source: trunk/packages/invirt-base/python/invirt/config.py @ 1421

from __future__ import with_statement

import json
from invirt.common import *
from os import rename
from os.path import getmtime
from contextlib import closing
import yaml

try:    loader = yaml.CSafeLoader
except: loader = yaml.SafeLoader

src_path   = '/etc/invirt/master.yaml'
cache_path = '/var/lib/invirt/cache.json'
lock_path  = '/var/lib/invirt/cache.lock'

def load_master():
    with closing(file(src_path)) as f:
        return yaml.load(f, loader)

def get_src_mtime():
    return getmtime(src_path)

def load(force_refresh = False):
    """
    Try loading the configuration from the faster-to-load JSON cache at
    cache_path.  If it doesn't exist or is outdated, load the configuration
    instead from the original YAML file at src_path and regenerate the cache.
    I assume I have the permissions to write to the cache directory.
    """

    # Namespace container for state variables, so that they can be updated by
    # closures.
    ns = struct()

    if force_refresh:
        do_refresh = True
    else:
        src_mtime = get_src_mtime()
        try:            cache_mtime = getmtime(cache_path)
        except OSError: do_refresh  = True
        else:           do_refresh  = src_mtime + 1 >= cache_mtime

        # We chose not to simply say
        #
        #   do_refresh = src_mtime >= cache_time
        #
        # because between the getmtime(src_path) and the time the cache is
        # rewritten, the master configuration may have been updated, so future
        # checks here would find a cache with a newer mtime than the master
        # (and thus treat the cache as containing the latest version of the
        # master).  The +1 means that for at least a full second following the
        # update to the master, this function will refresh the cache, giving us
        # 1 second to write the cache.  Note that if it takes longer than 1
        # second to write the cache, then this situation could still arise.
        #
        # The getmtime calls should logically be part of the same transaction
        # as the rest of this function (cache read + conditional cache
        # refresh), but to wrap everything in an flock would cause the
        # following cache read to be less streamlined.

    if not do_refresh:
        # Try reading from the cache first.  This must be transactionally
        # isolated from concurrent writes to prevent reading an incomplete
        # (changing) version of the data (but the transaction can share the
        # lock with other concurrent reads).  This isolation is accomplished
        # using an atomic filesystem rename in the refreshing stage.
        try: 
            with closing(file(cache_path)) as f:
                ns.cfg = json.read(f.read())
        except: do_refresh = True

    if do_refresh:
        # Atomically reload the source and regenerate the cache.  The read and
        # write must be a single transaction, or a stale version may be
        # written (if another read/write of a more recent configuration
        # is interleaved).  The final atomic rename is to keep this
        # transactionally isolated from the above cache read.  If we fail to
        # acquire the lock, just try to load the master configuration.
        try:
            with lock_file(lock_path):
                ns.cfg = load_master()
                try: 
                    with closing(file(cache_path + '.tmp', 'w')) as f:
                        f.write(json.write(ns.cfg))
                except: pass # silent failure
                else: rename(cache_path + '.tmp', cache_path)
        except IOError:
            ns.cfg = load_master()
    return ns.cfg

dicts = load()
structs = dicts2struct(dicts)

# vim:et:sw=4:ts=4