9 SRC_PATH = '/etc/invirt/master.yaml'
10 SRC_DIRPATH = '/etc/invirt/conf.d'
11 CACHE_PATH = '/var/lib/invirt/cache.json'
12 LOCK_PATH = '/var/lib/invirt/cache.lock'
17 Splice dict-tree d2 into d1. Return d1.
19 d2 may be None for an empty dict-tree, because yaml.load produces that.
22 >>> d = {'a': {'b': 1}, 'c': 2}
23 >>> augment(d, {'a': {'d': 3}})
24 {'a': {'b', 1, 'd': 3}, 'c': 2}
26 {'a': {'b', 1, 'd': 3}, 'c': 2}
32 if k in d1 and isinstance(d1[k], dict):
38 def run_parts_list(dirname):
40 Reimplements Debian's run-parts --list.
42 One difference from run-parts's behavior: run-parts --list /foo/
43 will give output like /foo//bar, but run_parts_list('/foo/') gives
44 /foo/bar in deference to Python conventions.
46 Matches documented behavior of run-parts in debianutils v2.28.2, dated 2007.
50 lanana_re = re.compile('^[a-z0-9]+$')
51 lsb_re = re.compile('^_?([a-z0-9_.]+-)+[a-z0-9]+$')
52 deb_cron_re = re.compile('^[a-z0-9][a-z0-9-]*$')
53 for name in os.listdir(dirname):
54 if lanana_re.match(name) or lsb_re.match(name) or deb_cron_re.match(name):
55 yield os.path.join(dirname, name)
59 yield from run_parts_list(SRC_DIRPATH)
63 for filename in list_files():
64 with open(filename) as f:
65 augment(cfg, yaml.load(f, yaml.CSafeLoader))
69 return max(max(os.path.getmtime(filename) for filename in list_files()),
70 os.path.getmtime(SRC_DIRPATH))
72 def load(force_refresh=False):
74 Try loading the configuration from the faster-to-load JSON cache at
75 CACHE_PATH. If it doesn't exist or is outdated, load the configuration
76 instead from the original YAML file at SRC_PATH and regenerate the cache.
77 I assume I have the permissions to write to the cache directory.
85 src_mtime = get_src_mtime()
87 cache_mtime = os.path.getmtime(CACHE_PATH)
91 do_refresh = src_mtime + 1 >= cache_mtime
93 # We chose not to simply say
95 # do_refresh = src_mtime >= cache_time
97 # because between the getmtime(SRC_PATH) and the time the cache is
98 # rewritten, the master configuration may have been updated, so future
99 # checks here would find a cache with a newer mtime than the master
100 # (and thus treat the cache as containing the latest version of the
101 # master). The +1 means that for at least a full second following the
102 # update to the master, this function will refresh the cache, giving us
103 # 1 second to write the cache. Note that if it takes longer than 1
104 # second to write the cache, then this situation could still arise.
106 # The getmtime calls should logically be part of the same transaction
107 # as the rest of this function (cache read + conditional cache
108 # refresh), but to wrap everything in an flock would cause the
109 # following cache read to be less streamlined.
112 # Try reading from the cache first. This must be transactionally
113 # isolated from concurrent writes to prevent reading an incomplete
114 # (changing) version of the data (but the transaction can share the
115 # lock with other concurrent reads). This isolation is accomplished
116 # using an atomic filesystem rename in the refreshing stage.
118 with open(CACHE_PATH) as f:
124 # Atomically reload the source and regenerate the cache. The read and
125 # write must be a single transaction, or a stale version may be
126 # written (if another read/write of a more recent configuration
127 # is interleaved). The final atomic rename is to keep this
128 # transactionally isolated from the above cache read. If we fail to
129 # acquire the lock, just try to load the master configuration.
131 with invirt.common.open_locked(LOCK_PATH):
134 with open(CACHE_PATH + '.tmp', 'w') as f:
137 pass # silent failure
139 os.rename(CACHE_PATH + '.tmp', CACHE_PATH)