X-Git-Url: http://xvm.mit.edu/gitweb/invirt/packages/invirt-base.git/blobdiff_plain/a11a45e1b18add76f2a440412ef6379a9874b082..bed209caf968dd992a7f849f8b982b4b46a0f876:/files/usr/share/python-support/sipb-xen-base/invirt/config.py

diff --git a/files/usr/share/python-support/sipb-xen-base/invirt/config.py b/files/usr/share/python-support/sipb-xen-base/invirt/config.py
index c27ba6b..8648c70 100644
--- a/files/usr/share/python-support/sipb-xen-base/invirt/config.py
+++ b/files/usr/share/python-support/sipb-xen-base/invirt/config.py
@@ -1,9 +1,11 @@
 import json
 from invirt.common import *
+from os import rename
 from os.path import getmtime
 
 default_src_path   = '/etc/invirt/master.yaml'
 default_cache_path = '/var/lib/invirt/cache.json'
+lock_file          = '/var/lib/invirt/cache.lock'
 
 def load(src_path = default_src_path,
          cache_path = default_cache_path,
@@ -14,33 +16,70 @@ def load(src_path = default_src_path,
     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 = getmtime(src_path)
         try:            cache_mtime = getmtime(cache_path)
         except OSError: do_refresh  = True
-        else:           do_refresh  = src_mtime > cache_mtime
+        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
-        try: cfg = with_closing(file(cache_path))(lambda f: json.read(f.read()))
+        # 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: ns.cfg = with_closing(file(cache_path)) (
+                lambda f: 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.
-        @with_lock_file('/var/lib/invirt/cache.lock')
-        def cfg():
-            import yaml
-            try:    default_loader = yaml.CSafeLoader
-            except: default_loader = yaml.SafeLoader
-            cfg = with_closing(file(src_path))(lambda f: yaml.load(f, default_loader))
-            try: with_closing(file(cache_path, 'w'))(lambda f: f.write(json.write(cfg)))
-            except: pass # silent failure
-            return cfg
-    return cfg
+        # 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.
+        import yaml
+        try:    loader = yaml.CSafeLoader
+        except: loader = yaml.SafeLoader
+        try:
+            @with_lock_file(lock_file)
+            def refresh_cache():
+                ns.cfg = with_closing(file(src_path)) (
+                        lambda f: yaml.load(f, loader))
+                try: with_closing(file(cache_path + '.tmp', 'w')) (
+                        lambda f: f.write(json.write(ns.cfg)))
+                except: pass # silent failure
+                else: rename(cache_path + '.tmp', cache_path)
+        except IOError:
+            ns.cfg = with_closing(file(src_path)) (
+                    lambda f: yaml.load(f, loader))
+    return ns.cfg
 
 dicts = load()
 structs = dicts2struct(dicts)