of times a
1025
# second so small variations add up.
1026
assert value.__class__ is str
1027
result.append(self._history[int(value)])
1030
def get_graph(self):
1032
for version_id, index in self._cache.iteritems():
1033
graph.append((version_id, index[4]))
1036
def get_ancestry(self, versions):
1037
"""See VersionedFile.get_ancestry."""
1038
# get a graph of all the mentioned versions:
1040
pending = set(versions)
1042
version = pending.pop()
1043
parents = self._cache[version][4]
1044
# got the parents ok
1046
parents = [parent for parent in parents if parent in self._cache]
1047
for parent in parents:
1048
# if not completed and not a ghost
1049
if parent not in graph:
1051
graph[version] = parents
1052
return topo_sort(graph.items())
1054
def get_ancestry_with_ghosts(self, versions):
1055
"""See VersionedFile.get_ancestry_with_ghosts."""
1056
# get a graph of all the mentioned versions:
1058
pending = set(versions)
1060
version = pending.pop()
1062
parents = self._cache[version][4]
1068
# got the parents ok
1069
for parent in parents:
1070
if parent not in graph:
1072
graph[version] = parents
1073
return topo_sort(graph.items())
1075
def num_versions(self):
1076
return len(self._history)
1078
__len__ = num_versions
1080
def get_versions(self):
1081
return self._history
1083
def idx_to_name(self, idx):
1084
return self._history[idx]
1086
def lookup(self, version_id):
1087
assert version_id in self._cache
1088
return self._cache[version_id][5]
1090
def _version_list_to_index(self, versions):
1092
for version in versions:
1093
if version in self._cache:
1094
# -- inlined lookup() --
1095
result_list.append(str(self._cache[version][5]))
1096
# -- end lookup () --
1098
result_list.append('.' + version.encode('utf-8'))
1099
return ' '.join(result_list)
1101
def add_version(self, version_id, options, pos, size, parents):
1102
"""Add a version record to the index."""
1103
self._cache_version(version_id, options, pos, size, parents)
1105
content = "\n%s %s %s %s %s :" % (version_id.encode('utf-8'),
1109
self._version_list_to_index(parents))
1110
assert isinstance(content, str), 'content must be utf-8 encoded'
1111
self._transport.append(self._filename, StringIO(content))
1113
def has_version(self, version_id):
1114
"""True if the version is in the index."""
1115
return self._cache.has_key(version_id)
1117
def get_position(self, version_id):
1118
"""Return data position and size of specified version."""
1119
return (self._cache[version_id][2], \
1120
self._cache[version_id][3])
1122
def get_method(self, version_id):
1123
"""Return compression method of specified version."""
1124
options = self._cache[version_id][1]
1125
if 'fulltext' in options:
1128
assert 'line-delta' in options
1131
def get_options(self, version_id):
1132
return self._cache[version_id][1]
1134
def get_parents(self, version_id):
1135
"""Return parents of specified version ignoring ghosts."""
1136
return [parent for parent in self._cache[version_id][4]
1137
if parent in self._cache]
1139
def get_parents_with_ghosts(self, version_id):
1140
"""Return parents of specified version wth ghosts."""
1141
return self._cache[version_id][4]
1143
def check_versions_present(self, version_ids):
1144
"""Check that all specified versions are present."""
1145
version_ids = set(version_ids)
1146
for version_id in list(version_ids):
1147
if version_id in self._cache:
1148
version_ids.remove(version_id)
1150
raise RevisionNotPresent(list(version_ids)[0], self.filename)
1153
class _KnitData(_KnitComponentFile):
1154
"""Contents of the knit data file"""
1156
HEADER = "# bzr knit data 8\n"
1158
def __init__(self, transport, filename, mode, create=False, file_mode=None):
1159
_KnitComponentFile.__init__(self, transport, filename, mode)
1161
self._checked = False
1163
self._transport.put(self._filename, StringIO(''), mode=file_mode)
1166
def clear_cache(self):
1167
"""Clear the record cache."""
1170
def _open_file(self):
1171
if self._file is None:
1173
self._file = self._transport.get(self._filename)
1178
def _record_to_data(self, version_id, digest, lines):
1179
"""Convert version_id, digest, lines into a raw data block.
1181
:return: (len, a StringIO instance with the raw data ready to read.)
1184
data_file = GzipFile(None, mode='wb', fileobj=sio)
1185
data_file.writelines(chain(
1186
["version %s %d %s\n" % (version_id.encode('utf-8'),
1190
["end %s\n" % version_id.encode('utf-8')]))
1197
def add_raw_record(self, raw_data):
1198
"""Append a prepared record to the data file."""
1199
assert isinstance(raw_data, str), 'data must be plain bytes'
1200
start_pos = self._transport.append(self._filename, StringIO(raw_data))
1201
return start_pos, len(raw_data)
1203
def add_record(self, version_id, digest, lines):
1204
"""Write new text record to disk. Returns the position in the
1205
file where it was written."""
1206
size, sio = self._record_to_data(version_id, digest, lines)
1208
self._records[version_id] = (digest, lines)
1210
start_pos = self._transport.append(self._filename, sio)
1211
return start_pos, size
1213
def _parse_record_header(self, version_id, raw_data):
1214
"""Parse a record header for consistency.
1216
:return: the header and the decompressor stream.
1217
as (stream, header_record)
1219
df = GzipFile(mode='rb', fileobj=StringIO(raw_data))
1220
rec = df.readline().split()
1222
raise KnitCorrupt(self._filename, 'unexpected number of elements in record header')
1223
if rec[1].decode('utf-8')!= version_id:
1224
raise KnitCorrupt(self._filename,
1225
'unexpected version, wanted %r, got %r' % (
1226
version_id, rec[1]))
1229
def _parse_record(self, version_id, data):
1231
# 4168 calls in 2880 217 internal
1232
# 4168 calls to _parse_record_header in 2121
1233
# 4168 calls to readlines in 330
1234
df, rec = self._parse_record_header(version_id, data)
1235
record_contents = df.readlines()
1236
l = record_contents.pop()
1237
assert len(record_contents) == int(rec[2])
1238
if l.decode('utf-8') != 'end %s\n' % version_id:
1239
raise KnitCorrupt(self._filename, 'unexpected version end line %r, wanted %r'
1242
return record_contents, rec[3]
1244
def read_records_iter_raw(self, records):
1245
"""Read text records from data file and yield raw data.
1247
This unpacks enough of the text record to validate the id is
1248
as expected but thats all.
1250
It will actively recompress currently cached records on the
1251
basis that that is cheaper than I/O activity.
1254
for version_id, pos, size in records:
1255
if version_id not in self._records:
1256
needed_records.append((version_id, pos, size))
1258
# setup an iterator of the external records:
1259
# uses readv so nice and fast we hope.
1260
if len(needed_records):
1261
# grab the disk data needed.
1262
raw_records = self._transport.readv(self._filename,
1263
[(pos, size) for version_id, pos, size in needed_records])
1265
for version_id, pos, size in records:
1266
if version_id in self._records:
1267
# compress a new version
1268
size, sio = self._record_to_data(version_id,
1269
self._records[version_id][0],
1270
self._records[version_id][1])
1271
yield version_id, sio.getvalue()
1273
pos, data = raw_records.next()
1274
# validate the header
1275
df, rec = self._parse_record_header(version_id, data)
1277
yield version_id, data
1280
def read_records_iter(self, records):
1281
"""Read text records from data file and yield result.
1283
Each passed record is a tuple of (version_id, pos, len) and
1284
will be read in the given order. Yields (version_id,
1288
# 60890 calls for 4168 extractions in 5045, 683 internal.
1289
# 4168 calls to readv in 1411
1290
# 4168 calls to parse_record in 2880
1293
for version_id, pos, size in records:
1294
if version_id not in self._records:
1295
needed_records.append((version_id, pos, size))
1297
if len(needed_records):
1298
# We take it that the transport optimizes the fetching as good
1299
# as possible (ie, reads continous ranges.)
1300
response = self._transport.readv(self._filename,
1301
[(pos, size) for version_id, pos, size in needed_records])
1303
for (record_id, pos, size), (pos, data) in izip(iter(needed_records), response):
1304
content, digest = self._parse_record(record_id, data)
1305
self._records[record_id] = (digest, content)
1307
for version_id, pos, size in records:
1308
yield version_id, list(self._records[version_id][1]), self._records[version_id][0]
1310
def read_records(self, records):
1311
"""Read records into a dictionary."""
1313
for record_id, content, digest in self.read_records_iter(records):
1314
components[record_id] = (content, digest)
1318
class InterKnit(InterVersionedFile):
1319
"""Optimised code paths for knit to knit operations."""
1321
_matching_file_factory = KnitVersionedFile
1324
def is_compatible(source, target):
1325
"""Be compatible with knits. """
1327
return (isinstance(source, KnitVersionedFile) and
1328
isinstance(target, KnitVersionedFile))
1329
except AttributeError:
1332
def join(self, pb=None, msg=None, version_ids=None, ignore_missing=False):
1333
"""See InterVersionedFile.join."""
1334
assert isinstance(self.source, KnitVersionedFile)
1335
assert isinstance(self.target, KnitVersionedFile)
1337
if version_ids is None:
1338
version_ids = self.source.versions()
1340
if not ignore_missing:
1341
self.source._check_versions_present(version_ids)
1343
version_ids = set(self.source.versions()).intersection(
1349
pb = bzrlib.ui.ui_factory.nested_progress_bar()
1351
version_ids = list(version_ids)
1352
if None in version_ids:
1353
version_ids.remove(None)
1355
self.source_ancestry = set(self.source.get_ancestry(version_ids))
1356
this_versions = set(self.target._index.get_versions())
1357
needed_versions = self.source_ancestry - this_versions
1358
cross_check_versions = self.source_ancestry.intersection(this_versions)
1359
mismatched_versions = set()
1360
for version in cross_check_versions:
1361
# scan to include needed parents.
1362
n1 = set(self.target.get_parents_with_ghosts(version))
1363
n2 = set(self.source.get_parents_with_ghosts(version))
1365
# FIXME TEST this check for cycles being introduced works
1366
# the logic is we have a cycle if in our graph we are an
1367
# ancestor of any of the n2 revisions.
1373
parent_ancestors = self.source.get_ancestry(parent)
1374
if version in parent_ancestors:
1375
raise errors.GraphCycleError([parent, version])
1376
# ensure this parent will be available later.
1377
new_parents = n2.difference(n1)
1378
needed_versions.update(new_parents.difference(this_versions))
1379
mismatched_versions.add(version)
1381
if not needed_versions and not cross_check_versions:
1383
full_list = topo_sort(self.source.get_graph())
1385
version_list = [i for i in full_list if (not self.target.has_version(i)
1386
and i in needed_versions)]
1390
copy_queue_records = []
1392
for version_id in version_list:
1393
options = self.source._index.get_options(version_id)
1394
parents = self.source._index.get_parents_with_ghosts(version_id)
1395
# check that its will be a consistent copy:
1396
for parent in parents:
1397
# if source has the parent, we must :
1398
# * already have it or
1399
# * have it scheduled already
1400
# otherwise we dont care
1401
assert (self.target.has_version(parent) or
1402
parent in copy_set or
1403
not self.source.has_version(parent))
1404
data_pos, data_size = self.source._index.get_position(version_id)
1405
copy_queue_records.append((version_id, data_pos, data_size))
1406
copy_queue.append((version_id, options, parents))
1407
copy_set.add(version_id)
1409
# data suck the join:
1411
total = len(version_list)
1412
# we want the raw gzip for bulk copying, but the record validated
1413
# just enough to be sure its the right one.
1414
# TODO: consider writev or write combining to reduce
1415
# death of a thousand cuts feeling.
1416
for (version_id, raw_data), \
1417
(version_id2, options, parents) in \
1418
izip(self.source._data.read_records_iter_raw(copy_queue_records),
1420
assert version_id == version_id2, 'logic error, inconsistent results'
1422
pb.update("Joining knit", count, total)
1423
pos, size = self.target._data.add_raw_record(raw_data)
1424
self.target._index.add_version(version_id, options, pos, size, parents)
1426
for version in mismatched_versions:
1427
# FIXME RBC 20060309 is this needed?
1428
n1 = set(self.target.get_parents_with_ghosts(version))
1429
n2 = set(self.source.get_parents_with_ghosts(version))
1430
# write a combined record to our history preserving the current
1431
# parents as first in the list
1432
new_parents = self.target.get_parents_with_ghosts(version) + list(n2.difference(n1))
1433
self.target.fix_parents(version, new_parents)
1439
InterVersionedFile.register_optimiser(InterKnit)
1442
class SequenceMatcher(difflib.SequenceMatcher):
1443
"""Knit tuned sequence matcher.
1445
This is based on profiling of difflib which indicated some improvements
1446
for our usage pattern.
1449
def find_longest_match(self, alo, ahi, blo, bhi):
1450
"""Find longest matching block in a[alo:ahi] and b[blo:bhi].
1452
If isjunk is not defined:
1454
Return (i,j,k) such that a[i:i+k] is equal to b[j:j+k], where
1455
alo <= i <= i+k <= ahi
1456
blo <= j <= j+k <= bhi
1457
and for all (i',j',k') meeting those conditions,
1460
and if i == i', j <= j'
1462
In other words, of all maximal matching blocks, return one that
1463
starts earliest in a, and of all those maximal matching blocks that
1464
start earliest in a, return the one that starts earliest in b.
1466
>>> s = SequenceMatcher(None, " abcd", "abcd abcd")
1467
>>> s.find_longest_match(0, 5, 0, 9)
1470
If isjunk is defined, first the longest matching block is
1471
determined as above, but with the additional restriction that no
1472
junk element appears in the block. Then that block is extended as
1473
far as possible by matching (only) junk elements on both sides. So
1474
the resulting block never matches on junk except as identical junk
1475
happens to be adjacent to an "interesting" match.
1477
Here's the same example as before, but considering blanks to be
1478
junk. That prevents " abcd" from matching the " abcd" at the tail
1479
end of the second sequence directly. Instead only the "abcd" can
1480
match, and matches the leftmost "abcd" in the second sequence:
1482
>>> s = SequenceMatcher(lambda x: x==" ", " abcd", "abcd abcd")
1483
>>> s.find_longest_match(0, 5, 0, 9)
1486
If no blocks match, return (alo, blo, 0).
1488
>>> s = SequenceMatcher(None, "ab", "c")
1489
>>> s.find_longest_match(0, 2, 0, 1)
1493
# CAUTION: stripping common prefix or suffix would be incorrect.
1497
# Longest matching block is "ab", but if common prefix is
1498
# stripped, it's "a" (tied with "b"). UNIX(tm) diff does so
1499
# strip, so ends up claiming that ab is changed to acab by
1500
# inserting "ca" in the middle. That's minimal but unintuitive:
1501
# "it's obvious" that someone inserted "ac" at the front.
1502
# Windiff ends up at the same place as diff, but by pairing up
1503
# the unique 'b's and then matching the first two 'a's.
1505
a, b, b2j, isbjunk = self.a, self.b, self.b2j, self.isbjunk
1506
besti, bestj, bestsize = alo, blo, 0
1507
# find longest junk-free match
1508
# during an iteration of the loop, j2len[j] = length of longest
1509
# junk-free match ending with a[i-1] and b[j]
1513
for i in xrange(alo, ahi):
1514
# look at all instances of a[i] in b; note that because
1515
# b2j has no junk keys, the loop is skipped if a[i] is junk
1516
j2lenget = j2len.get
1519
# changing b2j.get(a[i], nothing) to a try:Keyerror pair produced the
1520
# following improvement
1521
# 704 0 4650.5320 2620.7410 bzrlib.knit:1336(find_longest_match)
1522
# +326674 0 1655.1210 1655.1210 +<method 'get' of 'dict' objects>
1523
# +76519 0 374.6700 374.6700 +<method 'has_key' of 'dict' objects>
1525
# 704 0 3733.2820 2209.6520 bzrlib.knit:1336(find_longest_match)
1526
# +211400 0 1147.3520 1147.3520 +<method 'get' of 'dict' objects>
1527
# +76519 0 376.2780 376.2780 +<method 'has_key' of 'dict' objects>
1539
k = newj2len[j] = 1 + j2lenget(-1 + j, 0)
1541
besti, bestj, bestsize = 1 + i-k, 1 + j-k, k
1544
# Extend the best by non-junk elements on each end. In particular,
1545
# "popular" non-junk elements aren't in b2j, which greatly speeds
1546
# the inner loop above, but also means "the best" match so far
1547
# doesn't contain any junk *or* popular non-junk elements.
1548
while besti > alo and bestj > blo and \
1549
not isbjunk(b[bestj-1]) and \
1550
a[besti-1] == b[bestj-1]:
1551
besti, bestj, bestsize = besti-1, bestj-1, bestsize+1
1552
while besti+bestsize < ahi and bestj+bestsize < bhi and \
1553
not isbjunk(b[bestj+bestsize]) and \
1554
a[besti+bestsize] == b[bestj+bestsize]:
1557
# Now that we have a wholly interesting match (albeit possibly
1558
# empty!), we may as well suck up the matching junk on each
1559
# side of it too. Can't think of a good reason not to, and it
1560
# saves post-processing the (possibly considerable) expense of
1561
# figuring out what to do with it. In the case of an empty
1562
# interesting match, this is clearly the right thing to do,
1563
# because no other kind of match is possible in the regions.
1564
while besti > alo and bestj > blo and \
1565
isbjunk(b[bestj-1]) and \
1566
a[besti-1] == b[bestj-1]:
1567
besti, bestj, bestsize = besti-1, bestj-1, bestsize+1
1568
while besti+bestsize < ahi and bestj+bestsize < bhi and \
1569
isbjunk(b[bestj+bestsize]) and \
1570
a[besti+bestsize] == b[bestj+bestsize]:
1571
bestsize = bestsize + 1
1573
return besti, bestj, bestsize