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)
289
self._data = _KnitData(transport, relpath + DATA_SUFFIX,
290
access_mode, create=not len(self.versions()))
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)
422
"""See VersionedFile.get_suffixes()."""
423
return [DATA_SUFFIX, INDEX_SUFFIX]
425
def has_ghost(self, version_id):
426
"""True if there is a ghost reference in the file to version_id."""
428
if self.has_version(version_id):
430
# optimisable if needed by memoising the _ghosts set.
431
items = self._index.get_graph()
432
for node, parents in items:
433
for parent in parents:
434
if parent not in self._index._cache:
435
if parent == version_id:
440
"""See VersionedFile.versions."""
441
return self._index.get_versions()
443
def has_version(self, version_id):
444
"""See VersionedFile.has_version."""
445
return self._index.has_version(version_id)
447
__contains__ = has_version
449
def _merge_annotations(self, content, parents, parent_texts={},
450
delta=None, annotated=None):
451
"""Merge annotations for content. This is done by comparing
452
the annotations based on changed to the text.
456
for parent_id in parents:
457
merge_content = self._get_content(parent_id, parent_texts)
458
seq = SequenceMatcher(None, merge_content.text(), content.text())
459
if delta_seq is None:
460
# setup a delta seq to reuse.
462
for i, j, n in seq.get_matching_blocks():
465
# this appears to copy (origin, text) pairs across to the new
466
# content for any line that matches the last-checked parent.
467
# FIXME: save the sequence control data for delta compression
468
# against the most relevant parent rather than rediffing.
469
content._lines[j:j+n] = merge_content._lines[i:i+n]
472
reference_content = self._get_content(parents[0], parent_texts)
473
new_texts = content.text()
474
old_texts = reference_content.text()
475
delta_seq = SequenceMatcher(None, old_texts, new_texts)
476
return self._make_line_delta(delta_seq, content)
478
def _make_line_delta(self, delta_seq, new_content):
479
"""Generate a line delta from delta_seq and new_content."""
481
for op in delta_seq.get_opcodes():
484
diff_hunks.append((op[1], op[2], op[4]-op[3], new_content._lines[op[3]:op[4]]))
487
def _get_components(self, version_id):
488
"""Return a list of (version_id, method, data) tuples that
489
makes up version specified by version_id of the knit.
491
The components should be applied in the order of the returned
494
The basis knit will be used to the largest extent possible
495
since it is assumed that accesses to it is faster.
498
# 4168 calls in 14912, 2289 internal
499
# 4168 in 9711 to read_records
500
# 52554 in 1250 to get_parents
501
# 170166 in 865 to list.append
503
# needed_revisions holds a list of (method, version_id) of
504
# versions that is needed to be fetched to construct the final
505
# version of the file.
507
# basis_revisions is a list of versions that needs to be
508
# fetched but exists in the basis knit.
510
basis = self.basis_knit
517
if basis and basis._index.has_version(cursor):
519
basis_versions.append(cursor)
520
method = picked_knit._index.get_method(cursor)
521
needed_versions.append((method, cursor))
522
if method == 'fulltext':
524
cursor = picked_knit.get_parents(cursor)[0]
529
for comp_id in basis_versions:
530
data_pos, data_size = basis._index.get_data_position(comp_id)
531
records.append((piece_id, data_pos, data_size))
532
components.update(basis._data.read_records(records))
535
for comp_id in [vid for method, vid in needed_versions
536
if vid not in basis_versions]:
537
data_pos, data_size = self._index.get_position(comp_id)
538
records.append((comp_id, data_pos, data_size))
539
components.update(self._data.read_records(records))
541
# get_data_records returns a mapping with the version id as
542
# index and the value as data. The order the components need
543
# to be applied is held by needed_versions (reversed).
545
for method, comp_id in reversed(needed_versions):
546
out.append((comp_id, method, components[comp_id]))
550
def _get_content(self, version_id, parent_texts={}):
551
"""Returns a content object that makes up the specified
553
if not self.has_version(version_id):
554
raise RevisionNotPresent(version_id, self.filename)
556
cached_version = parent_texts.get(version_id, None)
557
if cached_version is not None:
558
return cached_version
560
if self.basis_knit and version_id in self.basis_knit:
561
return self.basis_knit._get_content(version_id)
564
components = self._get_components(version_id)
565
for component_id, method, (data, digest) in components:
566
version_idx = self._index.lookup(component_id)
567
if method == 'fulltext':
568
assert content is None
569
content = self.factory.parse_fulltext(data, version_idx)
570
elif method == 'line-delta':
571
delta = self.factory.parse_line_delta(data, version_idx)
572
content._lines = self._apply_delta(content._lines, delta)
574
if 'no-eol' in self._index.get_options(version_id):
575
line = content._lines[-1][1].rstrip('\n')
576
content._lines[-1] = (content._lines[-1][0], line)
578
if sha_strings(content.text()) != digest:
579
import pdb;pdb.set_trace()
580
raise KnitCorrupt(self.filename, 'sha-1 does not match %s' % version_id)
584
def _check_versions_present(self, version_ids):
585
"""Check that all specified versions are present."""
586
version_ids = set(version_ids)
587
for r in list(version_ids):
588
if self._index.has_version(r):
589
version_ids.remove(r)
591
raise RevisionNotPresent(list(version_ids)[0], self.filename)
593
def _add_lines_with_ghosts(self, version_id, parents, lines, parent_texts):
594
"""See VersionedFile.add_lines_with_ghosts()."""
595
self._check_add(version_id, lines)
596
return self._add(version_id, lines[:], parents, self.delta, parent_texts)
598
def _add_lines(self, version_id, parents, lines, parent_texts):
599
"""See VersionedFile.add_lines."""
600
self._check_add(version_id, lines)
601
self._check_versions_present(parents)
602
return self._add(version_id, lines[:], parents, self.delta, parent_texts)
604
def _check_add(self, version_id, lines):
605
"""check that version_id and lines are safe to add."""
606
assert self.writable, "knit is not opened for write"
607
### FIXME escape. RBC 20060228
608
if contains_whitespace(version_id):
609
raise InvalidRevisionId(version_id)
610
if self.has_version(version_id):
611
raise RevisionAlreadyPresent(version_id, self.filename)
613
if False or __debug__:
615
assert '\n' not in l[:-1]
617
def _add(self, version_id, lines, parents, delta, parent_texts):
618
"""Add a set of lines on top of version specified by parents.
620
If delta is true, compress the text as a line-delta against
623
Any versions not present will be converted into ghosts.
625
# 461 0 6546.0390 43.9100 bzrlib.knit:489(_add)
626
# +400 0 889.4890 418.9790 +bzrlib.knit:192(lower_fulltext)
627
# +461 0 1364.8070 108.8030 +bzrlib.knit:996(add_record)
628
# +461 0 193.3940 41.5720 +bzrlib.knit:898(add_version)
629
# +461 0 134.0590 18.3810 +bzrlib.osutils:361(sha_strings)
630
# +461 0 36.3420 15.4540 +bzrlib.knit:146(make)
631
# +1383 0 8.0370 8.0370 +<len>
632
# +61 0 13.5770 7.9190 +bzrlib.knit:199(lower_line_delta)
633
# +61 0 963.3470 7.8740 +bzrlib.knit:427(_get_content)
634
# +61 0 973.9950 5.2950 +bzrlib.knit:136(line_delta)
635
# +61 0 1918.1800 5.2640 +bzrlib.knit:359(_merge_annotations)
639
if parent_texts is None:
641
for parent in parents:
642
if not self.has_version(parent):
643
ghosts.append(parent)
645
present_parents.append(parent)
647
if delta and not len(present_parents):
650
digest = sha_strings(lines)
653
if lines[-1][-1] != '\n':
654
options.append('no-eol')
655
lines[-1] = lines[-1] + '\n'
657
if len(present_parents) and delta:
658
# To speed the extract of texts the delta chain is limited
659
# to a fixed number of deltas. This should minimize both
660
# I/O and the time spend applying deltas.
662
delta_parents = present_parents
664
parent = delta_parents[0]
665
method = self._index.get_method(parent)
666
if method == 'fulltext':
668
delta_parents = self._index.get_parents(parent)
670
if method == 'line-delta':
673
lines = self.factory.make(lines, version_id)
674
if delta or (self.factory.annotated and len(present_parents) > 0):
675
# Merge annotations from parent texts if so is needed.
676
delta_hunks = self._merge_annotations(lines, present_parents, parent_texts,
677
delta, self.factory.annotated)
680
options.append('line-delta')
681
store_lines = self.factory.lower_line_delta(delta_hunks)
683
options.append('fulltext')
684
store_lines = self.factory.lower_fulltext(lines)
686
where, size = self._data.add_record(version_id, digest, store_lines)
687
self._index.add_version(version_id, options, where, size, parents)
690
def check(self, progress_bar=None):
691
"""See VersionedFile.check()."""
693
def _clone_text(self, new_version_id, old_version_id, parents):
694
"""See VersionedFile.clone_text()."""
695
# FIXME RBC 20060228 make fast by only inserting an index with null delta.
696
self.add_lines(new_version_id, parents, self.get_lines(old_version_id))
698
def get_lines(self, version_id):
699
"""See VersionedFile.get_lines()."""
700
return self._get_content(version_id).text()
702
def iter_lines_added_or_present_in_versions(self, version_ids=None):
703
"""See VersionedFile.iter_lines_added_or_present_in_versions()."""
704
if version_ids is None:
705
version_ids = self.versions()
706
# we dont care about inclusions, the caller cares.
707
# but we need to setup a list of records to visit.
708
# we need version_id, position, length
709
version_id_records = []
710
requested_versions = list(version_ids)
711
# filter for available versions
712
for version_id in requested_versions:
713
if not self.has_version(version_id):
714
raise RevisionNotPresent(version_id, self.filename)
715
# get a in-component-order queue:
717
for version_id in self.versions():
718
if version_id in requested_versions:
719
version_ids.append(version_id)
720
data_pos, length = self._index.get_position(version_id)
721
version_id_records.append((version_id, data_pos, length))
723
pb = bzrlib.ui.ui_factory.nested_progress_bar()
725
total = len(version_id_records)
727
pb.update('Walking content.', count, total)
728
for version_id, data, sha_value in \
729
self._data.read_records_iter(version_id_records):
730
pb.update('Walking content.', count, total)
731
method = self._index.get_method(version_id)
732
version_idx = self._index.lookup(version_id)
733
assert method in ('fulltext', 'line-delta')
734
if method == 'fulltext':
735
content = self.factory.parse_fulltext(data, version_idx)
736
for line in content.text():
739
delta = self.factory.parse_line_delta(data, version_idx)
740
for start, end, count, lines in delta:
741
for origin, line in lines:
744
pb.update('Walking content.', total, total)
747
pb.update('Walking content.', total, total)
751
def num_versions(self):
752
"""See VersionedFile.num_versions()."""
753
return self._index.num_versions()
755
__len__ = num_versions
757
def annotate_iter(self, version_id):
758
"""See VersionedFile.annotate_iter."""
759
content = self._get_content(version_id)
760
for origin, text in content.annotate_iter():
763
def get_parents(self, version_id):
764
"""See VersionedFile.get_parents."""
767
# 52554 calls in 1264 872 internal down from 3674
769
return self._index.get_parents(version_id)
771
raise RevisionNotPresent(version_id, self.filename)
773
def get_parents_with_ghosts(self, version_id):
774
"""See VersionedFile.get_parents."""
776
return self._index.get_parents_with_ghosts(version_id)
778
raise RevisionNotPresent(version_id, self.filename)
780
def get_ancestry(self, versions):
781
"""See VersionedFile.get_ancestry."""
782
if isinstance(versions, basestring):
783
versions = [versions]
786
self._check_versions_present(versions)
787
return self._index.get_ancestry(versions)
789
def get_ancestry_with_ghosts(self, versions):
790
"""See VersionedFile.get_ancestry_with_ghosts."""
791
if isinstance(versions, basestring):
792
versions = [versions]
795
self._check_versions_present(versions)
796
return self._index.get_ancestry_with_ghosts(versions)
798
#@deprecated_method(zero_eight)
799
def walk(self, version_ids):
800
"""See VersionedFile.walk."""
801
# We take the short path here, and extract all relevant texts
802
# and put them in a weave and let that do all the work. Far
803
# from optimal, but is much simpler.
804
# FIXME RB 20060228 this really is inefficient!
805
from bzrlib.weave import Weave
807
w = Weave(self.filename)
808
ancestry = self.get_ancestry(version_ids)
809
sorted_graph = topo_sort(self._index.get_graph())
810
version_list = [vid for vid in sorted_graph if vid in ancestry]
812
for version_id in version_list:
813
lines = self.get_lines(version_id)
814
w.add_lines(version_id, self.get_parents(version_id), lines)
816
for lineno, insert_id, dset, line in w.walk(version_ids):
817
yield lineno, insert_id, dset, line
820
class _KnitComponentFile(object):
821
"""One of the files used to implement a knit database"""
823
def __init__(self, transport, filename, mode):
824
self._transport = transport
825
self._filename = filename
828
def write_header(self):
829
if self._transport.append(self._filename, StringIO(self.HEADER)):
830
raise KnitCorrupt(self._filename, 'misaligned after writing header')
832
def check_header(self, fp):
834
if line != self.HEADER:
835
raise KnitHeaderError(badline=line)
838
"""Commit is a nop."""
841
return '%s(%s)' % (self.__class__.__name__, self._filename)
844
class _KnitIndex(_KnitComponentFile):
845
"""Manages knit index file.
847
The index is already kept in memory and read on startup, to enable
848
fast lookups of revision information. The cursor of the index
849
file is always pointing to the end, making it easy to append
852
_cache is a cache for fast mapping from version id to a Index
855
_history is a cache for fast mapping from indexes to version ids.
857
The index data format is dictionary compressed when it comes to
858
parent references; a index entry may only have parents that with a
859
lover index number. As a result, the index is topological sorted.
861
Duplicate entries may be written to the index for a single version id
862
if this is done then the latter one completely replaces the former:
863
this allows updates to correct version and parent information.
864
Note that the two entries may share the delta, and that successive
865
annotations and references MUST point to the first entry.
867
The index file on disc contains a header, followed by one line per knit
868
record. The same revision can be present in an index file more than once.
869
The first occurence gets assigned a sequence number starting from 0.
871
The format of a single line is
872
REVISION_ID FLAGS BYTE_OFFSET LENGTH( PARENT_ID|PARENT_SEQUENCE_ID)* :\n
873
REVISION_ID is a utf8-encoded revision id
874
FLAGS is a comma separated list of flags about the record. Values include
875
no-eol, line-delta, fulltext.
876
BYTE_OFFSET is the ascii representation of the byte offset in the data file
877
that the the compressed data starts at.
878
LENGTH is the ascii representation of the length of the data file.
879
PARENT_ID a utf-8 revision id prefixed by a '.' that is a parent of
881
PARENT_SEQUENCE_ID the ascii representation of the sequence number of a
882
revision id already in the knit that is a parent of REVISION_ID.
883
The ' :' marker is the end of record marker.
886
when a write is interrupted to the index file, it will result in a line that
887
does not end in ' :'. If the ' :' is not present at the end of a line, or at
888
the end of the file, then the record that is missing it will be ignored by
891
When writing new records to the index file, the data is preceeded by '\n'
892
to ensure that records always start on new lines even if the last write was
893
interrupted. As a result its normal for the last line in the index to be
894
missing a trailing newline. One can be added with no harmful effects.
897
HEADER = "# bzr knit index 7\n"
899
# speed of knit parsing went from 280 ms to 280 ms with slots addition.
900
# __slots__ = ['_cache', '_history', '_transport', '_filename']
902
def _cache_version(self, version_id, options, pos, size, parents):
903
"""Cache a version record in the history array and index cache.
905
This is inlined into __init__ for performance. KEEP IN SYNC.
906
(It saves 60ms, 25% of the __init__ overhead on local 4000 record
909
# only want the _history index to reference the 1st index entry
911
if version_id not in self._cache:
912
index = len(self._history)
913
self._history.append(version_id)
915
index = self._cache[version_id][5]
916
self._cache[version_id] = (version_id,
923
def __init__(self, transport, filename, mode, create=False):
924
_KnitComponentFile.__init__(self, transport, filename, mode)
926
# position in _history is the 'official' index for a revision
927
# but the values may have come from a newer entry.
928
# so - wc -l of a knit index is != the number of uniqe names
931
pb = bzrlib.ui.ui_factory.nested_progress_bar()
936
pb.update('read knit index', count, total)
937
fp = self._transport.get(self._filename)
938
self.check_header(fp)
939
# readlines reads the whole file at once:
940
# bad for transports like http, good for local disk
941
# we save 60 ms doing this one change (
942
# from calling readline each time to calling
944
# probably what we want for nice behaviour on
945
# http is a incremental readlines that yields, or
946
# a check for local vs non local indexes,
947
for l in fp.readlines():
949
if len(rec) < 5 or rec[-1] != ':':
951
# FIXME: in the future we should determine if its a
952
# short write - and ignore it
953
# or a different failure, and raise. RBC 20060407
957
#pb.update('read knit index', count, total)
958
# See self._parse_parents
960
for value in rec[4:-1]:
962
# uncompressed reference
963
parents.append(value[1:])
965
# this is 15/4000ms faster than isinstance,
967
# this function is called thousands of times a
968
# second so small variations add up.
969
assert value.__class__ is str
970
parents.append(self._history[int(value)])
971
# end self._parse_parents
972
# self._cache_version(rec[0],
977
# --- self._cache_version
978
# only want the _history index to reference the 1st
979
# index entry for version_id
981
if version_id not in self._cache:
982
index = len(self._history)
983
self._history.append(version_id)
985
index = self._cache[version_id][5]
986
self._cache[version_id] = (version_id,
992
# --- self._cache_version
993
except NoSuchFile, e:
994
if mode != 'w' or not create:
998
pb.update('read knit index', total, total)
1001
def _parse_parents(self, compressed_parents):
1002
"""convert a list of string parent values into version ids.
1004
ints are looked up in the index.
1005
.FOO values are ghosts and converted in to FOO.
1007
NOTE: the function is retained here for clarity, and for possible
1008
use in partial index reads. However bulk processing now has
1009
it inlined in __init__ for inner-loop optimisation.
1012
for value in compressed_parents:
1013
if value[-1] == '.':
1014
# uncompressed reference
1015
result.append(value[1:])
1017
# this is 15/4000ms faster than isinstance,
1018
# this function is called thousands of times a
1019
# second so small variations add up.
1020
assert value.__class__ is str
1021
result.append(self._history[int(value)])
1024
def get_graph(self):
1026
for version_id, index in self._cache.iteritems():
1027
graph.append((version_id, index[4]))
1030
def get_ancestry(self, versions):
1031
"""See VersionedFile.get_ancestry."""
1032
# get a graph of all the mentioned versions:
1034
pending = set(versions)
1036
version = pending.pop()
1037
parents = self._cache[version][4]
1038
# got the parents ok
1040
parents = [parent for parent in parents if parent in self._cache]
1041
for parent in parents:
1042
# if not completed and not a ghost
1043
if parent not in graph:
1045
graph[version] = parents
1046
return topo_sort(graph.items())
1048
def get_ancestry_with_ghosts(self, versions):
1049
"""See VersionedFile.get_ancestry_with_ghosts."""
1050
# get a graph of all the mentioned versions:
1052
pending = set(versions)
1054
version = pending.pop()
1056
parents = self._cache[version][4]
1062
# got the parents ok
1063
for parent in parents:
1064
if parent not in graph:
1066
graph[version] = parents
1067
return topo_sort(graph.items())
1069
def num_versions(self):
1070
return len(self._history)
1072
__len__ = num_versions
1074
def get_versions(self):
1075
return self._history
1077
def idx_to_name(self, idx):
1078
return self._history[idx]
1080
def lookup(self, version_id):
1081
assert version_id in self._cache
1082
return self._cache[version_id][5]
1084
def _version_list_to_index(self, versions):
1086
for version in versions:
1087
if version in self._cache:
1088
# -- inlined lookup() --
1089
result_list.append(str(self._cache[version][5]))
1090
# -- end lookup () --
1092
result_list.append('.' + version.encode('utf-8'))
1093
return ' '.join(result_list)
1095
def add_version(self, version_id, options, pos, size, parents):
1096
"""Add a version record to the index."""
1097
self._cache_version(version_id, options, pos, size, parents)
1099
content = "\n%s %s %s %s %s :" % (version_id.encode('utf-8'),
1103
self._version_list_to_index(parents))
1104
assert isinstance(content, str), 'content must be utf-8 encoded'
1105
self._transport.append(self._filename, StringIO(content))
1107
def has_version(self, version_id):
1108
"""True if the version is in the index."""
1109
return self._cache.has_key(version_id)
1111
def get_position(self, version_id):
1112
"""Return data position and size of specified version."""
1113
return (self._cache[version_id][2], \
1114
self._cache[version_id][3])
1116
def get_method(self, version_id):
1117
"""Return compression method of specified version."""
1118
options = self._cache[version_id][1]
1119
if 'fulltext' in options:
1122
assert 'line-delta' in options
1125
def get_options(self, version_id):
1126
return self._cache[version_id][1]
1128
def get_parents(self, version_id):
1129
"""Return parents of specified version ignoring ghosts."""
1130
return [parent for parent in self._cache[version_id][4]
1131
if parent in self._cache]
1133
def get_parents_with_ghosts(self, version_id):
1134
"""Return parents of specified version wth ghosts."""
1135
return self._cache[version_id][4]
1137
def check_versions_present(self, version_ids):
1138
"""Check that all specified versions are present."""
1139
version_ids = set(version_ids)
1140
for version_id in list(version_ids):
1141
if version_id in self._cache:
1142
version_ids.remove(version_id)
1144
raise RevisionNotPresent(list(version_ids)[0], self.filename)
1147
class _KnitData(_KnitComponentFile):
1148
"""Contents of the knit data file"""
1150
HEADER = "# bzr knit data 7\n"
1152
def __init__(self, transport, filename, mode, create=False):
1153
_KnitComponentFile.__init__(self, transport, filename, mode)
1155
self._checked = False
1157
self._transport.put(self._filename, StringIO(''))
1160
def clear_cache(self):
1161
"""Clear the record cache."""
1164
def _open_file(self):
1165
if self._file is None:
1167
self._file = self._transport.get(self._filename)
1172
def _record_to_data(self, version_id, digest, lines):
1173
"""Convert version_id, digest, lines into a raw data block.
1175
:return: (len, a StringIO instance with the raw data ready to read.)
1178
data_file = GzipFile(None, mode='wb', fileobj=sio)
1179
data_file.writelines(chain(
1180
["version %s %d %s\n" % (version_id.encode('utf-8'),
1184
["end %s\n" % version_id.encode('utf-8')]))
1191
def add_raw_record(self, raw_data):
1192
"""Append a prepared record to the data file."""
1193
assert isinstance(raw_data, str), 'data must be plain bytes'
1194
start_pos = self._transport.append(self._filename, StringIO(raw_data))
1195
return start_pos, len(raw_data)
1197
def add_record(self, version_id, digest, lines):
1198
"""Write new text record to disk. Returns the position in the
1199
file where it was written."""
1200
size, sio = self._record_to_data(version_id, digest, lines)
1202
self._records[version_id] = (digest, lines)
1204
start_pos = self._transport.append(self._filename, sio)
1205
return start_pos, size
1207
def _parse_record_header(self, version_id, raw_data):
1208
"""Parse a record header for consistency.
1210
:return: the header and the decompressor stream.
1211
as (stream, header_record)
1213
df = GzipFile(mode='rb', fileobj=StringIO(raw_data))
1214
rec = df.readline().split()
1216
raise KnitCorrupt(self._filename, 'unexpected number of elements in record header')
1217
if rec[1].decode('utf-8')!= version_id:
1218
raise KnitCorrupt(self._filename,
1219
'unexpected version, wanted %r, got %r' % (
1220
version_id, rec[1]))
1223
def _parse_record(self, version_id, data):
1225
# 4168 calls in 2880 217 internal
1226
# 4168 calls to _parse_record_header in 2121
1227
# 4168 calls to readlines in 330
1228
df, rec = self._parse_record_header(version_id, data)
1229
record_contents = df.readlines()
1230
l = record_contents.pop()
1231
assert len(record_contents) == int(rec[2])
1232
if l.decode('utf-8') != 'end %s\n' % version_id:
1233
raise KnitCorrupt(self._filename, 'unexpected version end line %r, wanted %r'
1236
return record_contents, rec[3]
1238
def read_records_iter_raw(self, records):
1239
"""Read text records from data file and yield raw data.
1241
This unpacks enough of the text record to validate the id is
1242
as expected but thats all.
1244
It will actively recompress currently cached records on the
1245
basis that that is cheaper than I/O activity.
1248
for version_id, pos, size in records:
1249
if version_id not in self._records:
1250
needed_records.append((version_id, pos, size))
1252
# setup an iterator of the external records:
1253
# uses readv so nice and fast we hope.
1254
if len(needed_records):
1255
# grab the disk data needed.
1256
raw_records = self._transport.readv(self._filename,
1257
[(pos, size) for version_id, pos, size in needed_records])
1259
for version_id, pos, size in records:
1260
if version_id in self._records:
1261
# compress a new version
1262
size, sio = self._record_to_data(version_id,
1263
self._records[version_id][0],
1264
self._records[version_id][1])
1265
yield version_id, sio.getvalue()
1267
pos, data = raw_records.next()
1268
# validate the header
1269
df, rec = self._parse_record_header(version_id, data)
1271
yield version_id, data
1274
def read_records_iter(self, records):
1275
"""Read text records from data file and yield result.
1277
Each passed record is a tuple of (version_id, pos, len) and
1278
will be read in the given order. Yields (version_id,
1282
# 60890 calls for 4168 extractions in 5045, 683 internal.
1283
# 4168 calls to readv in 1411
1284
# 4168 calls to parse_record in 2880
1287
for version_id, pos, size in records:
1288
if version_id not in self._records:
1289
needed_records.append((version_id, pos, size))
1291
if len(needed_records):
1292
# We take it that the transport optimizes the fetching as good
1293
# as possible (ie, reads continous ranges.)
1294
response = self._transport.readv(self._filename,
1295
[(pos, size) for version_id, pos, size in needed_records])
1297
for (record_id, pos, size), (pos, data) in izip(iter(needed_records), response):
1298
content, digest = self._parse_record(record_id, data)
1299
self._records[record_id] = (digest, content)
1301
for version_id, pos, size in records:
1302
yield version_id, list(self._records[version_id][1]), self._records[version_id][0]
1304
def read_records(self, records):
1305
"""Read records into a dictionary."""
1307
for record_id, content, digest in self.read_records_iter(records):
1308
components[record_id] = (content, digest)
1312
class InterKnit(InterVersionedFile):
1313
"""Optimised code paths for knit to knit operations."""
1315
_matching_file_factory = KnitVersionedFile
1318
def is_compatible(source, target):
1319
"""Be compatible with knits. """
1321
return (isinstance(source, KnitVersionedFile) and
1322
isinstance(target, KnitVersionedFile))
1323
except AttributeError:
1326
def join(self, pb=None, msg=None, version_ids=None, ignore_missing=False):
1327
"""See InterVersionedFile.join."""
1328
assert isinstance(self.source, KnitVersionedFile)
1329
assert isinstance(self.target, KnitVersionedFile)
1331
if version_ids is None:
1332
version_ids = self.source.versions()
1334
if not ignore_missing:
1335
self.source._check_versions_present(version_ids)
1337
version_ids = set(self.source.versions()).intersection(
1343
pb = bzrlib.ui.ui_factory.nested_progress_bar()
1345
version_ids = list(version_ids)
1346
if None in version_ids:
1347
version_ids.remove(None)
1349
self.source_ancestry = set(self.source.get_ancestry(version_ids))
1350
this_versions = set(self.target._index.get_versions())
1351
needed_versions = self.source_ancestry - this_versions
1352
cross_check_versions = self.source_ancestry.intersection(this_versions)
1353
mismatched_versions = set()
1354
for version in cross_check_versions:
1355
# scan to include needed parents.
1356
n1 = set(self.target.get_parents_with_ghosts(version))
1357
n2 = set(self.source.get_parents_with_ghosts(version))
1359
# FIXME TEST this check for cycles being introduced works
1360
# the logic is we have a cycle if in our graph we are an
1361
# ancestor of any of the n2 revisions.
1367
parent_ancestors = self.source.get_ancestry(parent)
1368
if version in parent_ancestors:
1369
raise errors.GraphCycleError([parent, version])
1370
# ensure this parent will be available later.
1371
new_parents = n2.difference(n1)
1372
needed_versions.update(new_parents.difference(this_versions))
1373
mismatched_versions.add(version)
1375
if not needed_versions and not cross_check_versions:
1377
full_list = topo_sort(self.source.get_graph())
1379
version_list = [i for i in full_list if (not self.target.has_version(i)
1380
and i in needed_versions)]
1384
copy_queue_records = []
1386
for version_id in version_list:
1387
options = self.source._index.get_options(version_id)
1388
parents = self.source._index.get_parents_with_ghosts(version_id)
1389
# check that its will be a consistent copy:
1390
for parent in parents:
1391
# if source has the parent, we must :
1392
# * already have it or
1393
# * have it scheduled already
1394
# otherwise we dont care
1395
assert (self.target.has_version(parent) or
1396
parent in copy_set or
1397
not self.source.has_version(parent))
1398
data_pos, data_size = self.source._index.get_position(version_id)
1399
copy_queue_records.append((version_id, data_pos, data_size))
1400
copy_queue.append((version_id, options, parents))
1401
copy_set.add(version_id)
1403
# data suck the join:
1405
total = len(version_list)
1406
# we want the raw gzip for bulk copying, but the record validated
1407
# just enough to be sure its the right one.
1408
# TODO: consider writev or write combining to reduce
1409
# death of a thousand cuts feeling.
1410
for (version_id, raw_data), \
1411
(version_id2, options, parents) in \
1412
izip(self.source._data.read_records_iter_raw(copy_queue_records),
1414
assert version_id == version_id2, 'logic error, inconsistent results'
1416
pb.update("Joining knit", count, total)
1417
pos, size = self.target._data.add_raw_record(raw_data)
1418
self.target._index.add_version(version_id, options, pos, size, parents)
1420
for version in mismatched_versions:
1421
# FIXME RBC 20060309 is this needed?
1422
n1 = set(self.target.get_parents_with_ghosts(version))
1423
n2 = set(self.source.get_parents_with_ghosts(version))
1424
# write a combined record to our history preserving the current
1425
# parents as first in the list
1426
new_parents = self.target.get_parents_with_ghosts(version) + list(n2.difference(n1))
1427
self.target.fix_parents(version, new_parents)
1433
InterVersionedFile.register_optimiser(InterKnit)
1436
class SequenceMatcher(difflib.SequenceMatcher):
1437
"""Knit tuned sequence matcher.
1439
This is based on profiling of difflib which indicated some improvements
1440
for our usage pattern.
1443
def find_longest_match(self, alo, ahi, blo, bhi):
1444
"""Find longest matching block in a[alo:ahi] and b[blo:bhi].
1446
If isjunk is not defined:
1448
Return (i,j,k) such that a[i:i+k] is equal to b[j:j+k], where
1449
alo <= i <= i+k <= ahi
1450
blo <= j <= j+k <= bhi
1451
and for all (i',j',k') meeting those conditions,
1454
and if i == i', j <= j'
1456
In other words, of all maximal matching blocks, return one that
1457
starts earliest in a, and of all those maximal matching blocks that
1458
start earliest in a, return the one that starts earliest in b.
1460
>>> s = SequenceMatcher(None, " abcd", "abcd abcd")
1461
>>> s.find_longest_match(0, 5, 0, 9)
1464
If isjunk is defined, first the longest matching block is
1465
determined as above, but with the additional restriction that no
1466
junk element appears in the block. Then that block is extended as
1467
far as possible by matching (only) junk elements on both sides. So
1468
the resulting block never matches on junk except as identical junk
1469
happens to be adjacent to an "interesting" match.
1471
Here's the same example as before, but considering blanks to be
1472
junk. That prevents " abcd" from matching the " abcd" at the tail
1473
end of the second sequence directly. Instead only the "abcd" can
1474
match, and matches the leftmost "abcd" in the second sequence:
1476
>>> s = SequenceMatcher(lambda x: x==" ", " abcd", "abcd abcd")
1477
>>> s.find_longest_match(0, 5, 0, 9)
1480
If no blocks match, return (alo, blo, 0).
1482
>>> s = SequenceMatcher(None, "ab", "c")
1483
>>> s.find_longest_match(0, 2, 0, 1)
1487
# CAUTION: stripping common prefix or suffix would be incorrect.
1491
# Longest matching block is "ab", but if common prefix is
1492
# stripped, it's "a" (tied with "b"). UNIX(tm) diff does so
1493
# strip, so ends up claiming that ab is changed to acab by
1494
# inserting "ca" in the middle. That's minimal but unintuitive:
1495
# "it's obvious" that someone inserted "ac" at the front.
1496
# Windiff ends up at the same place as diff, but by pairing up
1497
# the unique 'b's and then matching the first two 'a's.
1499
a, b, b2j, isbjunk = self.a, self.b, self.b2j, self.isbjunk
1500
besti, bestj, bestsize = alo, blo, 0
1501
# find longest junk-free match
1502
# during an iteration of the loop, j2len[j] = length of longest
1503
# junk-free match ending with a[i-1] and b[j]
1507
for i in xrange(alo, ahi):
1508
# look at all instances of a[i] in b; note that because
1509
# b2j has no junk keys, the loop is skipped if a[i] is junk
1510
j2lenget = j2len.get
1513
# changing b2j.get(a[i], nothing) to a try:Keyerror pair produced the
1514
# following improvement
1515
# 704 0 4650.5320 2620.7410 bzrlib.knit:1336(find_longest_match)
1516
# +326674 0 1655.1210 1655.1210 +<method 'get' of 'dict' objects>
1517
# +76519 0 374.6700 374.6700 +<method 'has_key' of 'dict' objects>
1519
# 704 0 3733.2820 2209.6520 bzrlib.knit:1336(find_longest_match)
1520
# +211400 0 1147.3520 1147.3520 +<method 'get' of 'dict' objects>
1521
# +76519 0 376.2780 376.2780 +<method 'has_key' of 'dict' objects>
1533
k = newj2len[j] = 1 + j2lenget(-1 + j, 0)
1535
besti, bestj, bestsize = 1 + i-k, 1 + j-k, k
1538
# Extend the best by non-junk elements on each end. In particular,
1539
# "popular" non-junk elements aren't in b2j, which greatly speeds
1540
# the inner loop above, but also means "the best" match so far
1541
# doesn't contain any junk *or* popular non-junk elements.
1542
while besti > alo and bestj > blo and \
1543
not isbjunk(b[bestj-1]) and \
1544
a[besti-1] == b[bestj-1]:
1545
besti, bestj, bestsize = besti-1, bestj-1, bestsize+1
1546
while besti+bestsize < ahi and bestj+bestsize < bhi and \
1547
not isbjunk(b[bestj+bestsize]) and \
1548
a[besti+bestsize] == b[bestj+bestsize]:
1551
# Now that we have a wholly interesting match (albeit possibly
1552
# empty!), we may as well suck up the matching junk on each
1553
# side of it too. Can't think of a good reason not to, and it
1554
# saves post-processing the (possibly considerable) expense of
1555
# figuring out what to do with it. In the case of an empty
1556
# interesting match, this is clearly the right thing to do,
1557
# because no other kind of match is possible in the regions.
1558
while besti > alo and bestj > blo and \
1559
isbjunk(b[bestj-1]) and \
1560
a[besti-1] == b[bestj-1]:
1561
besti, bestj, bestsize = besti-1, bestj-1, bestsize+1
1562
while besti+bestsize < ahi and bestj+bestsize < bhi and \
1563
isbjunk(b[bestj+bestsize]) and \
1564
a[besti+bestsize] == b[bestj+bestsize]:
1565
bestsize = bestsize + 1
1567
return besti, bestj, bestsize
added
removed
bzrlib/knit.py
