1
# Copyright (C) 2008, 2009 Canonical Ltd
1
# Copyright (C) 2008, 2009, 2010 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
57
from bzrlib.static_tuple import StaticTuple
57
60
# If each line is 50 bytes, and you have 255 internal pages, with 255-way fan
58
61
# out, it takes 3.1MB to cache the layer.
59
62
_PAGE_CACHE_SIZE = 4*1024*1024
60
# We are caching bytes so len(value) is perfectly accurate
61
_page_cache = lru_cache.LRUSizeCache(_PAGE_CACHE_SIZE)
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
63
89
# If a ChildNode falls below this many bytes, we check for a remap
64
90
_INTERESTING_NEW_SIZE = 50
80
106
class CHKMap(object):
81
107
"""A persistent map from string to string backed by a CHK store."""
109
__slots__ = ('_store', '_root_node', '_search_key_func')
83
111
def __init__(self, store, root_key, search_key_func=None):
84
112
"""Create a CHKMap object.
111
139
# Check preconditions first.
112
new_items = set([key for (old, key, value) in delta if key is not None
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])
114
143
existing_new = list(self.iteritems(key_filter=new_items))
116
145
raise errors.InconsistentDeltaDelta(delta,
144
173
:param node: A tuple key or node object.
145
174
:return: A node object.
147
if type(node) is tuple:
176
if type(node) is StaticTuple:
148
177
bytes = self._read_bytes(node)
149
178
return _deserialise(bytes, node,
150
179
search_key_func=self._search_key_func)
154
183
def _read_bytes(self, key):
156
return _page_cache[key]
185
return _get_cache()[key]
158
187
stream = self._store.get_record_stream([key], 'unordered', True)
159
188
bytes = stream.next().get_bytes_as('fulltext')
160
_page_cache[key] = bytes
189
_get_cache()[key] = bytes
163
192
def _dump_tree(self, include_keys=False):
191
220
for key, value in sorted(node._items.iteritems()):
192
221
# Don't use prefix nor indent here to line up when used in
193
222
# tests in conjunction with assertEqualDiff
194
result.append(' %r %r' % (key, value))
223
result.append(' %r %r' % (tuple(key), value))
215
244
root_key = klass._create_directly(store, initial_value,
216
245
maximum_size=maximum_size, key_width=key_width,
217
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'
235
267
node = LeafNode(search_key_func=search_key_func)
236
268
node.set_maximum_size(maximum_size)
237
269
node._key_width = key_width
238
node._items = dict(initial_value)
270
as_st = StaticTuple.from_sequence
271
node._items = dict([(as_st(key), val) for key, val
272
in initial_value.iteritems()])
239
273
node._raw_size = sum([node._key_value_len(key, value)
240
for key,value in initial_value.iteritems()])
274
for key,value in node._items.iteritems()])
241
275
node._len = len(node._items)
242
276
node._compute_search_prefix()
243
277
node._compute_serialised_prefix()
479
513
def iteritems(self, key_filter=None):
480
514
"""Iterate over the entire CHKMap's contents."""
481
515
self._ensure_root()
516
if key_filter is not None:
517
as_st = StaticTuple.from_sequence
518
key_filter = [as_st(key) for key in key_filter]
482
519
return self._root_node.iteritems(self._store, key_filter=key_filter)
485
522
"""Return the key for this map."""
486
if type(self._root_node) is tuple:
523
if type(self._root_node) is StaticTuple:
487
524
return self._root_node
489
526
return self._root_node._key
498
535
:param key: A key to map.
499
536
:param value: The value to assign to key.
538
key = StaticTuple.from_sequence(key)
501
539
# Need a root object.
502
540
self._ensure_root()
503
541
prefix, node_details = self._root_node.map(self._store, key, value)
514
552
def _node_key(self, node):
515
553
"""Get the key for a node whether it's a tuple or node."""
516
554
if type(node) is tuple:
555
node = StaticTuple.from_sequence(node)
556
if type(node) is StaticTuple:
521
561
def unmap(self, key, check_remap=True):
522
562
"""remove key from the map."""
563
key = StaticTuple.from_sequence(key)
523
564
self._ensure_root()
524
565
if type(self._root_node) is InternalNode:
525
566
unmapped = self._root_node.unmap(self._store, key,
553
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'
556
601
def __init__(self, key_width=1):
557
602
"""Create a node.
647
692
the key/value pairs.
695
__slots__ = ('_common_serialised_prefix',)
650
697
def __init__(self, search_key_func=None):
651
698
Node.__init__(self)
652
699
# All of the keys in this leaf node share this common prefix
653
700
self._common_serialised_prefix = None
654
self._serialise_key = '\x00'.join
655
701
if search_key_func is None:
656
702
self._search_key_func = _search_key_plain
695
741
:param bytes: The bytes of the node.
696
742
:param key: The key that the serialised node has.
744
key = static_tuple.expect_static_tuple(key)
698
745
return _deserialise_leaf_node(bytes, key,
699
746
search_key_func=search_key_func)
840
887
raise AssertionError('%r must be known' % self._search_prefix)
841
888
return self._search_prefix, [("", self)]
890
_serialise_key = '\x00'.join
843
892
def serialise(self, store):
844
893
"""Serialise the LeafNode to store.
870
919
lines.append(serialized[prefix_len:])
871
920
lines.extend(value_lines)
872
921
sha1, _, _ = store.add_lines((None,), (), lines)
873
self._key = ("sha1:" + sha1,)
922
self._key = StaticTuple("sha1:" + sha1,).intern()
874
923
bytes = ''.join(lines)
875
924
if len(bytes) != self._current_size():
876
925
raise AssertionError('Invalid _current_size')
877
_page_cache.add(self._key, bytes)
926
_get_cache().add(self._key, bytes)
878
927
return [self._key]
944
993
LeafNode or InternalNode.
996
__slots__ = ('_node_width',)
947
998
def __init__(self, prefix='', search_key_func=None):
948
999
Node.__init__(self)
949
1000
# The size of an internalnode with default values and no children.
991
1042
:param key: The key that the serialised node has.
992
1043
:return: An InternalNode instance.
1045
key = static_tuple.expect_static_tuple(key)
994
1046
return _deserialise_internal_node(bytes, key,
995
1047
search_key_func=search_key_func)
1021
1073
# for whatever we are missing
1022
1074
shortcut = True
1023
1075
for prefix, node in self._items.iteritems():
1024
if node.__class__ is tuple:
1076
if node.__class__ is StaticTuple:
1025
1077
keys[node] = (prefix, None)
1027
1079
yield node, None
1056
1108
# A given key can only match 1 child node, if it isn't
1057
1109
# there, then we can just return nothing
1059
if node.__class__ is tuple:
1111
if node.__class__ is StaticTuple:
1060
1112
keys[node] = (search_prefix, [key])
1062
1114
# This is loaded, and the only thing that can match,
1089
1141
# We can ignore this one
1091
1143
node_key_filter = prefix_to_keys[search_prefix]
1092
if node.__class__ is tuple:
1144
if node.__class__ is StaticTuple:
1093
1145
keys[node] = (search_prefix, node_key_filter)
1095
1147
yield node, node_key_filter
1104
1156
if sub_prefix in length_filter:
1105
1157
node_key_filter.extend(prefix_to_keys[sub_prefix])
1106
1158
if node_key_filter: # this key matched something, yield it
1107
if node.__class__ is tuple:
1159
if node.__class__ is StaticTuple:
1108
1160
keys[node] = (prefix, node_key_filter)
1110
1162
yield node, node_key_filter
1144
1196
prefix, node_key_filter = keys[record.key]
1145
1197
node_and_filters.append((node, node_key_filter))
1146
1198
self._items[prefix] = node
1147
_page_cache.add(record.key, bytes)
1199
_get_cache().add(record.key, bytes)
1148
1200
for info in node_and_filters:
1259
1311
lines.append('%s\n' % (self._search_prefix,))
1260
1312
prefix_len = len(self._search_prefix)
1261
1313
for prefix, node in sorted(self._items.items()):
1262
if type(node) is tuple:
1314
if type(node) is StaticTuple:
1265
1317
key = node._key[0]
1269
1321
% (serialised, self._search_prefix))
1270
1322
lines.append(serialised[prefix_len:])
1271
1323
sha1, _, _ = store.add_lines((None,), (), lines)
1272
self._key = ("sha1:" + sha1,)
1273
_page_cache.add(self._key, ''.join(lines))
1324
self._key = StaticTuple("sha1:" + sha1,).intern()
1325
_get_cache().add(self._key, ''.join(lines))
1274
1326
yield self._key
1276
1328
def _search_key(self, key):
1425
1477
def __init__(self, store, new_root_keys, old_root_keys,
1426
1478
search_key_func, pb=None):
1479
# TODO: Should we add a StaticTuple barrier here? It would be nice to
1480
# force callers to use StaticTuple, because there will often be
1481
# lots of keys passed in here. And even if we cast it locally,
1482
# that just meanst that we will have *both* a StaticTuple and a
1483
# tuple() in memory, referring to the same object. (so a net
1484
# increase in memory, not a decrease.)
1427
1485
self._store = store
1428
1486
self._new_root_keys = new_root_keys
1429
1487
self._old_root_keys = old_root_keys
1431
1489
# All uninteresting chks that we have seen. By the time they are added
1432
1490
# here, they should be either fully ignored, or queued up for
1492
# TODO: This might grow to a large size if there are lots of merge
1493
# parents, etc. However, it probably doesn't scale to O(history)
1494
# like _processed_new_refs does.
1434
1495
self._all_old_chks = set(self._old_root_keys)
1435
1496
# All items that we have seen from the old_root_keys
1436
1497
self._all_old_items = set()
1437
1498
# These are interesting items which were either read, or already in the
1438
1499
# interesting queue (so we don't need to walk them again)
1500
# TODO: processed_new_refs becomes O(all_chks), consider switching to
1439
1502
self._processed_new_refs = set()
1440
1503
self._search_key_func = search_key_func
1448
1511
self._state = None
1450
1513
def _read_nodes_from_store(self, keys):
1451
# We chose not to use _page_cache, because we think in terms of records
1452
# to be yielded. Also, we expect to touch each page only 1 time during
1453
# this code. (We may want to evaluate saving the raw bytes into the
1454
# page cache, which would allow a working tree update after the fetch
1455
# to not have to read the bytes again.)
1514
# We chose not to use _get_cache(), because we think in
1515
# terms of records to be yielded. Also, we expect to touch each page
1516
# only 1 time during this code. (We may want to evaluate saving the
1517
# raw bytes into the page cache, which would allow a working tree
1518
# update after the fetch to not have to read the bytes again.)
1519
as_st = StaticTuple.from_sequence
1456
1520
stream = self._store.get_record_stream(keys, 'unordered', True)
1457
1521
for record in stream:
1458
1522
if self._pb is not None:
1465
1529
if type(node) is InternalNode:
1466
1530
# Note we don't have to do node.refs() because we know that
1467
1531
# there are no children that have been pushed into this node
1532
# Note: Using as_st() here seemed to save 1.2MB, which would
1533
# indicate that we keep 100k prefix_refs around while
1534
# processing. They *should* be shorter lived than that...
1535
# It does cost us ~10s of processing time
1536
#prefix_refs = [as_st(item) for item in node._items.iteritems()]
1468
1537
prefix_refs = node._items.items()
1471
1540
prefix_refs = []
1541
# Note: We don't use a StaticTuple here. Profiling showed a
1542
# minor memory improvement (0.8MB out of 335MB peak 0.2%)
1543
# But a significant slowdown (15s / 145s, or 10%)
1472
1544
items = node._items.items()
1473
1545
yield record, node, prefix_refs, items
1482
1554
if p_r[1] not in all_old_chks]
1483
1555
new_refs = [p_r[1] for p_r in prefix_refs]
1484
1556
all_old_chks.update(new_refs)
1557
# TODO: This might be a good time to turn items into StaticTuple
1558
# instances and possibly intern them. However, this does not
1559
# impact 'initial branch' performance, so I'm not worrying
1485
1561
self._all_old_items.update(items)
1486
1562
# Queue up the uninteresting references
1487
1563
# Don't actually put them in the 'to-read' queue until we have
1540
1616
# current design allows for this, as callers will do the work
1541
1617
# to make the results unique. We might profile whether we
1542
1618
# gain anything by ensuring unique return values for items
1619
# TODO: This might be a good time to cast to StaticTuple, as
1620
# self._new_item_queue will hold the contents of multiple
1621
# records for an extended lifetime
1543
1622
new_items = [item for item in items
1544
1623
if item not in self._all_old_items]
1545
1624
self._new_item_queue.extend(new_items)
1571
1650
yield None, new_items
1572
1651
refs = refs.difference(all_old_chks)
1652
processed_new_refs.update(refs)
1654
# TODO: Using a SimpleSet for self._processed_new_refs and
1655
# saved as much as 10MB of peak memory. However, it requires
1656
# implementing a non-pyrex version.
1574
1657
next_refs = set()
1575
1658
next_refs_update = next_refs.update
1576
1659
# Inlining _read_nodes_from_store improves 'bzr branch bzr.dev'
1577
1660
# from 1m54s to 1m51s. Consider it.
1578
1661
for record, _, p_refs, items in self._read_nodes_from_store(refs):
1579
items = [item for item in items
1580
if item not in all_old_items]
1663
# using the 'if' check saves about 145s => 141s, when
1664
# streaming initial branch of Launchpad data.
1665
items = [item for item in items
1666
if item not in all_old_items]
1581
1667
yield record, items
1582
1668
next_refs_update([p_r[1] for p_r in p_refs])
1670
# set1.difference(set/dict) walks all of set1, and checks if it
1671
# exists in 'other'.
1672
# set1.difference(iterable) walks all of iterable, and does a
1673
# 'difference_update' on a clone of set1. Pick wisely based on the
1674
# expected sizes of objects.
1675
# in our case it is expected that 'new_refs' will always be quite
1583
1677
next_refs = next_refs.difference(all_old_chks)
1584
1678
next_refs = next_refs.difference(processed_new_refs)
1585
1679
processed_new_refs.update(next_refs)
1592
1686
self._old_queue = []
1593
1687
all_old_chks = self._all_old_chks
1594
1688
for record, _, prefix_refs, items in self._read_nodes_from_store(refs):
1689
# TODO: Use StaticTuple here?
1595
1690
self._all_old_items.update(items)
1596
1691
refs = [r for _,r in prefix_refs if r not in all_old_chks]
1597
1692
self._old_queue.extend(refs)
1634
1729
from bzrlib._chk_map_pyx import (
1635
1731
_search_key_16,
1636
1732
_search_key_255,
1637
1733
_deserialise_leaf_node,
1638
1734
_deserialise_internal_node,
1736
except ImportError, e:
1737
osutils.failed_to_load_extension(e)
1641
1738
from bzrlib._chk_map_py import (
1642
1740
_search_key_16,
1643
1741
_search_key_255,
1644
1742
_deserialise_leaf_node,
1647
1745
search_key_registry.register('hash-16-way', _search_key_16)
1648
1746
search_key_registry.register('hash-255-way', _search_key_255)
1749
def _check_key(key):
1750
"""Helper function to assert that a key is properly formatted.
1752
This generally shouldn't be used in production code, but it can be helpful
1755
if type(key) is not StaticTuple:
1756
raise TypeError('key %r is not StaticTuple but %s' % (key, type(key)))
1758
raise ValueError('key %r should have length 1, not %d' % (key, len(key),))
1759
if type(key[0]) is not str:
1760
raise TypeError('key %r should hold a str, not %r'
1761
% (key, type(key[0])))
1762
if not key[0].startswith('sha1:'):
1763
raise ValueError('key %r should point to a sha1:' % (key,))
added
removed
bzrlib/chk_map.py
