1
# Copyright (C) 2008-2011 Canonical Ltd
1
# Copyright (C) 2008, 2009 Canonical Ltd
3
3
# This program is free software; you can redistribute it and/or modify
4
4
# it under the terms of the GNU General Public License as published by
40
from __future__ import absolute_import
45
42
from bzrlib import lazy_import
46
43
lazy_import.lazy_import(globals(), """
47
44
from bzrlib import (
51
49
from bzrlib import (
59
from bzrlib.static_tuple import StaticTuple
62
57
# If each line is 50 bytes, and you have 255 internal pages, with 255-way fan
63
58
# out, it takes 3.1MB to cache the layer.
64
59
_PAGE_CACHE_SIZE = 4*1024*1024
65
# Per thread caches for 2 reasons:
66
# - in the server we may be serving very different content, so we get less
68
# - we avoid locking on every cache lookup.
69
_thread_caches = threading.local()
71
_thread_caches.page_cache = None
74
"""Get the per-thread page cache.
76
We need a function to do this because in a new thread the _thread_caches
77
threading.local object does not have the cache initialized yet.
79
page_cache = getattr(_thread_caches, 'page_cache', None)
80
if page_cache is None:
81
# We are caching bytes so len(value) is perfectly accurate
82
page_cache = lru_cache.LRUSizeCache(_PAGE_CACHE_SIZE)
83
_thread_caches.page_cache = page_cache
60
# We are caching bytes so len(value) is perfectly accurate
61
_page_cache = lru_cache.LRUSizeCache(_PAGE_CACHE_SIZE)
91
66
# If a ChildNode falls below this many bytes, we check for a remap
92
67
_INTERESTING_NEW_SIZE = 50
93
68
# If a ChildNode shrinks by more than this amount, we check for a remap
94
69
_INTERESTING_SHRINKAGE_LIMIT = 20
70
# If we delete more than this many nodes applying a delta, we check for a remap
71
_INTERESTING_DELETES_LIMIT = 5
97
74
def _search_key_plain(key):
106
83
class CHKMap(object):
107
84
"""A persistent map from string to string backed by a CHK store."""
109
__slots__ = ('_store', '_root_node', '_search_key_func')
111
86
def __init__(self, store, root_key, search_key_func=None):
112
87
"""Create a CHKMap object.
135
110
into the map; if old_key is not None, then the old mapping
136
111
of old_key is removed.
139
114
# 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])
115
new_items = set([key for (old, key, value) in delta if key is not None
143
117
existing_new = list(self.iteritems(key_filter=new_items))
145
119
raise errors.InconsistentDeltaDelta(delta,
148
122
for old, new, value in delta:
149
123
if old is not None and old != new:
150
124
self.unmap(old, check_remap=False)
152
126
for old, new, value in delta:
153
127
if new is not None:
154
128
self.map(new, value)
129
if delete_count > _INTERESTING_DELETES_LIMIT:
130
trace.mutter("checking remap as %d deletions", delete_count)
156
131
self._check_remap()
157
132
return self._save()
159
134
def _ensure_root(self):
160
135
"""Ensure that the root node is an object not a key."""
161
if type(self._root_node) is StaticTuple:
136
if type(self._root_node) is tuple:
162
137
# Demand-load the root
163
138
self._root_node = self._get_node(self._root_node)
172
147
:param node: A tuple key or node object.
173
148
:return: A node object.
175
if type(node) is StaticTuple:
150
if type(node) is tuple:
176
151
bytes = self._read_bytes(node)
177
152
return _deserialise(bytes, node,
178
153
search_key_func=self._search_key_func)
182
157
def _read_bytes(self, key):
184
return _get_cache()[key]
159
return _page_cache[key]
186
161
stream = self._store.get_record_stream([key], 'unordered', True)
187
162
bytes = stream.next().get_bytes_as('fulltext')
188
_get_cache()[key] = bytes
163
_page_cache[key] = bytes
191
166
def _dump_tree(self, include_keys=False):
219
194
for key, value in sorted(node._items.iteritems()):
220
195
# Don't use prefix nor indent here to line up when used in
221
196
# tests in conjunction with assertEqualDiff
222
result.append(' %r %r' % (tuple(key), value))
197
result.append(' %r %r' % (key, value))
243
218
root_key = klass._create_directly(store, initial_value,
244
219
maximum_size=maximum_size, key_width=key_width,
245
220
search_key_func=search_key_func)
246
if type(root_key) is not StaticTuple:
247
raise AssertionError('we got a %s instead of a StaticTuple'
266
238
node = LeafNode(search_key_func=search_key_func)
267
239
node.set_maximum_size(maximum_size)
268
240
node._key_width = key_width
269
as_st = StaticTuple.from_sequence
270
node._items = dict([(as_st(key), val) for key, val
271
in initial_value.iteritems()])
241
node._items = dict(initial_value)
272
242
node._raw_size = sum([node._key_value_len(key, value)
273
for key,value in node._items.iteritems()])
243
for key,value in initial_value.iteritems()])
274
244
node._len = len(node._items)
275
245
node._compute_search_prefix()
276
246
node._compute_serialised_prefix()
512
482
def iteritems(self, key_filter=None):
513
483
"""Iterate over the entire CHKMap's contents."""
514
484
self._ensure_root()
515
if key_filter is not None:
516
as_st = StaticTuple.from_sequence
517
key_filter = [as_st(key) for key in key_filter]
518
485
return self._root_node.iteritems(self._store, key_filter=key_filter)
521
488
"""Return the key for this map."""
522
if type(self._root_node) is StaticTuple:
489
if type(self._root_node) is tuple:
523
490
return self._root_node
525
492
return self._root_node._key
534
501
:param key: A key to map.
535
502
:param value: The value to assign to key.
537
key = StaticTuple.from_sequence(key)
538
504
# Need a root object.
539
505
self._ensure_root()
540
506
prefix, node_details = self._root_node.map(self._store, key, value)
551
517
def _node_key(self, node):
552
518
"""Get the key for a node whether it's a tuple or node."""
553
519
if type(node) is tuple:
554
node = StaticTuple.from_sequence(node)
555
if type(node) is StaticTuple:
560
524
def unmap(self, key, check_remap=True):
561
525
"""remove key from the map."""
562
key = StaticTuple.from_sequence(key)
563
526
self._ensure_root()
564
527
if type(self._root_node) is InternalNode:
565
528
unmapped = self._root_node.unmap(self._store, key,
572
535
"""Check if nodes can be collapsed."""
573
536
self._ensure_root()
574
537
if type(self._root_node) is InternalNode:
575
self._root_node = self._root_node._check_remap(self._store)
538
self._root_node._check_remap(self._store)
578
541
"""Save the map completely.
580
543
:return: The key of the root node.
582
if type(self._root_node) is StaticTuple:
545
if type(self._root_node) is tuple:
584
547
return self._root_node
585
548
keys = list(self._root_node.serialise(self._store))
593
556
adding the header bytes, and without prefix compression.
596
__slots__ = ('_key', '_len', '_maximum_size', '_key_width',
597
'_raw_size', '_items', '_search_prefix', '_search_key_func'
600
559
def __init__(self, key_width=1):
601
560
"""Create a node.
691
650
the key/value pairs.
694
__slots__ = ('_common_serialised_prefix',)
696
653
def __init__(self, search_key_func=None):
697
654
Node.__init__(self)
698
655
# All of the keys in this leaf node share this common prefix
699
656
self._common_serialised_prefix = None
657
self._serialise_key = '\x00'.join
700
658
if search_key_func is None:
701
659
self._search_key_func = _search_key_plain
740
698
:param bytes: The bytes of the node.
741
699
:param key: The key that the serialised node has.
743
key = static_tuple.expect_static_tuple(key)
744
701
return _deserialise_leaf_node(bytes, key,
745
702
search_key_func=search_key_func)
886
843
raise AssertionError('%r must be known' % self._search_prefix)
887
844
return self._search_prefix, [("", self)]
889
_serialise_key = '\x00'.join
891
846
def serialise(self, store):
892
847
"""Serialise the LeafNode to store.
918
873
lines.append(serialized[prefix_len:])
919
874
lines.extend(value_lines)
920
875
sha1, _, _ = store.add_lines((None,), (), lines)
921
self._key = StaticTuple("sha1:" + sha1,).intern()
876
self._key = ("sha1:" + sha1,)
922
877
bytes = ''.join(lines)
923
878
if len(bytes) != self._current_size():
924
879
raise AssertionError('Invalid _current_size')
925
_get_cache()[self._key] = bytes
880
_page_cache.add(self._key, bytes)
926
881
return [self._key]
992
947
LeafNode or InternalNode.
995
__slots__ = ('_node_width',)
997
950
def __init__(self, prefix='', search_key_func=None):
998
951
Node.__init__(self)
999
952
# The size of an internalnode with default values and no children.
1041
994
:param key: The key that the serialised node has.
1042
995
:return: An InternalNode instance.
1044
key = static_tuple.expect_static_tuple(key)
1045
997
return _deserialise_internal_node(bytes, key,
1046
998
search_key_func=search_key_func)
1072
1024
# for whatever we are missing
1073
1025
shortcut = True
1074
1026
for prefix, node in self._items.iteritems():
1075
if node.__class__ is StaticTuple:
1027
if node.__class__ is tuple:
1076
1028
keys[node] = (prefix, None)
1078
1030
yield node, None
1107
1059
# A given key can only match 1 child node, if it isn't
1108
1060
# there, then we can just return nothing
1110
if node.__class__ is StaticTuple:
1062
if node.__class__ is tuple:
1111
1063
keys[node] = (search_prefix, [key])
1113
1065
# This is loaded, and the only thing that can match,
1140
1092
# We can ignore this one
1142
1094
node_key_filter = prefix_to_keys[search_prefix]
1143
if node.__class__ is StaticTuple:
1095
if node.__class__ is tuple:
1144
1096
keys[node] = (search_prefix, node_key_filter)
1146
1098
yield node, node_key_filter
1155
1107
if sub_prefix in length_filter:
1156
1108
node_key_filter.extend(prefix_to_keys[sub_prefix])
1157
1109
if node_key_filter: # this key matched something, yield it
1158
if node.__class__ is StaticTuple:
1110
if node.__class__ is tuple:
1159
1111
keys[node] = (prefix, node_key_filter)
1161
1113
yield node, node_key_filter
1195
1147
prefix, node_key_filter = keys[record.key]
1196
1148
node_and_filters.append((node, node_key_filter))
1197
1149
self._items[prefix] = node
1198
_get_cache()[record.key] = bytes
1150
_page_cache.add(record.key, bytes)
1199
1151
for info in node_and_filters:
1310
1262
lines.append('%s\n' % (self._search_prefix,))
1311
1263
prefix_len = len(self._search_prefix)
1312
1264
for prefix, node in sorted(self._items.items()):
1313
if type(node) is StaticTuple:
1265
if type(node) is tuple:
1316
1268
key = node._key[0]
1320
1272
% (serialised, self._search_prefix))
1321
1273
lines.append(serialised[prefix_len:])
1322
1274
sha1, _, _ = store.add_lines((None,), (), lines)
1323
self._key = StaticTuple("sha1:" + sha1,).intern()
1324
_get_cache()[self._key] = ''.join(lines)
1275
self._key = ("sha1:" + sha1,)
1276
_page_cache.add(self._key, ''.join(lines))
1325
1277
yield self._key
1327
1279
def _search_key(self, key):
1371
1323
return self._search_prefix
1373
1325
def unmap(self, store, key, check_remap=True):
1374
"""Remove key from this node and its children."""
1326
"""Remove key from this node and it's children."""
1375
1327
if not len(self._items):
1376
1328
raise AssertionError("can't unmap in an empty InternalNode.")
1377
1329
children = [node for node, _
1476
1428
def __init__(self, store, new_root_keys, old_root_keys,
1477
1429
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.)
1484
1430
self._store = store
1485
1431
self._new_root_keys = new_root_keys
1486
1432
self._old_root_keys = old_root_keys
1488
1434
# All uninteresting chks that we have seen. By the time they are added
1489
1435
# 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
1437
self._all_old_chks = set(self._old_root_keys)
1495
1438
# All items that we have seen from the old_root_keys
1496
1439
self._all_old_items = set()
1497
1440
# These are interesting items which were either read, or already in the
1498
1441
# interesting queue (so we don't need to walk them again)
1499
# TODO: processed_new_refs becomes O(all_chks), consider switching to
1501
1442
self._processed_new_refs = set()
1502
1443
self._search_key_func = search_key_func
1510
1451
self._state = None
1512
1453
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
1454
# We chose not to use _page_cache, because we think in terms of records
1455
# to be yielded. Also, we expect to touch each page only 1 time during
1456
# this code. (We may want to evaluate saving the raw bytes into the
1457
# page cache, which would allow a working tree update after the fetch
1458
# to not have to read the bytes again.)
1519
1459
stream = self._store.get_record_stream(keys, 'unordered', True)
1520
1460
for record in stream:
1521
1461
if self._pb is not None:
1528
1468
if type(node) is InternalNode:
1529
1469
# Note we don't have to do node.refs() because we know that
1530
1470
# 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
1471
prefix_refs = node._items.items()
1539
1474
prefix_refs = []
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
1475
items = node._items.items()
1544
1476
yield record, node, prefix_refs, items
1553
1485
if p_r[1] not in all_old_chks]
1554
1486
new_refs = [p_r[1] for p_r in prefix_refs]
1555
1487
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
1488
self._all_old_items.update(items)
1561
1489
# Queue up the uninteresting references
1562
1490
# Don't actually put them in the 'to-read' queue until we have
1615
1543
# current design allows for this, as callers will do the work
1616
1544
# to make the results unique. We might profile whether we
1617
1545
# 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
1546
new_items = [item for item in items
1622
1547
if item not in self._all_old_items]
1623
1548
self._new_item_queue.extend(new_items)
1649
1574
yield None, new_items
1650
1575
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.
1656
1577
next_refs = set()
1657
1578
next_refs_update = next_refs.update
1658
1579
# Inlining _read_nodes_from_store improves 'bzr branch bzr.dev'
1659
1580
# from 1m54s to 1m51s. Consider it.
1660
1581
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]
1582
items = [item for item in items
1583
if item not in all_old_items]
1666
1584
yield record, items
1667
1585
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
1586
next_refs = next_refs.difference(all_old_chks)
1677
1587
next_refs = next_refs.difference(processed_new_refs)
1678
1588
processed_new_refs.update(next_refs)
1685
1595
self._old_queue = []
1686
1596
all_old_chks = self._all_old_chks
1687
1597
for record, _, prefix_refs, items in self._read_nodes_from_store(refs):
1688
# TODO: Use StaticTuple here?
1689
1598
self._all_old_items.update(items)
1690
1599
refs = [r for _,r in prefix_refs if r not in all_old_chks]
1691
1600
self._old_queue.extend(refs)
1744
1651
search_key_registry.register('hash-16-way', _search_key_16)
1745
1652
search_key_registry.register('hash-255-way', _search_key_255)
1748
def _check_key(key):
1749
"""Helper function to assert that a key is properly formatted.
1751
This generally shouldn't be used in production code, but it can be helpful
1754
if type(key) is not StaticTuple:
1755
raise TypeError('key %r is not StaticTuple but %s' % (key, type(key)))
1757
raise ValueError('key %r should have length 1, not %d' % (key, len(key),))
1758
if type(key[0]) is not str:
1759
raise TypeError('key %r should hold a str, not %r'
1760
% (key, type(key[0])))
1761
if not key[0].startswith('sha1:'):
1762
raise ValueError('key %r should point to a sha1:' % (key,))
added
removed
bzrlib/chk_map.py
