1
# Copyright (C) 2008, 2009, 2010 Canonical Ltd
3
# This program is free software; you can redistribute it and/or modify
4
# it under the terms of the GNU General Public License as published by
5
# the Free Software Foundation; either version 2 of the License, or
6
# (at your option) any later version.
8
# This program is distributed in the hope that it will be useful,
9
# but WITHOUT ANY WARRANTY; without even the implied warranty of
10
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11
# GNU General Public License for more details.
13
# You should have received a copy of the GNU General Public License
14
# along with this program; if not, write to the Free Software
15
# Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
17
"""Persistent maps from tuple_of_strings->string using CHK stores.
19
Overview and current status:
21
The CHKMap class implements a dict from tuple_of_strings->string by using a trie
22
with internal nodes of 8-bit fan out; The key tuples are mapped to strings by
23
joining them by \x00, and \x00 padding shorter keys out to the length of the
24
longest key. Leaf nodes are packed as densely as possible, and internal nodes
25
are all an additional 8-bits wide leading to a sparse upper tree.
27
Updates to a CHKMap are done preferentially via the apply_delta method, to
28
allow optimisation of the update operation; but individual map/unmap calls are
29
possible and supported. Individual changes via map/unmap are buffered in memory
30
until the _save method is called to force serialisation of the tree.
31
apply_delta records its changes immediately by performing an implicit _save.
36
Densely packed upper nodes.
43
from bzrlib import lazy_import
44
lazy_import.lazy_import(globals(), """
57
from bzrlib.static_tuple import StaticTuple
60
# If each line is 50 bytes, and you have 255 internal pages, with 255-way fan
61
# out, it takes 3.1MB to cache the layer.
62
_PAGE_CACHE_SIZE = 4*1024*1024
63
# Per thread caches for 2 reasons:
64
# - in the server we may be serving very different content, so we get less
66
# - we avoid locking on every cache lookup.
67
_thread_caches = threading.local()
69
_thread_caches.page_cache = None
72
"""Get the per-thread page cache.
74
We need a function to do this because in a new thread the _thread_caches
75
threading.local object does not have the cache initialized yet.
77
page_cache = getattr(_thread_caches, 'page_cache', None)
78
if page_cache is None:
79
# We are caching bytes so len(value) is perfectly accurate
80
page_cache = lru_cache.LRUSizeCache(_PAGE_CACHE_SIZE)
81
_thread_caches.page_cache = page_cache
89
# If a ChildNode falls below this many bytes, we check for a remap
90
_INTERESTING_NEW_SIZE = 50
91
# If a ChildNode shrinks by more than this amount, we check for a remap
92
_INTERESTING_SHRINKAGE_LIMIT = 20
93
# If we delete more than this many nodes applying a delta, we check for a remap
94
_INTERESTING_DELETES_LIMIT = 5
97
def _search_key_plain(key):
98
"""Map the key tuple into a search string that just uses the key bytes."""
99
return '\x00'.join(key)
102
search_key_registry = registry.Registry()
103
search_key_registry.register('plain', _search_key_plain)
106
class CHKMap(object):
107
"""A persistent map from string to string backed by a CHK store."""
109
__slots__ = ('_store', '_root_node', '_search_key_func')
111
def __init__(self, store, root_key, search_key_func=None):
112
"""Create a CHKMap object.
114
:param store: The store the CHKMap is stored in.
115
:param root_key: The root key of the map. None to create an empty
117
:param search_key_func: A function mapping a key => bytes. These bytes
118
are then used by the internal nodes to split up leaf nodes into
122
if search_key_func is None:
123
search_key_func = _search_key_plain
124
self._search_key_func = search_key_func
126
self._root_node = LeafNode(search_key_func=search_key_func)
128
self._root_node = self._node_key(root_key)
130
def apply_delta(self, delta):
131
"""Apply a delta to the map.
133
:param delta: An iterable of old_key, new_key, new_value tuples.
134
If new_key is not None, then new_key->new_value is inserted
135
into the map; if old_key is not None, then the old mapping
136
of old_key is removed.
139
# Check preconditions first.
140
as_st = StaticTuple.from_sequence
141
new_items = set([as_st(key) for (old, key, value) in delta
142
if key is not None and old is None])
143
existing_new = list(self.iteritems(key_filter=new_items))
145
raise errors.InconsistentDeltaDelta(delta,
146
"New items are already in the map %r." % existing_new)
148
for old, new, value in delta:
149
if old is not None and old != new:
150
self.unmap(old, check_remap=False)
152
for old, new, value in delta:
155
if delete_count > _INTERESTING_DELETES_LIMIT:
156
trace.mutter("checking remap as %d deletions", delete_count)
160
def _ensure_root(self):
161
"""Ensure that the root node is an object not a key."""
162
if type(self._root_node) is StaticTuple:
163
# Demand-load the root
164
self._root_node = self._get_node(self._root_node)
166
def _get_node(self, node):
169
Note that this does not update the _items dict in objects containing a
170
reference to this node. As such it does not prevent subsequent IO being
173
:param node: A tuple key or node object.
174
:return: A node object.
176
if type(node) is StaticTuple:
177
bytes = self._read_bytes(node)
178
return _deserialise(bytes, node,
179
search_key_func=self._search_key_func)
183
def _read_bytes(self, key):
185
return _get_cache()[key]
187
stream = self._store.get_record_stream([key], 'unordered', True)
188
bytes = stream.next().get_bytes_as('fulltext')
189
_get_cache()[key] = bytes
192
def _dump_tree(self, include_keys=False):
193
"""Return the tree in a string representation."""
195
res = self._dump_tree_node(self._root_node, prefix='', indent='',
196
include_keys=include_keys)
197
res.append('') # Give a trailing '\n'
198
return '\n'.join(res)
200
def _dump_tree_node(self, node, prefix, indent, include_keys=True):
201
"""For this node and all children, generate a string representation."""
206
node_key = node.key()
207
if node_key is not None:
208
key_str = ' %s' % (node_key[0],)
211
result.append('%s%r %s%s' % (indent, prefix, node.__class__.__name__,
213
if type(node) is InternalNode:
214
# Trigger all child nodes to get loaded
215
list(node._iter_nodes(self._store))
216
for prefix, sub in sorted(node._items.iteritems()):
217
result.extend(self._dump_tree_node(sub, prefix, indent + ' ',
218
include_keys=include_keys))
220
for key, value in sorted(node._items.iteritems()):
221
# Don't use prefix nor indent here to line up when used in
222
# tests in conjunction with assertEqualDiff
223
result.append(' %r %r' % (tuple(key), value))
227
def from_dict(klass, store, initial_value, maximum_size=0, key_width=1,
228
search_key_func=None):
229
"""Create a CHKMap in store with initial_value as the content.
231
:param store: The store to record initial_value in, a VersionedFiles
232
object with 1-tuple keys supporting CHK key generation.
233
:param initial_value: A dict to store in store. Its keys and values
235
:param maximum_size: The maximum_size rule to apply to nodes. This
236
determines the size at which no new data is added to a single node.
237
:param key_width: The number of elements in each key_tuple being stored
239
:param search_key_func: A function mapping a key => bytes. These bytes
240
are then used by the internal nodes to split up leaf nodes into
242
:return: The root chk of the resulting CHKMap.
244
root_key = klass._create_directly(store, initial_value,
245
maximum_size=maximum_size, key_width=key_width,
246
search_key_func=search_key_func)
247
if type(root_key) is not StaticTuple:
248
raise AssertionError('we got a %s instead of a StaticTuple'
253
def _create_via_map(klass, store, initial_value, maximum_size=0,
254
key_width=1, search_key_func=None):
255
result = klass(store, None, search_key_func=search_key_func)
256
result._root_node.set_maximum_size(maximum_size)
257
result._root_node._key_width = key_width
259
for key, value in initial_value.items():
260
delta.append((None, key, value))
261
root_key = result.apply_delta(delta)
265
def _create_directly(klass, store, initial_value, maximum_size=0,
266
key_width=1, search_key_func=None):
267
node = LeafNode(search_key_func=search_key_func)
268
node.set_maximum_size(maximum_size)
269
node._key_width = key_width
270
as_st = StaticTuple.from_sequence
271
node._items = dict([(as_st(key), val) for key, val
272
in initial_value.iteritems()])
273
node._raw_size = sum([node._key_value_len(key, value)
274
for key,value in node._items.iteritems()])
275
node._len = len(node._items)
276
node._compute_search_prefix()
277
node._compute_serialised_prefix()
280
and node._current_size() > maximum_size):
281
prefix, node_details = node._split(store)
282
if len(node_details) == 1:
283
raise AssertionError('Failed to split using node._split')
284
node = InternalNode(prefix, search_key_func=search_key_func)
285
node.set_maximum_size(maximum_size)
286
node._key_width = key_width
287
for split, subnode in node_details:
288
node.add_node(split, subnode)
289
keys = list(node.serialise(store))
292
def iter_changes(self, basis):
293
"""Iterate over the changes between basis and self.
295
:return: An iterator of tuples: (key, old_value, new_value). Old_value
296
is None for keys only in self; new_value is None for keys only in
300
# Read both trees in lexographic, highest-first order.
301
# Any identical nodes we skip
302
# Any unique prefixes we output immediately.
303
# values in a leaf node are treated as single-value nodes in the tree
304
# which allows them to be not-special-cased. We know to output them
305
# because their value is a string, not a key(tuple) or node.
307
# corner cases to beware of when considering this function:
308
# *) common references are at different heights.
309
# consider two trees:
310
# {'a': LeafNode={'aaa':'foo', 'aab':'bar'}, 'b': LeafNode={'b'}}
311
# {'a': InternalNode={'aa':LeafNode={'aaa':'foo', 'aab':'bar'},
312
# 'ab':LeafNode={'ab':'bar'}}
313
# 'b': LeafNode={'b'}}
314
# the node with aaa/aab will only be encountered in the second tree
315
# after reading the 'a' subtree, but it is encountered in the first
316
# tree immediately. Variations on this may have read internal nodes
317
# like this. we want to cut the entire pending subtree when we
318
# realise we have a common node. For this we use a list of keys -
319
# the path to a node - and check the entire path is clean as we
321
if self._node_key(self._root_node) == self._node_key(basis._root_node):
325
excluded_keys = set()
326
self_node = self._root_node
327
basis_node = basis._root_node
328
# A heap, each element is prefix, node(tuple/NodeObject/string),
329
# key_path (a list of tuples, tail-sharing down the tree.)
332
def process_node(node, path, a_map, pending):
333
# take a node and expand it
334
node = a_map._get_node(node)
335
if type(node) == LeafNode:
336
path = (node._key, path)
337
for key, value in node._items.items():
338
# For a LeafNode, the key is a serialized_key, rather than
339
# a search_key, but the heap is using search_keys
340
search_key = node._search_key_func(key)
341
heapq.heappush(pending, (search_key, key, value, path))
343
# type(node) == InternalNode
344
path = (node._key, path)
345
for prefix, child in node._items.items():
346
heapq.heappush(pending, (prefix, None, child, path))
347
def process_common_internal_nodes(self_node, basis_node):
348
self_items = set(self_node._items.items())
349
basis_items = set(basis_node._items.items())
350
path = (self_node._key, None)
351
for prefix, child in self_items - basis_items:
352
heapq.heappush(self_pending, (prefix, None, child, path))
353
path = (basis_node._key, None)
354
for prefix, child in basis_items - self_items:
355
heapq.heappush(basis_pending, (prefix, None, child, path))
356
def process_common_leaf_nodes(self_node, basis_node):
357
self_items = set(self_node._items.items())
358
basis_items = set(basis_node._items.items())
359
path = (self_node._key, None)
360
for key, value in self_items - basis_items:
361
prefix = self._search_key_func(key)
362
heapq.heappush(self_pending, (prefix, key, value, path))
363
path = (basis_node._key, None)
364
for key, value in basis_items - self_items:
365
prefix = basis._search_key_func(key)
366
heapq.heappush(basis_pending, (prefix, key, value, path))
367
def process_common_prefix_nodes(self_node, self_path,
368
basis_node, basis_path):
369
# Would it be more efficient if we could request both at the same
371
self_node = self._get_node(self_node)
372
basis_node = basis._get_node(basis_node)
373
if (type(self_node) == InternalNode
374
and type(basis_node) == InternalNode):
375
# Matching internal nodes
376
process_common_internal_nodes(self_node, basis_node)
377
elif (type(self_node) == LeafNode
378
and type(basis_node) == LeafNode):
379
process_common_leaf_nodes(self_node, basis_node)
381
process_node(self_node, self_path, self, self_pending)
382
process_node(basis_node, basis_path, basis, basis_pending)
383
process_common_prefix_nodes(self_node, None, basis_node, None)
386
excluded_keys = set()
387
def check_excluded(key_path):
388
# Note that this is N^2, it depends on us trimming trees
389
# aggressively to not become slow.
390
# A better implementation would probably have a reverse map
391
# back to the children of a node, and jump straight to it when
392
# a common node is detected, the proceed to remove the already
393
# pending children. bzrlib.graph has a searcher module with a
395
while key_path is not None:
396
key, key_path = key_path
397
if key in excluded_keys:
402
while self_pending or basis_pending:
405
# self is exhausted: output remainder of basis
406
for prefix, key, node, path in basis_pending:
407
if check_excluded(path):
409
node = basis._get_node(node)
412
yield (key, node, None)
414
# subtree - fastpath the entire thing.
415
for key, value in node.iteritems(basis._store):
416
yield (key, value, None)
418
elif not basis_pending:
419
# basis is exhausted: output remainder of self.
420
for prefix, key, node, path in self_pending:
421
if check_excluded(path):
423
node = self._get_node(node)
426
yield (key, None, node)
428
# subtree - fastpath the entire thing.
429
for key, value in node.iteritems(self._store):
430
yield (key, None, value)
433
# XXX: future optimisation - yield the smaller items
434
# immediately rather than pushing everything on/off the
435
# heaps. Applies to both internal nodes and leafnodes.
436
if self_pending[0][0] < basis_pending[0][0]:
438
prefix, key, node, path = heapq.heappop(self_pending)
439
if check_excluded(path):
443
yield (key, None, node)
445
process_node(node, path, self, self_pending)
447
elif self_pending[0][0] > basis_pending[0][0]:
449
prefix, key, node, path = heapq.heappop(basis_pending)
450
if check_excluded(path):
454
yield (key, node, None)
456
process_node(node, path, basis, basis_pending)
459
# common prefix: possibly expand both
460
if self_pending[0][1] is None:
465
if basis_pending[0][1] is None:
470
if not read_self and not read_basis:
471
# compare a common value
472
self_details = heapq.heappop(self_pending)
473
basis_details = heapq.heappop(basis_pending)
474
if self_details[2] != basis_details[2]:
475
yield (self_details[1],
476
basis_details[2], self_details[2])
478
# At least one side wasn't a simple value
479
if (self._node_key(self_pending[0][2]) ==
480
self._node_key(basis_pending[0][2])):
481
# Identical pointers, skip (and don't bother adding to
482
# excluded, it won't turn up again.
483
heapq.heappop(self_pending)
484
heapq.heappop(basis_pending)
486
# Now we need to expand this node before we can continue
487
if read_self and read_basis:
488
# Both sides start with the same prefix, so process
490
self_prefix, _, self_node, self_path = heapq.heappop(
492
basis_prefix, _, basis_node, basis_path = heapq.heappop(
494
if self_prefix != basis_prefix:
495
raise AssertionError(
496
'%r != %r' % (self_prefix, basis_prefix))
497
process_common_prefix_nodes(
498
self_node, self_path,
499
basis_node, basis_path)
502
prefix, key, node, path = heapq.heappop(self_pending)
503
if check_excluded(path):
505
process_node(node, path, self, self_pending)
507
prefix, key, node, path = heapq.heappop(basis_pending)
508
if check_excluded(path):
510
process_node(node, path, basis, basis_pending)
513
def iteritems(self, key_filter=None):
514
"""Iterate over the entire CHKMap's contents."""
516
if key_filter is not None:
517
as_st = StaticTuple.from_sequence
518
key_filter = [as_st(key) for key in key_filter]
519
return self._root_node.iteritems(self._store, key_filter=key_filter)
522
"""Return the key for this map."""
523
if type(self._root_node) is StaticTuple:
524
return self._root_node
526
return self._root_node._key
530
return len(self._root_node)
532
def map(self, key, value):
533
"""Map a key tuple to value.
535
:param key: A key to map.
536
:param value: The value to assign to key.
538
key = StaticTuple.from_sequence(key)
539
# Need a root object.
541
prefix, node_details = self._root_node.map(self._store, key, value)
542
if len(node_details) == 1:
543
self._root_node = node_details[0][1]
545
self._root_node = InternalNode(prefix,
546
search_key_func=self._search_key_func)
547
self._root_node.set_maximum_size(node_details[0][1].maximum_size)
548
self._root_node._key_width = node_details[0][1]._key_width
549
for split, node in node_details:
550
self._root_node.add_node(split, node)
552
def _node_key(self, node):
553
"""Get the key for a node whether it's a tuple or node."""
554
if type(node) is tuple:
555
node = StaticTuple.from_sequence(node)
556
if type(node) is StaticTuple:
561
def unmap(self, key, check_remap=True):
562
"""remove key from the map."""
563
key = StaticTuple.from_sequence(key)
565
if type(self._root_node) is InternalNode:
566
unmapped = self._root_node.unmap(self._store, key,
567
check_remap=check_remap)
569
unmapped = self._root_node.unmap(self._store, key)
570
self._root_node = unmapped
572
def _check_remap(self):
573
"""Check if nodes can be collapsed."""
575
if type(self._root_node) is InternalNode:
576
self._root_node._check_remap(self._store)
579
"""Save the map completely.
581
:return: The key of the root node.
583
if type(self._root_node) is StaticTuple:
585
return self._root_node
586
keys = list(self._root_node.serialise(self._store))
591
"""Base class defining the protocol for CHK Map nodes.
593
:ivar _raw_size: The total size of the serialized key:value data, before
594
adding the header bytes, and without prefix compression.
597
__slots__ = ('_key', '_len', '_maximum_size', '_key_width',
598
'_raw_size', '_items', '_search_prefix', '_search_key_func'
601
def __init__(self, key_width=1):
604
:param key_width: The width of keys for this node.
607
# Current number of elements
609
self._maximum_size = 0
610
self._key_width = key_width
611
# current size in bytes
613
# The pointers/values this node has - meaning defined by child classes.
615
# The common search prefix
616
self._search_prefix = None
619
items_str = str(sorted(self._items))
620
if len(items_str) > 20:
621
items_str = items_str[:16] + '...]'
622
return '%s(key:%s len:%s size:%s max:%s prefix:%s items:%s)' % (
623
self.__class__.__name__, self._key, self._len, self._raw_size,
624
self._maximum_size, self._search_prefix, items_str)
633
def maximum_size(self):
634
"""What is the upper limit for adding references to a node."""
635
return self._maximum_size
637
def set_maximum_size(self, new_size):
638
"""Set the size threshold for nodes.
640
:param new_size: The size at which no data is added to a node. 0 for
643
self._maximum_size = new_size
646
def common_prefix(cls, prefix, key):
647
"""Given 2 strings, return the longest prefix common to both.
649
:param prefix: This has been the common prefix for other keys, so it is
650
more likely to be the common prefix in this case as well.
651
:param key: Another string to compare to
653
if key.startswith(prefix):
656
# Is there a better way to do this?
657
for pos, (left, right) in enumerate(zip(prefix, key)):
661
common = prefix[:pos+1]
665
def common_prefix_for_keys(cls, keys):
666
"""Given a list of keys, find their common prefix.
668
:param keys: An iterable of strings.
669
:return: The longest common prefix of all keys.
673
if common_prefix is None:
676
common_prefix = cls.common_prefix(common_prefix, key)
677
if not common_prefix:
678
# if common_prefix is the empty string, then we know it won't
684
# Singleton indicating we have not computed _search_prefix yet
687
class LeafNode(Node):
688
"""A node containing actual key:value pairs.
690
:ivar _items: A dict of key->value items. The key is in tuple form.
691
:ivar _size: The number of bytes that would be used by serializing all of
695
__slots__ = ('_common_serialised_prefix', '_serialise_key')
697
def __init__(self, search_key_func=None):
699
# All of the keys in this leaf node share this common prefix
700
self._common_serialised_prefix = None
701
self._serialise_key = '\x00'.join
702
if search_key_func is None:
703
self._search_key_func = _search_key_plain
705
self._search_key_func = search_key_func
708
items_str = str(sorted(self._items))
709
if len(items_str) > 20:
710
items_str = items_str[:16] + '...]'
712
'%s(key:%s len:%s size:%s max:%s prefix:%s keywidth:%s items:%s)' \
713
% (self.__class__.__name__, self._key, self._len, self._raw_size,
714
self._maximum_size, self._search_prefix, self._key_width, items_str)
716
def _current_size(self):
717
"""Answer the current serialised size of this node.
719
This differs from self._raw_size in that it includes the bytes used for
722
if self._common_serialised_prefix is None:
726
# We will store a single string with the common prefix
727
# And then that common prefix will not be stored in any of the
729
prefix_len = len(self._common_serialised_prefix)
730
bytes_for_items = (self._raw_size - (prefix_len * self._len))
731
return (9 # 'chkleaf:\n'
732
+ len(str(self._maximum_size)) + 1
733
+ len(str(self._key_width)) + 1
734
+ len(str(self._len)) + 1
739
def deserialise(klass, bytes, key, search_key_func=None):
740
"""Deserialise bytes, with key key, into a LeafNode.
742
:param bytes: The bytes of the node.
743
:param key: The key that the serialised node has.
745
key = static_tuple.expect_static_tuple(key)
746
return _deserialise_leaf_node(bytes, key,
747
search_key_func=search_key_func)
749
def iteritems(self, store, key_filter=None):
750
"""Iterate over items in the node.
752
:param key_filter: A filter to apply to the node. It should be a
753
list/set/dict or similar repeatedly iterable container.
755
if key_filter is not None:
756
# Adjust the filter - short elements go to a prefix filter. All
757
# other items are looked up directly.
758
# XXX: perhaps defaultdict? Profiling<rinse and repeat>
760
for key in key_filter:
761
if len(key) == self._key_width:
762
# This filter is meant to match exactly one key, yield it
765
yield key, self._items[key]
767
# This key is not present in this map, continue
770
# Short items, we need to match based on a prefix
771
length_filter = filters.setdefault(len(key), set())
772
length_filter.add(key)
774
filters = filters.items()
775
for item in self._items.iteritems():
776
for length, length_filter in filters:
777
if item[0][:length] in length_filter:
781
for item in self._items.iteritems():
784
def _key_value_len(self, key, value):
785
# TODO: Should probably be done without actually joining the key, but
786
# then that can be done via the C extension
787
return (len(self._serialise_key(key)) + 1
788
+ len(str(value.count('\n'))) + 1
791
def _search_key(self, key):
792
return self._search_key_func(key)
794
def _map_no_split(self, key, value):
795
"""Map a key to a value.
797
This assumes either the key does not already exist, or you have already
798
removed its size and length from self.
800
:return: True if adding this node should cause us to split.
802
self._items[key] = value
803
self._raw_size += self._key_value_len(key, value)
805
serialised_key = self._serialise_key(key)
806
if self._common_serialised_prefix is None:
807
self._common_serialised_prefix = serialised_key
809
self._common_serialised_prefix = self.common_prefix(
810
self._common_serialised_prefix, serialised_key)
811
search_key = self._search_key(key)
812
if self._search_prefix is _unknown:
813
self._compute_search_prefix()
814
if self._search_prefix is None:
815
self._search_prefix = search_key
817
self._search_prefix = self.common_prefix(
818
self._search_prefix, search_key)
820
and self._maximum_size
821
and self._current_size() > self._maximum_size):
822
# Check to see if all of the search_keys for this node are
823
# identical. We allow the node to grow under that circumstance
824
# (we could track this as common state, but it is infrequent)
825
if (search_key != self._search_prefix
826
or not self._are_search_keys_identical()):
830
def _split(self, store):
831
"""We have overflowed.
833
Split this node into multiple LeafNodes, return it up the stack so that
834
the next layer creates a new InternalNode and references the new nodes.
836
:return: (common_serialised_prefix, [(node_serialised_prefix, node)])
838
if self._search_prefix is _unknown:
839
raise AssertionError('Search prefix must be known')
840
common_prefix = self._search_prefix
841
split_at = len(common_prefix) + 1
843
for key, value in self._items.iteritems():
844
search_key = self._search_key(key)
845
prefix = search_key[:split_at]
846
# TODO: Generally only 1 key can be exactly the right length,
847
# which means we can only have 1 key in the node pointed
848
# at by the 'prefix\0' key. We might want to consider
849
# folding it into the containing InternalNode rather than
850
# having a fixed length-1 node.
851
# Note this is probably not true for hash keys, as they
852
# may get a '\00' node anywhere, but won't have keys of
854
if len(prefix) < split_at:
855
prefix += '\x00'*(split_at - len(prefix))
856
if prefix not in result:
857
node = LeafNode(search_key_func=self._search_key_func)
858
node.set_maximum_size(self._maximum_size)
859
node._key_width = self._key_width
860
result[prefix] = node
862
node = result[prefix]
863
sub_prefix, node_details = node.map(store, key, value)
864
if len(node_details) > 1:
865
if prefix != sub_prefix:
866
# This node has been split and is now found via a different
869
new_node = InternalNode(sub_prefix,
870
search_key_func=self._search_key_func)
871
new_node.set_maximum_size(self._maximum_size)
872
new_node._key_width = self._key_width
873
for split, node in node_details:
874
new_node.add_node(split, node)
875
result[prefix] = new_node
876
return common_prefix, result.items()
878
def map(self, store, key, value):
879
"""Map key to value."""
880
if key in self._items:
881
self._raw_size -= self._key_value_len(key, self._items[key])
884
if self._map_no_split(key, value):
885
return self._split(store)
887
if self._search_prefix is _unknown:
888
raise AssertionError('%r must be known' % self._search_prefix)
889
return self._search_prefix, [("", self)]
891
def serialise(self, store):
892
"""Serialise the LeafNode to store.
894
:param store: A VersionedFiles honouring the CHK extensions.
895
:return: An iterable of the keys inserted by this operation.
897
lines = ["chkleaf:\n"]
898
lines.append("%d\n" % self._maximum_size)
899
lines.append("%d\n" % self._key_width)
900
lines.append("%d\n" % self._len)
901
if self._common_serialised_prefix is None:
903
if len(self._items) != 0:
904
raise AssertionError('If _common_serialised_prefix is None'
905
' we should have no items')
907
lines.append('%s\n' % (self._common_serialised_prefix,))
908
prefix_len = len(self._common_serialised_prefix)
909
for key, value in sorted(self._items.items()):
910
# Always add a final newline
911
value_lines = osutils.chunks_to_lines([value + '\n'])
912
serialized = "%s\x00%s\n" % (self._serialise_key(key),
914
if not serialized.startswith(self._common_serialised_prefix):
915
raise AssertionError('We thought the common prefix was %r'
916
' but entry %r does not have it in common'
917
% (self._common_serialised_prefix, serialized))
918
lines.append(serialized[prefix_len:])
919
lines.extend(value_lines)
920
sha1, _, _ = store.add_lines((None,), (), lines)
921
self._key = StaticTuple("sha1:" + sha1,).intern()
922
bytes = ''.join(lines)
923
if len(bytes) != self._current_size():
924
raise AssertionError('Invalid _current_size')
925
_get_cache().add(self._key, bytes)
929
"""Return the references to other CHK's held by this node."""
932
def _compute_search_prefix(self):
933
"""Determine the common search prefix for all keys in this node.
935
:return: A bytestring of the longest search key prefix that is
936
unique within this node.
938
search_keys = [self._search_key_func(key) for key in self._items]
939
self._search_prefix = self.common_prefix_for_keys(search_keys)
940
return self._search_prefix
942
def _are_search_keys_identical(self):
943
"""Check to see if the search keys for all entries are the same.
945
When using a hash as the search_key it is possible for non-identical
946
keys to collide. If that happens enough, we may try overflow a
947
LeafNode, but as all are collisions, we must not split.
949
common_search_key = None
950
for key in self._items:
951
search_key = self._search_key(key)
952
if common_search_key is None:
953
common_search_key = search_key
954
elif search_key != common_search_key:
958
def _compute_serialised_prefix(self):
959
"""Determine the common prefix for serialised keys in this node.
961
:return: A bytestring of the longest serialised key prefix that is
962
unique within this node.
964
serialised_keys = [self._serialise_key(key) for key in self._items]
965
self._common_serialised_prefix = self.common_prefix_for_keys(
967
return self._common_serialised_prefix
969
def unmap(self, store, key):
970
"""Unmap key from the node."""
972
self._raw_size -= self._key_value_len(key, self._items[key])
974
trace.mutter("key %s not found in %r", key, self._items)
979
# Recompute from scratch
980
self._compute_search_prefix()
981
self._compute_serialised_prefix()
985
class InternalNode(Node):
986
"""A node that contains references to other nodes.
988
An InternalNode is responsible for mapping search key prefixes to child
991
:ivar _items: serialised_key => node dictionary. node may be a tuple,
992
LeafNode or InternalNode.
995
__slots__ = ('_node_width',)
997
def __init__(self, prefix='', search_key_func=None):
999
# The size of an internalnode with default values and no children.
1000
# How many octets key prefixes within this node are.
1001
self._node_width = 0
1002
self._search_prefix = prefix
1003
if search_key_func is None:
1004
self._search_key_func = _search_key_plain
1006
self._search_key_func = search_key_func
1008
def add_node(self, prefix, node):
1009
"""Add a child node with prefix prefix, and node node.
1011
:param prefix: The search key prefix for node.
1012
:param node: The node being added.
1014
if self._search_prefix is None:
1015
raise AssertionError("_search_prefix should not be None")
1016
if not prefix.startswith(self._search_prefix):
1017
raise AssertionError("prefixes mismatch: %s must start with %s"
1018
% (prefix,self._search_prefix))
1019
if len(prefix) != len(self._search_prefix) + 1:
1020
raise AssertionError("prefix wrong length: len(%s) is not %d" %
1021
(prefix, len(self._search_prefix) + 1))
1022
self._len += len(node)
1023
if not len(self._items):
1024
self._node_width = len(prefix)
1025
if self._node_width != len(self._search_prefix) + 1:
1026
raise AssertionError("node width mismatch: %d is not %d" %
1027
(self._node_width, len(self._search_prefix) + 1))
1028
self._items[prefix] = node
1031
def _current_size(self):
1032
"""Answer the current serialised size of this node."""
1033
return (self._raw_size + len(str(self._len)) + len(str(self._key_width)) +
1034
len(str(self._maximum_size)))
1037
def deserialise(klass, bytes, key, search_key_func=None):
1038
"""Deserialise bytes to an InternalNode, with key key.
1040
:param bytes: The bytes of the node.
1041
:param key: The key that the serialised node has.
1042
:return: An InternalNode instance.
1044
key = static_tuple.expect_static_tuple(key)
1045
return _deserialise_internal_node(bytes, key,
1046
search_key_func=search_key_func)
1048
def iteritems(self, store, key_filter=None):
1049
for node, node_filter in self._iter_nodes(store, key_filter=key_filter):
1050
for item in node.iteritems(store, key_filter=node_filter):
1053
def _iter_nodes(self, store, key_filter=None, batch_size=None):
1054
"""Iterate over node objects which match key_filter.
1056
:param store: A store to use for accessing content.
1057
:param key_filter: A key filter to filter nodes. Only nodes that might
1058
contain a key in key_filter will be returned.
1059
:param batch_size: If not None, then we will return the nodes that had
1060
to be read using get_record_stream in batches, rather than reading
1062
:return: An iterable of nodes. This function does not have to be fully
1063
consumed. (There will be no pending I/O when items are being returned.)
1065
# Map from chk key ('sha1:...',) to (prefix, key_filter)
1066
# prefix is the key in self._items to use, key_filter is the key_filter
1067
# entries that would match this node
1070
if key_filter is None:
1071
# yielding all nodes, yield whatever we have, and queue up a read
1072
# for whatever we are missing
1074
for prefix, node in self._items.iteritems():
1075
if node.__class__ is StaticTuple:
1076
keys[node] = (prefix, None)
1079
elif len(key_filter) == 1:
1080
# Technically, this path could also be handled by the first check
1081
# in 'self._node_width' in length_filters. However, we can handle
1082
# this case without spending any time building up the
1083
# prefix_to_keys, etc state.
1085
# This is a bit ugly, but TIMEIT showed it to be by far the fastest
1086
# 0.626us list(key_filter)[0]
1087
# is a func() for list(), 2 mallocs, and a getitem
1088
# 0.489us [k for k in key_filter][0]
1089
# still has the mallocs, avoids the func() call
1090
# 0.350us iter(key_filter).next()
1091
# has a func() call, and mallocs an iterator
1092
# 0.125us for key in key_filter: pass
1093
# no func() overhead, might malloc an iterator
1094
# 0.105us for key in key_filter: break
1095
# no func() overhead, might malloc an iterator, probably
1096
# avoids checking an 'else' clause as part of the for
1097
for key in key_filter:
1099
search_prefix = self._search_prefix_filter(key)
1100
if len(search_prefix) == self._node_width:
1101
# This item will match exactly, so just do a dict lookup, and
1102
# see what we can return
1105
node = self._items[search_prefix]
1107
# A given key can only match 1 child node, if it isn't
1108
# there, then we can just return nothing
1110
if node.__class__ is StaticTuple:
1111
keys[node] = (search_prefix, [key])
1113
# This is loaded, and the only thing that can match,
1118
# First, convert all keys into a list of search prefixes
1119
# Aggregate common prefixes, and track the keys they come from
1122
for key in key_filter:
1123
search_prefix = self._search_prefix_filter(key)
1124
length_filter = length_filters.setdefault(
1125
len(search_prefix), set())
1126
length_filter.add(search_prefix)
1127
prefix_to_keys.setdefault(search_prefix, []).append(key)
1129
if (self._node_width in length_filters
1130
and len(length_filters) == 1):
1131
# all of the search prefixes match exactly _node_width. This
1132
# means that everything is an exact match, and we can do a
1133
# lookup into self._items, rather than iterating over the items
1135
search_prefixes = length_filters[self._node_width]
1136
for search_prefix in search_prefixes:
1138
node = self._items[search_prefix]
1140
# We can ignore this one
1142
node_key_filter = prefix_to_keys[search_prefix]
1143
if node.__class__ is StaticTuple:
1144
keys[node] = (search_prefix, node_key_filter)
1146
yield node, node_key_filter
1148
# The slow way. We walk every item in self._items, and check to
1149
# see if there are any matches
1150
length_filters = length_filters.items()
1151
for prefix, node in self._items.iteritems():
1152
node_key_filter = []
1153
for length, length_filter in length_filters:
1154
sub_prefix = prefix[:length]
1155
if sub_prefix in length_filter:
1156
node_key_filter.extend(prefix_to_keys[sub_prefix])
1157
if node_key_filter: # this key matched something, yield it
1158
if node.__class__ is StaticTuple:
1159
keys[node] = (prefix, node_key_filter)
1161
yield node, node_key_filter
1163
# Look in the page cache for some more bytes
1167
bytes = _get_cache()[key]
1171
node = _deserialise(bytes, key,
1172
search_key_func=self._search_key_func)
1173
prefix, node_key_filter = keys[key]
1174
self._items[prefix] = node
1176
yield node, node_key_filter
1177
for key in found_keys:
1180
# demand load some pages.
1181
if batch_size is None:
1182
# Read all the keys in
1183
batch_size = len(keys)
1184
key_order = list(keys)
1185
for batch_start in range(0, len(key_order), batch_size):
1186
batch = key_order[batch_start:batch_start + batch_size]
1187
# We have to fully consume the stream so there is no pending
1188
# I/O, so we buffer the nodes for now.
1189
stream = store.get_record_stream(batch, 'unordered', True)
1190
node_and_filters = []
1191
for record in stream:
1192
bytes = record.get_bytes_as('fulltext')
1193
node = _deserialise(bytes, record.key,
1194
search_key_func=self._search_key_func)
1195
prefix, node_key_filter = keys[record.key]
1196
node_and_filters.append((node, node_key_filter))
1197
self._items[prefix] = node
1198
_get_cache().add(record.key, bytes)
1199
for info in node_and_filters:
1202
def map(self, store, key, value):
1203
"""Map key to value."""
1204
if not len(self._items):
1205
raise AssertionError("can't map in an empty InternalNode.")
1206
search_key = self._search_key(key)
1207
if self._node_width != len(self._search_prefix) + 1:
1208
raise AssertionError("node width mismatch: %d is not %d" %
1209
(self._node_width, len(self._search_prefix) + 1))
1210
if not search_key.startswith(self._search_prefix):
1211
# This key doesn't fit in this index, so we need to split at the
1212
# point where it would fit, insert self into that internal node,
1213
# and then map this key into that node.
1214
new_prefix = self.common_prefix(self._search_prefix,
1216
new_parent = InternalNode(new_prefix,
1217
search_key_func=self._search_key_func)
1218
new_parent.set_maximum_size(self._maximum_size)
1219
new_parent._key_width = self._key_width
1220
new_parent.add_node(self._search_prefix[:len(new_prefix)+1],
1222
return new_parent.map(store, key, value)
1223
children = [node for node, _
1224
in self._iter_nodes(store, key_filter=[key])]
1229
child = self._new_child(search_key, LeafNode)
1230
old_len = len(child)
1231
if type(child) is LeafNode:
1232
old_size = child._current_size()
1235
prefix, node_details = child.map(store, key, value)
1236
if len(node_details) == 1:
1237
# child may have shrunk, or might be a new node
1238
child = node_details[0][1]
1239
self._len = self._len - old_len + len(child)
1240
self._items[search_key] = child
1243
if type(child) is LeafNode:
1244
if old_size is None:
1245
# The old node was an InternalNode which means it has now
1246
# collapsed, so we need to check if it will chain to a
1247
# collapse at this level.
1248
trace.mutter("checking remap as InternalNode -> LeafNode")
1249
new_node = self._check_remap(store)
1251
# If the LeafNode has shrunk in size, we may want to run
1252
# a remap check. Checking for a remap is expensive though
1253
# and the frequency of a successful remap is very low.
1254
# Shrinkage by small amounts is common, so we only do the
1255
# remap check if the new_size is low or the shrinkage
1256
# amount is over a configurable limit.
1257
new_size = child._current_size()
1258
shrinkage = old_size - new_size
1259
if (shrinkage > 0 and new_size < _INTERESTING_NEW_SIZE
1260
or shrinkage > _INTERESTING_SHRINKAGE_LIMIT):
1262
"checking remap as size shrunk by %d to be %d",
1263
shrinkage, new_size)
1264
new_node = self._check_remap(store)
1265
if new_node._search_prefix is None:
1266
raise AssertionError("_search_prefix should not be None")
1267
return new_node._search_prefix, [('', new_node)]
1268
# child has overflown - create a new intermediate node.
1269
# XXX: This is where we might want to try and expand our depth
1270
# to refer to more bytes of every child (which would give us
1271
# multiple pointers to child nodes, but less intermediate nodes)
1272
child = self._new_child(search_key, InternalNode)
1273
child._search_prefix = prefix
1274
for split, node in node_details:
1275
child.add_node(split, node)
1276
self._len = self._len - old_len + len(child)
1278
return self._search_prefix, [("", self)]
1280
def _new_child(self, search_key, klass):
1281
"""Create a new child node of type klass."""
1283
child.set_maximum_size(self._maximum_size)
1284
child._key_width = self._key_width
1285
child._search_key_func = self._search_key_func
1286
self._items[search_key] = child
1289
def serialise(self, store):
1290
"""Serialise the node to store.
1292
:param store: A VersionedFiles honouring the CHK extensions.
1293
:return: An iterable of the keys inserted by this operation.
1295
for node in self._items.itervalues():
1296
if type(node) is StaticTuple:
1297
# Never deserialised.
1299
if node._key is not None:
1302
for key in node.serialise(store):
1304
lines = ["chknode:\n"]
1305
lines.append("%d\n" % self._maximum_size)
1306
lines.append("%d\n" % self._key_width)
1307
lines.append("%d\n" % self._len)
1308
if self._search_prefix is None:
1309
raise AssertionError("_search_prefix should not be None")
1310
lines.append('%s\n' % (self._search_prefix,))
1311
prefix_len = len(self._search_prefix)
1312
for prefix, node in sorted(self._items.items()):
1313
if type(node) is StaticTuple:
1317
serialised = "%s\x00%s\n" % (prefix, key)
1318
if not serialised.startswith(self._search_prefix):
1319
raise AssertionError("prefixes mismatch: %s must start with %s"
1320
% (serialised, self._search_prefix))
1321
lines.append(serialised[prefix_len:])
1322
sha1, _, _ = store.add_lines((None,), (), lines)
1323
self._key = StaticTuple("sha1:" + sha1,).intern()
1324
_get_cache().add(self._key, ''.join(lines))
1327
def _search_key(self, key):
1328
"""Return the serialised key for key in this node."""
1329
# search keys are fixed width. All will be self._node_width wide, so we
1331
return (self._search_key_func(key) + '\x00'*self._node_width)[:self._node_width]
1333
def _search_prefix_filter(self, key):
1334
"""Serialise key for use as a prefix filter in iteritems."""
1335
return self._search_key_func(key)[:self._node_width]
1337
def _split(self, offset):
1338
"""Split this node into smaller nodes starting at offset.
1340
:param offset: The offset to start the new child nodes at.
1341
:return: An iterable of (prefix, node) tuples. prefix is a byte
1342
prefix for reaching node.
1344
if offset >= self._node_width:
1345
for node in self._items.values():
1346
for result in node._split(offset):
1349
for key, node in self._items.items():
1353
"""Return the references to other CHK's held by this node."""
1354
if self._key is None:
1355
raise AssertionError("unserialised nodes have no refs.")
1357
for value in self._items.itervalues():
1358
if type(value) is StaticTuple:
1361
refs.append(value.key())
1364
def _compute_search_prefix(self, extra_key=None):
1365
"""Return the unique key prefix for this node.
1367
:return: A bytestring of the longest search key prefix that is
1368
unique within this node.
1370
self._search_prefix = self.common_prefix_for_keys(self._items)
1371
return self._search_prefix
1373
def unmap(self, store, key, check_remap=True):
1374
"""Remove key from this node and it's children."""
1375
if not len(self._items):
1376
raise AssertionError("can't unmap in an empty InternalNode.")
1377
children = [node for node, _
1378
in self._iter_nodes(store, key_filter=[key])]
1384
unmapped = child.unmap(store, key)
1386
search_key = self._search_key(key)
1387
if len(unmapped) == 0:
1388
# All child nodes are gone, remove the child:
1389
del self._items[search_key]
1392
# Stash the returned node
1393
self._items[search_key] = unmapped
1394
if len(self._items) == 1:
1395
# this node is no longer needed:
1396
return self._items.values()[0]
1397
if type(unmapped) is InternalNode:
1400
return self._check_remap(store)
1404
def _check_remap(self, store):
1405
"""Check if all keys contained by children fit in a single LeafNode.
1407
:param store: A store to use for reading more nodes
1408
:return: Either self, or a new LeafNode which should replace self.
1410
# Logic for how we determine when we need to rebuild
1411
# 1) Implicitly unmap() is removing a key which means that the child
1412
# nodes are going to be shrinking by some extent.
1413
# 2) If all children are LeafNodes, it is possible that they could be
1414
# combined into a single LeafNode, which can then completely replace
1415
# this internal node with a single LeafNode
1416
# 3) If *one* child is an InternalNode, we assume it has already done
1417
# all the work to determine that its children cannot collapse, and
1418
# we can then assume that those nodes *plus* the current nodes don't
1419
# have a chance of collapsing either.
1420
# So a very cheap check is to just say if 'unmapped' is an
1421
# InternalNode, we don't have to check further.
1423
# TODO: Another alternative is to check the total size of all known
1424
# LeafNodes. If there is some formula we can use to determine the
1425
# final size without actually having to read in any more
1426
# children, it would be nice to have. However, we have to be
1427
# careful with stuff like nodes that pull out the common prefix
1428
# of each key, as adding a new key can change the common prefix
1429
# and cause size changes greater than the length of one key.
1430
# So for now, we just add everything to a new Leaf until it
1431
# splits, as we know that will give the right answer
1432
new_leaf = LeafNode(search_key_func=self._search_key_func)
1433
new_leaf.set_maximum_size(self._maximum_size)
1434
new_leaf._key_width = self._key_width
1435
# A batch_size of 16 was chosen because:
1436
# a) In testing, a 4k page held 14 times. So if we have more than 16
1437
# leaf nodes we are unlikely to hold them in a single new leaf
1438
# node. This still allows for 1 round trip
1439
# b) With 16-way fan out, we can still do a single round trip
1440
# c) With 255-way fan out, we don't want to read all 255 and destroy
1441
# the page cache, just to determine that we really don't need it.
1442
for node, _ in self._iter_nodes(store, batch_size=16):
1443
if type(node) is InternalNode:
1444
# Without looking at any leaf nodes, we are sure
1446
for key, value in node._items.iteritems():
1447
if new_leaf._map_no_split(key, value):
1449
trace.mutter("remap generated a new LeafNode")
1453
def _deserialise(bytes, key, search_key_func):
1454
"""Helper for repositorydetails - convert bytes to a node."""
1455
if bytes.startswith("chkleaf:\n"):
1456
node = LeafNode.deserialise(bytes, key, search_key_func=search_key_func)
1457
elif bytes.startswith("chknode:\n"):
1458
node = InternalNode.deserialise(bytes, key,
1459
search_key_func=search_key_func)
1461
raise AssertionError("Unknown node type.")
1465
class CHKMapDifference(object):
1466
"""Iterate the stored pages and key,value pairs for (new - old).
1468
This class provides a generator over the stored CHK pages and the
1469
(key, value) pairs that are in any of the new maps and not in any of the
1472
Note that it may yield chk pages that are common (especially root nodes),
1473
but it won't yield (key,value) pairs that are common.
1476
def __init__(self, store, new_root_keys, old_root_keys,
1477
search_key_func, pb=None):
1478
# TODO: Should we add a StaticTuple barrier here? It would be nice to
1479
# force callers to use StaticTuple, because there will often be
1480
# lots of keys passed in here. And even if we cast it locally,
1481
# that just meanst that we will have *both* a StaticTuple and a
1482
# tuple() in memory, referring to the same object. (so a net
1483
# increase in memory, not a decrease.)
1485
self._new_root_keys = new_root_keys
1486
self._old_root_keys = old_root_keys
1488
# All uninteresting chks that we have seen. By the time they are added
1489
# here, they should be either fully ignored, or queued up for
1491
# TODO: This might grow to a large size if there are lots of merge
1492
# parents, etc. However, it probably doesn't scale to O(history)
1493
# like _processed_new_refs does.
1494
self._all_old_chks = set(self._old_root_keys)
1495
# All items that we have seen from the old_root_keys
1496
self._all_old_items = set()
1497
# These are interesting items which were either read, or already in the
1498
# interesting queue (so we don't need to walk them again)
1499
# TODO: processed_new_refs becomes O(all_chks), consider switching to
1501
self._processed_new_refs = set()
1502
self._search_key_func = search_key_func
1504
# The uninteresting and interesting nodes to be searched
1505
self._old_queue = []
1506
self._new_queue = []
1507
# Holds the (key, value) items found when processing the root nodes,
1508
# waiting for the uninteresting nodes to be walked
1509
self._new_item_queue = []
1512
def _read_nodes_from_store(self, keys):
1513
# We chose not to use _get_cache(), because we think in
1514
# terms of records to be yielded. Also, we expect to touch each page
1515
# only 1 time during this code. (We may want to evaluate saving the
1516
# raw bytes into the page cache, which would allow a working tree
1517
# update after the fetch to not have to read the bytes again.)
1518
as_st = StaticTuple.from_sequence
1519
stream = self._store.get_record_stream(keys, 'unordered', True)
1520
for record in stream:
1521
if self._pb is not None:
1523
if record.storage_kind == 'absent':
1524
raise errors.NoSuchRevision(self._store, record.key)
1525
bytes = record.get_bytes_as('fulltext')
1526
node = _deserialise(bytes, record.key,
1527
search_key_func=self._search_key_func)
1528
if type(node) is InternalNode:
1529
# Note we don't have to do node.refs() because we know that
1530
# there are no children that have been pushed into this node
1531
# Note: Using as_st() here seemed to save 1.2MB, which would
1532
# indicate that we keep 100k prefix_refs around while
1533
# processing. They *should* be shorter lived than that...
1534
# It does cost us ~10s of processing time
1535
#prefix_refs = [as_st(item) for item in node._items.iteritems()]
1536
prefix_refs = node._items.items()
1540
# Note: We don't use a StaticTuple here. Profiling showed a
1541
# minor memory improvement (0.8MB out of 335MB peak 0.2%)
1542
# But a significant slowdown (15s / 145s, or 10%)
1543
items = node._items.items()
1544
yield record, node, prefix_refs, items
1546
def _read_old_roots(self):
1547
old_chks_to_enqueue = []
1548
all_old_chks = self._all_old_chks
1549
for record, node, prefix_refs, items in \
1550
self._read_nodes_from_store(self._old_root_keys):
1551
# Uninteresting node
1552
prefix_refs = [p_r for p_r in prefix_refs
1553
if p_r[1] not in all_old_chks]
1554
new_refs = [p_r[1] for p_r in prefix_refs]
1555
all_old_chks.update(new_refs)
1556
# TODO: This might be a good time to turn items into StaticTuple
1557
# instances and possibly intern them. However, this does not
1558
# impact 'initial branch' performance, so I'm not worrying
1560
self._all_old_items.update(items)
1561
# Queue up the uninteresting references
1562
# Don't actually put them in the 'to-read' queue until we have
1563
# finished checking the interesting references
1564
old_chks_to_enqueue.extend(prefix_refs)
1565
return old_chks_to_enqueue
1567
def _enqueue_old(self, new_prefixes, old_chks_to_enqueue):
1568
# At this point, we have read all the uninteresting and interesting
1569
# items, so we can queue up the uninteresting stuff, knowing that we've
1570
# handled the interesting ones
1571
for prefix, ref in old_chks_to_enqueue:
1572
not_interesting = True
1573
for i in xrange(len(prefix), 0, -1):
1574
if prefix[:i] in new_prefixes:
1575
not_interesting = False
1578
# This prefix is not part of the remaining 'interesting set'
1580
self._old_queue.append(ref)
1582
def _read_all_roots(self):
1583
"""Read the root pages.
1585
This is structured as a generator, so that the root records can be
1586
yielded up to whoever needs them without any buffering.
1588
# This is the bootstrap phase
1589
if not self._old_root_keys:
1590
# With no old_root_keys we can just shortcut and be ready
1591
# for _flush_new_queue
1592
self._new_queue = list(self._new_root_keys)
1594
old_chks_to_enqueue = self._read_old_roots()
1595
# filter out any root keys that are already known to be uninteresting
1596
new_keys = set(self._new_root_keys).difference(self._all_old_chks)
1597
# These are prefixes that are present in new_keys that we are
1599
new_prefixes = set()
1600
# We are about to yield all of these, so we don't want them getting
1601
# added a second time
1602
processed_new_refs = self._processed_new_refs
1603
processed_new_refs.update(new_keys)
1604
for record, node, prefix_refs, items in \
1605
self._read_nodes_from_store(new_keys):
1606
# At this level, we now know all the uninteresting references
1607
# So we filter and queue up whatever is remaining
1608
prefix_refs = [p_r for p_r in prefix_refs
1609
if p_r[1] not in self._all_old_chks
1610
and p_r[1] not in processed_new_refs]
1611
refs = [p_r[1] for p_r in prefix_refs]
1612
new_prefixes.update([p_r[0] for p_r in prefix_refs])
1613
self._new_queue.extend(refs)
1614
# TODO: We can potentially get multiple items here, however the
1615
# current design allows for this, as callers will do the work
1616
# to make the results unique. We might profile whether we
1617
# gain anything by ensuring unique return values for items
1618
# TODO: This might be a good time to cast to StaticTuple, as
1619
# self._new_item_queue will hold the contents of multiple
1620
# records for an extended lifetime
1621
new_items = [item for item in items
1622
if item not in self._all_old_items]
1623
self._new_item_queue.extend(new_items)
1624
new_prefixes.update([self._search_key_func(item[0])
1625
for item in new_items])
1626
processed_new_refs.update(refs)
1628
# For new_prefixes we have the full length prefixes queued up.
1629
# However, we also need possible prefixes. (If we have a known ref to
1630
# 'ab', then we also need to include 'a'.) So expand the
1631
# new_prefixes to include all shorter prefixes
1632
for prefix in list(new_prefixes):
1633
new_prefixes.update([prefix[:i] for i in xrange(1, len(prefix))])
1634
self._enqueue_old(new_prefixes, old_chks_to_enqueue)
1636
def _flush_new_queue(self):
1637
# No need to maintain the heap invariant anymore, just pull things out
1639
refs = set(self._new_queue)
1640
self._new_queue = []
1641
# First pass, flush all interesting items and convert to using direct refs
1642
all_old_chks = self._all_old_chks
1643
processed_new_refs = self._processed_new_refs
1644
all_old_items = self._all_old_items
1645
new_items = [item for item in self._new_item_queue
1646
if item not in all_old_items]
1647
self._new_item_queue = []
1649
yield None, new_items
1650
refs = refs.difference(all_old_chks)
1651
processed_new_refs.update(refs)
1653
# TODO: Using a SimpleSet for self._processed_new_refs and
1654
# saved as much as 10MB of peak memory. However, it requires
1655
# implementing a non-pyrex version.
1657
next_refs_update = next_refs.update
1658
# Inlining _read_nodes_from_store improves 'bzr branch bzr.dev'
1659
# from 1m54s to 1m51s. Consider it.
1660
for record, _, p_refs, items in self._read_nodes_from_store(refs):
1662
# using the 'if' check saves about 145s => 141s, when
1663
# streaming initial branch of Launchpad data.
1664
items = [item for item in items
1665
if item not in all_old_items]
1667
next_refs_update([p_r[1] for p_r in p_refs])
1669
# set1.difference(set/dict) walks all of set1, and checks if it
1670
# exists in 'other'.
1671
# set1.difference(iterable) walks all of iterable, and does a
1672
# 'difference_update' on a clone of set1. Pick wisely based on the
1673
# expected sizes of objects.
1674
# in our case it is expected that 'new_refs' will always be quite
1676
next_refs = next_refs.difference(all_old_chks)
1677
next_refs = next_refs.difference(processed_new_refs)
1678
processed_new_refs.update(next_refs)
1681
def _process_next_old(self):
1682
# Since we don't filter uninteresting any further than during
1683
# _read_all_roots, process the whole queue in a single pass.
1684
refs = self._old_queue
1685
self._old_queue = []
1686
all_old_chks = self._all_old_chks
1687
for record, _, prefix_refs, items in self._read_nodes_from_store(refs):
1688
# TODO: Use StaticTuple here?
1689
self._all_old_items.update(items)
1690
refs = [r for _,r in prefix_refs if r not in all_old_chks]
1691
self._old_queue.extend(refs)
1692
all_old_chks.update(refs)
1694
def _process_queues(self):
1695
while self._old_queue:
1696
self._process_next_old()
1697
return self._flush_new_queue()
1700
for record in self._read_all_roots():
1702
for record, items in self._process_queues():
1706
def iter_interesting_nodes(store, interesting_root_keys,
1707
uninteresting_root_keys, pb=None):
1708
"""Given root keys, find interesting nodes.
1710
Evaluate nodes referenced by interesting_root_keys. Ones that are also
1711
referenced from uninteresting_root_keys are not considered interesting.
1713
:param interesting_root_keys: keys which should be part of the
1714
"interesting" nodes (which will be yielded)
1715
:param uninteresting_root_keys: keys which should be filtered out of the
1718
(interesting record, {interesting key:values})
1720
iterator = CHKMapDifference(store, interesting_root_keys,
1721
uninteresting_root_keys,
1722
search_key_func=store._search_key_func,
1724
return iterator.process()
1728
from bzrlib._chk_map_pyx import (
1731
_deserialise_leaf_node,
1732
_deserialise_internal_node,
1734
except ImportError, e:
1735
osutils.failed_to_load_extension(e)
1736
from bzrlib._chk_map_py import (
1739
_deserialise_leaf_node,
1740
_deserialise_internal_node,
1742
search_key_registry.register('hash-16-way', _search_key_16)
1743
search_key_registry.register('hash-255-way', _search_key_255)
1746
def _check_key(key):
1747
"""Helper function to assert that a key is properly formatted.
1749
This generally shouldn't be used in production code, but it can be helpful
1752
if type(key) is not StaticTuple:
1753
raise TypeError('key %r is not StaticTuple but %s' % (key, type(key)))
1755
raise ValueError('key %r should have length 1, not %d' % (key, len(key),))
1756
if type(key[0]) is not str:
1757
raise TypeError('key %r should hold a str, not %r'
1758
% (key, type(key[0])))
1759
if not key[0].startswith('sha1:'):
1760
raise ValueError('key %r should point to a sha1:' % (key,))
added
removed
bzrlib/chk_map.py
