|
Elektra
0.8.18
|
You might want to read about architecture. and data structures first.
In this section, we will explain the heart of Elektra. kdbOpen() is responsible for the setup and the construction of the data structures needed later. kdbGet() does, together with the plugins, all actions necessary to read in the configuration. kdbSet() orchestrates the plugins to write out the configuration correctly. kdbClose() finally frees all previously allocated data structures.
kdbOpen() retrieves the mount point configuration with kdbGet() using the default backend. During this process, the function sets up the data structures which are needed for later invocations of kdbGet() or kdbSet(). All backends are opened and mounted in the appropriate parts of the key hierarchy. The resulting backends are added both to the Split and the Trie object. kdbOpen() finally returns a KDB object that contains all this information.
The reading of the mount point configuration and the consequential self configuring of the system is called bootstrapping. Elektra builds itself up from the simple variant with a default backend only to the sophisticated configuration system presented in this thesis.
kdbOpen() creates a Split object. It adds all backend handles and parentKeys during bootstrapping. So the buildup of the Split object takes place once. The resulting object is then used for both kdbGet() and kdbSet(). This approach is much better testable because the Split object is first initialised using the mount point configuration – separated from the filtering of the backends for every specific kdbGet() and kdbSet() request.
Afterwards the key hierarchy is static. Every application using Elektra will build up the same key database. Application specific mount points are prohibited because changes of mount points would destroy the global key database. Elektra could not guarantee that every application retrieves the same configuration with the same key names any longer.
In kdbOpen(), nearly no checks are done regarding the expected behaviour of the backend. The contract checker guarantees that only appropriate mount points are written into the mount point configuration. kdbOpen() checks only if the opening of plugin was successful. If not, the backend enclosing the plugin is not mounted at all.
In Elektra version 0.6, removing keys was an explicit request. Only a single Key object could be removed from the database. For configuration files this method is inapplicable. For filesys, however, it was easy to implement.
In Elektra version 0.7, the behaviour changed. Removing keys was integrated into kdbSet(). The user tagged keys that should be removed. After the next kdbSet(), these keys were removed from the key database. On the one hand, backends writing configuration files simply ignored the keys marked for removal. On the other hand, filesys needed that information to remove the files. To make this approach work for filesys, the marked keys were located at the very end of the KeySet and sorted in reverse. With this trick, recursive removing worked well. But this approach had major defects in the usage of KeySet. Because marking a key to be removed changed the sort order of the key set ksLookupByName() did not find this key anymore.
So in the present version removing keys is consistent again. A KeySet describes the current configuration. The user can reduce the KeySet object by popping keys out. The kdbSet() function applies exactly this configuration as specified by the key set to the key database. Contrary to the previous versions, the popped keys of the key set will be permanently removed.
The new circumstance yields idempotent properties for kdbSet(). The same KeySet can be applied multiple times, but after the first time, the key database will not be changed anymore. (Note that kdbSet()) actually detects that there are no changes and will do nothing. To actually show the idempotent behaviour the KeySet has to be regenerated or the key database needs to be reopened.
It is, however, not known if keys should be removed permanently only by investigating the KeySet. But only if this knowledge is present, the core can decide if the key set needs to be written out or if the configuration is unchanged. So we decided to track how many keys are delivered in kdbGet(). If the size of the KeySet is lower than this number determined at the previous kdbGet(), Elektra's core knows that some keys were popped. Hence, the next kdbSet() invocation needs to change the concerned key database.
The situation is now much clearer. The semantics of popping a key will result in removing the key from the key database. And the intuitive idea that a KeySet will be applied to the key database is correct again.
It is critical for application startup time to retrieve the configuration as fast as possible. Hence, the design goal of the kdbGet() algorithm is to be efficient while still enabling plugins to have relaxed postconditions. To achieve this, the sequence of syscalls must be optimal. On the other hand, it is not tolerable to waste time or memory inside Elektra's core, especially during an initial request or when no update is available.
The synopsis of the function is:
int kdbGet(KDB *handle, KeySet *returned, Key * parentKey);
The user passes a key set, called returned. If the user invokes kdbGet() the first time, he or she will usually pass an empty key set. If the user wants to update the application's settings, returned will typically contain the configuration of the previous kdbGet() request. The parentKey holds the information below which key the configuration should be retrieved. The handle contains the data structures needed for the algorithm, like the Split and the Trie objects.
kdbGet() does a rather easy job, because kdbSet() already guarantees that only well formatted, non-corrupted and well-typed configuration is written out in the key database. The task is to query all backends in question for their configuration and then merge everything.
A backend may yield keys that it is not responsible for. It is not possible for a backend to know that another backend has been mounted below and the other backend is now responsible for some of the keys that are still in the storage. Additionally, plugins are not able to determine if they are responsible for a key or not. Consequently, it can happen that more than one backend delivers a key with the same name.
kdbGet() ensures that a key is uniquely identified by its name. Elektra's core will pop keys that are outside of the backend's responsibility. Hence, these keys will not be passed to the user and we get the desired behaviour: The nearest mounted backend to the key is responsible.
For example, a generator plugin in the backend A always emits following keys. (A) and (B) indicate from which backend the key comes from.
user/sw/generator/akey (A)
user/sw/generator/dir (A)
user/sw/generator/dir/outside1 (A)
user/sw/generator/dir/outside2 (A)
It will still return these keys even if the plugin is not responsible for some of them anymore. This can happen if another backend B is mounted to user/sw/generator/dir. In the example it yields the following keys:
user/sw/generator/dir (B)
user/sw/generator/dir/new (B)
user/sw/generator/dir/outside1 (B)
user/sw/generator/outside (B)
In this situation kdbGet() is responsible to pop all three keys at, and below, user/sw/generator/dir of backend A and the key user/sw/generator/outside of backend B. The user will get the resulting key set:
user/sw/generator/akey (A)
user/sw/generator/dir (B)
user/sw/generator/dir/new (B)
user/sw/generator/dir/outside1 (B)
Note that the key exactly at the mount point comes from the backend mounted at user/sw/generator/dir.
kdbOpen() already creates a Split object for the whole configuration tree. In this object, kdbOpen() will append a list of all backends available. A specific kdbGet() request usually includes only a part of the configuration. For example, the user is only interested in keys below user/sw/apps/userapp. All backends that cannot contribute to configuration below user/sw/apps/userapp will be omitted for that request. To achieve this, parts of the Split object are filtered out. After this step we know the list of backends involved. The Split object allocates a key set for each of these backends.
Afterwards the first plugin of each backend is called to determine if an update is needed. If no update is needed, the algorithm has finished and returns zero.
Now we know which backends do not need an update. For these backends, the previous configuration from returned is appointed from to the key sets of the Split object. The algorithm will not set the syncbits of the Split object for these backends because the storage of the backends already contains up-to-date configuration.
The other backends will be requested to retrieve their configuration. The initial empty KeySet from the Split object and the relevant file name in the key value of parentKey are passed to each remaining plugin. The plugins extend, validate and process the key set. When an error has occurred, the algorithm can stop immediately because the user's KeySet returned is not changed at this point. When this part finishes, the Split object contains the whole requested configuration separated in various key sets.
Subsequently the freshly received keys need some post-processing:
kdbGet() removes it.kdbSet() calls, kdbGet() needs to store the number of received keys of each backend.Trie. In this process, Elektra pops all duplicated and overlapping keys in favour of the responsible backend.The last step is to merge all these key sets together. This step changes the configuration visible to the user. After some cleanup the algorithm finally finishes.
The user can call kdbGet() often even if the configuration or parts of it are already up to date. This can happen when applications reread configuration in some events. Examples are signals (SIGHUP is the signal used for that on UNIX systems. It is sent when the program's controlling terminal is closed. Daemons do not have a terminal so the signal is reused for reloading configuration.), notifications, user requests and in the worst case periodical attempts to reread configuration.
The given goal is to keep the sequence of needed syscalls low. If no update is needed, it is sufficient to request the timestamp (On POSIX systems using stat()) of every file. No other syscall is needed. Elektra's core alone cannot check that because getting a timestamp is not defined within the standard C99. So instead the resolver plugin handles this problem. The resolver plugin returns 0 if nothing has changed.
This decision yields some advantages. Both the storage plugins and Elektra's core can conform to C99. Because the resolver plugin is the very first in the chain of plugins, it is guaranteed that no useless work is done.
Because Elektra provides self-contained configuration, kdbOpen() has to retrieve settings in the bootstrapping process below system/elektra as explained in bootstrapping. Because of the new way to keep track of removed keys, the internally executed kdbGet() creates a problem. Without countermeasures even the first kdbGet() of a user requesting the configuration below system/elektra fails because the resolver finds out that the configuration is already up to date. The configuration delivered by the user is empty at this point. As a result, the empty configuration will be appointed and returned to the user.
A simple way to resolve this issue is to reload the default backend after the internal configuration was fetched. Reloading resets the timestamps and kdbGet() works as expected.
Not performance, but robust and reliable behaviour is the most important issue for kdbSet(). The design was chosen so that some additional in-memory comparisons are preferred to a suboptimal sequence of syscalls. The algorithm makes sure that keys are written out only if it is necessary because applications can call kdbSet() with an unchanged KeySet. For the code to decide this, performance is important.
kdbSet() guarantees the following properties:
KeySet is propagated to permanent storage. Otherwise nothing is changed in the key database. Plugins developed during the thesis meet this requirement.The synopsis of the function is:
int kdbSet(KDB *handle, KeySet *returned, Key * parentKey);
The user passes the configuration using the KeySet returned. The key set will not be changed by kdbSet(). The parentKey provides a way to limit which part of the configuration is written out. For example, the parentKey user/sw/org/app/#0/current will induce kdbSet() to only modify the key databases below user/sw/org/app even if the KeySet returned also contains more configuration. Note that all backends with no keys in returned but that are below parentKey will completely wipe out their key database. The KDB handle contains the necessary data structures.
As a first step, kdbSet() divides the configuration passed in by the user to the key sets in the Split object. kdbSet() searches for every key if the sync flag is checked. Then kdbSet() decides if a key was removed from a backend by comparing the actual size of the key set with the size stored from the last kdbGet() call. We see that it is necessary to call kdbGet() first before invocations of kdbSet() are allowed.
We know that data of a backend has to be written out if at least one key was changed or removed. If no backend has any changes, the algorithm will terminate at this point. The careful reader notices that the process involves no file operations.
If some backends need synchronisation, the algorithm continues by filtering out all backends in the Split object that do not have changes. At this point, the Split object has a list of backends with their respective key sets.
Plugins in kdbSet() can change values. Other than in kdbGet(), the user is not interested in these changes. Instead, the values are transformed to be suitable for the storage. To make sure that the changed values are not passed to the user, the algorithm continues with a deep duplication of all key sets in the Split object.
All plugins of each included backend are executed one by one up to the resolver plugin. If this succeeds, the resolver plugin is responsible for committing these changes. After the successful commit, error codes of plugins are ignored. Only logging and notification plugins are affected.
For this thesis only file-based storage with atomic properties were developed. The replacement of a file with another file that has not yet been written is not trivial. The straightforward way is to lock a file and start writing to it. But this approach can result in broken or partially finished files in events like ''out of disc space'', signals or other asynchronous aborts of the program.
A temporary file solves this problem, because in problematic events the original file stays untouched. When the temporary file is written out properly, it is renamed and the original configuration file is overwritten. But another concurrent invocation of kdbSet() can try to do the same with the result that one of the newly written files is lost.
To avoid this problem, locks are needed again. It is not possible to lock the configuration file itself because it will be unlinked when the temporary file is renamed. So a third file for locking is needed. The resolver currently implements this approach.
An alternative to this approach without locks is to completely rely on the modification time. The modification time typically has only a resolution of one second. So any changes within that time slot will not be recognised. For this approach, however, the name of every temporary file must be unique because concurrent kdbSet() invocations each try to create one. The temporary file must also be unlinked in case of a rollback. The opened temporary file can be passed to the storage plugins using a file name in the directory */dev/fd*. This approach may be more practical than the currently implemented way because it does not need the additional lock file (Nevertheless, the other way was chosen to test if the algorithm is exception safe as described in exception safety).
The plugins within kdbSet() can fail for a variety of reasons. Conflicts occur most frequently. A conflict means that during executions of kdbGet() and kdbSet() another program has changed the key database. In order not to lose any data, kdbSet() fails without doing anything. In conflict situations Elektra leaves the programmer no choice. The programmer has to retrieve the configuration using kdbGet() again to be up to date with the key database. Afterwards it is up to the application to decide which configuration to use. In this situation it is the best to ask the user, by showing him the description and reason of the error, how to continue:
kdbGet() the application is already up to date with the new configuration. All configuration changes the user made before will be lost.kdbSet().ksAppend(). The source parameter is the preferred configuration. Note that the downside of the third option is that the merged configuration might not be valid.Sometimes a concrete key causes the problem that the whole key set cannot be stored. That can happen on validation or because of type errors. Such errors are usually caused by a mistake made by the user. So the user is responsible for changing the settings to make it valid again. In such situations, the internal cursor of the KeySet returned will point to the problematic key.
A completely different approach is to export the configuration when kdbSet() returned an error code. The user can then edit, change or merge this configuration with more powerful tools. Finally, the user can import the configuration into the global key database. The export and import mechanism is called "streaming" and will be explained in streaming.
1.8.8