source: trunk/packages/invirt-base/python/invirt/config.py @ 1499

Last change on this file since 1499 was 1423, checked in by price, 16 years ago

use run-parts semantics in conf.d directory

File size: 5.0 KB
Line 
1from __future__ import with_statement
2
3import json
4from invirt.common import *
5import os
6from os import rename
7from os.path import getmtime
8from contextlib import closing
9import yaml
10import re
11
12try:    loader = yaml.CSafeLoader
13except: loader = yaml.SafeLoader
14
15src_path    = '/etc/invirt/master.yaml'
16src_dirpath = '/etc/invirt/conf.d'
17cache_path  = '/var/lib/invirt/cache.json'
18lock_path   = '/var/lib/invirt/cache.lock'
19
20def augment(d1, d2):
21    """Splice dict-tree d2 into d1.  Return d1.
22
23    Example:
24    >>> d = {'a': {'b': 1}, 'c': 2}
25    >>> augment(d, {'a': {'d': 3}})
26    {'a': {'b', 1, 'd': 3}, 'c': 2}
27    >>> d
28    {'a': {'b', 1, 'd': 3}, 'c': 2}
29    """
30    for k in d2:
31        if k in d1 and isinstance(d1[k], dict):
32            augment(d1[k], d2[k])
33        else:
34            d1[k] = d2[k]
35    return d1
36
37def run_parts_list(dirname):
38    """Reimplements Debian's run-parts --list.
39
40    One difference from run-parts's behavior: run-parts --list /foo/
41    will give output like /foo//bar, because Python code tends to expect this.
42
43    Matches documented behavior of run-parts in debianutils v2.28.2, dated 2007.
44    """
45    # From run-parts(8).
46    lanana_re   = re.compile('^[a-z0-9]+$')
47    lsb_re      = re.compile('^_?([a-z0-9_.]+-)+[a-z0-9]+$')
48    deb_cron_re = re.compile('^[a-z0-9][a-z0-9-]*$')
49    for name in os.listdir(dirname):
50        if lanana_re.match(name) or lsb_re.match(name) or deb_cron_re.match(name):
51            yield os.path.join(dirname, name)
52
53def list_files():
54    yield src_path
55    for name in run_parts_list(src_dirpath):
56        yield name
57
58def load_master():
59    config = dict()
60    for filename in list_files():
61        with closing(file(filename)) as f:
62            augment(config, yaml.load(f, loader))
63    return config
64
65def get_src_mtime():
66    return max(max(getmtime(filename) for filename in list_files()),
67               getmtime(src_dirpath))
68
69def load(force_refresh = False):
70    """
71    Try loading the configuration from the faster-to-load JSON cache at
72    cache_path.  If it doesn't exist or is outdated, load the configuration
73    instead from the original YAML file at src_path and regenerate the cache.
74    I assume I have the permissions to write to the cache directory.
75    """
76
77    # Namespace container for state variables, so that they can be updated by
78    # closures.
79    ns = struct()
80
81    if force_refresh:
82        do_refresh = True
83    else:
84        src_mtime = get_src_mtime()
85        try:            cache_mtime = getmtime(cache_path)
86        except OSError: do_refresh  = True
87        else:           do_refresh  = src_mtime + 1 >= cache_mtime
88
89        # We chose not to simply say
90        #
91        #   do_refresh = src_mtime >= cache_time
92        #
93        # because between the getmtime(src_path) and the time the cache is
94        # rewritten, the master configuration may have been updated, so future
95        # checks here would find a cache with a newer mtime than the master
96        # (and thus treat the cache as containing the latest version of the
97        # master).  The +1 means that for at least a full second following the
98        # update to the master, this function will refresh the cache, giving us
99        # 1 second to write the cache.  Note that if it takes longer than 1
100        # second to write the cache, then this situation could still arise.
101        #
102        # The getmtime calls should logically be part of the same transaction
103        # as the rest of this function (cache read + conditional cache
104        # refresh), but to wrap everything in an flock would cause the
105        # following cache read to be less streamlined.
106
107    if not do_refresh:
108        # Try reading from the cache first.  This must be transactionally
109        # isolated from concurrent writes to prevent reading an incomplete
110        # (changing) version of the data (but the transaction can share the
111        # lock with other concurrent reads).  This isolation is accomplished
112        # using an atomic filesystem rename in the refreshing stage.
113        try: 
114            with closing(file(cache_path)) as f:
115                ns.cfg = json.read(f.read())
116        except: do_refresh = True
117
118    if do_refresh:
119        # Atomically reload the source and regenerate the cache.  The read and
120        # write must be a single transaction, or a stale version may be
121        # written (if another read/write of a more recent configuration
122        # is interleaved).  The final atomic rename is to keep this
123        # transactionally isolated from the above cache read.  If we fail to
124        # acquire the lock, just try to load the master configuration.
125        try:
126            with lock_file(lock_path):
127                ns.cfg = load_master()
128                try: 
129                    with closing(file(cache_path + '.tmp', 'w')) as f:
130                        f.write(json.write(ns.cfg))
131                except: pass # silent failure
132                else: rename(cache_path + '.tmp', cache_path)
133        except IOError:
134            ns.cfg = load_master()
135    return ns.cfg
136
137dicts = load()
138structs = dicts2struct(dicts)
139
140# vim:et:sw=4:ts=4
Note: See TracBrowser for help on using the repository browser.