Comparing changes
Open a pull request
Commits on Jan 24, 2024
-
Do not try and retrieve pid/ns for non UNIX sockets
This just generates uneeded syslog messages.
Commits on Feb 7, 2024
-
-
Add directory level refresher test
Ensure a directory level refresh is called when a get tree is done on that directory.
-
Add refresher path matching tests
Ensure refreshers are sent a path that matches the refreshed path. Some refreshers will used the passed in path to match against local copies, in this case the path must be terminated with a "/" if we don't otherwise have a key.
-
Call directory level matches for refreshers
When processing a traverse we need to search for both the exact key and values further down the tree. Usually a direct single value will be fetched with apteryx_get, but it is legal (and done by some users) to call apteryx_get_tree for a single value.
-
Send / to refreshers who require it
When sending refresher callback message send a path that has a trailing '/' if the refresher is either a wildcard or directory level refresher, and this call was not for a specific value.
-
Commits on Feb 8, 2024
-
Shutdown apteryx client library on lua_close
The apteryx client library is reference counted and so needs to be shutdown when the lua instance is closed.
Commits on Feb 9, 2024
Commits on Feb 18, 2024
-
Prepare more suitable paths to be passed to the refresher
We combine info from the refreshed path and the requested path. We now match both the current node and then node/ when looking for refreshers when doing a get tree. We no longer need the direct_request parameter to call_refreshers. Add more comprehensive unit tests for refresher paths.
Commits on Feb 19, 2024
-
Fix refreshing paths for wildcards
Do not call refresher if we have unresolved wildcards. Pass an actual wildcard node when recursing _refresh_paths so that strcmp matches properly. Fixup test asssertions for cases that should not have called a refresher and cases that were missing the final /. Add a bunch of test cases for root path queries and refreshers.
-
Call trunk wildcard refresher for first wildcard in query
This restores the previous behaviour and does not break the rule of a refresher not being passed a matching path. A refresher will be passed the longest path in a query that does not have a wildcard in it yet matches the refresher.
Commits on Feb 20, 2024
-
Tune lua garbage collector in alfred
The default values for the garbage collector in alfred can result in an ever-growing pile of garbage when repeated provide calls are made. Tune the garbage collector by reducing the pause value from the default 200 to 150 - the GC runs more frequently and prevents the pile from building up nearly as high. The lua_gc API differs between lua 5.3 and 5.4.
Commits on Mar 5, 2024
-
Fix local refresher for statistics
We should record the timestamp and last path on any local refreshers. We dont need to specify the CAS timestamp when adding stats to the DB, but using the wrong timestamp fn is even worse. Currently the only local refresher is for statistics and this fix avoids the stats refresher being thrashed.
Commits on Mar 11, 2024
-
Use MONOTONIC time for socket recv timeout
Get rid of the duplication of get_time_us functions.
-
Commits on Mar 13, 2024
-
Ensure workers have correct sigmask
The slow work thread should have all signals blocked. But should call callbacks with the worker_sigmask. Tidy up variable naming so this is more obvious if this code is changed later.
Commits on Apr 3, 2024
-
A worker pool is only needed for rpc instances that serve call backs
Since the glib-2.0 patch 4d2e77a55 GThreadPool: Always use the thread-spawning thread for the global shared thread pool GNOME/glib@4d2e77a creating a threadpool spawns a new thread. Previously, this was only the case for non-Linux platforms. This can cause issues for daemons that fork and depend on libraries that load apteryx on LT_INIT, as the glib threadpool spawned thread will not be copied on the fork. Additionally, now libraries that only use get calls will be running an extra thread from the threadpool. Move the threadpool creation into request_cb() to only spawn the pool managing thread for applications that require callbacks using workers.
Commits on Apr 15, 2024
-
Implement apteryx_watch_tree_masked
Allows users to create a watch callback that masks sets made by itself as defined by namespace and pid. Flags are now passed to apteryxd via the GUID. RPC messages contain the sender ns/pid. Use macros to alow users to activate the required functionality without them calling the _full fn.
Commits on Apr 16, 2024
-
Dont print error if apteryx_path_node passed NULL tree
It is commonly used to check if a tree has a path and that tree may be completely empty.
Commits on May 3, 2024
-
tests: don't try to sudo and set a value that doesn't exist
tcp_tw_recycle hasn't existed since 4.12, also it's a bit unfriendly to try to sudo
Commits on Sep 17, 2024
Commits on Sep 20, 2024
-
Large set tree + watch performance
Add a new larger set tree with a watch. This is the target for upcoming tree callback matching.
-
Add tree based callback matching
Add cb_match_tree. This function returns a GList of struct cb_tree_info with the usual cb_info_t structure populated, with a newly allocated GNode tree of the data that matches the callback.
-
Use tree matching for watchers
The new config_match_tree function allows us to match the whole change set in a set tree or prune against all watchers and validators. These callbacks can then be called with the matching tree as a single RPC - this can save a lot of processing for large changesets.
-
-
-
-
Handle error in parsing tree during handle_watch
Handle the unlikely case that the rpc_msg_decode_tree fails to build a tree.
-
Ensure tree passed to watch tree callback are valid to be used in apt…
…eryx_set_tree apteryx_set_tree and apteryx_query both check that the root key in the tree is a valid apteryx path. It handle this ensure that the GNode that is created when parsing the tree has a "/" if it is not a longer path.
Commits on Sep 30, 2024
-
Add tests and assert for invalid tree structures
When Apteryxd is handling a tree it requires the non-root nodes to be single key values. Add tests where the client sends compound keys and an assert to stop the process.
-
Break up compound non-root keys
Code in apteryxd is designed to handle GNode keys as single key values, so encode the tree in this way.
Commits on Oct 8, 2024
-
Return DB timestamp for refreshed paths
Previously we alway returned "now" if there was a chance of a path being refreshed. That was because refreshed paths always had their timestamps updated to "now" even if the refresher decided not to update anything so that we did not call the refresher again for stale child nodes. Now that we do not use the DB timestamps to decide whether to call the refreshers anymore we can fix this odd behaviour. The change in behaviour is; 1. A call to a refresher does not result in all children of the path having their timestamps set to "now" even if they were not modified. 2. A call to get a timestamp for a path will call refreshers to make sure the timestamp is up to date.
Commits on Oct 14, 2024
-
-
Lock tree during cb_match_tree
The cb tree can be changed while iterating, so it needs to be locked.
Commits on Oct 23, 2024
-
Add a function to remove duplicate values from a tree - as a side effect this will sort each level of the tree in alphabetic order.
-
apteryx_query result should not have duplicates
apteryx_query has an unusual behaviour in that some advanced queries can generate a tree with duplicate nodes. Remove these with apteryx_uniqify_tree.
-
This allows the g_node_prepend done in _rpc_msg_decode_tree to reassemble the tree in the same order as it was originally without being O(n*n).
-
-
Instated a check that now passes, made another test do what was intended.
-
-
Check apteryx_set_tree with multiple values
Ensure a client setting up multiple values for the same key in a tree ends up with the last value actually being used. There are ways of doing this setup which will end up with the wrong value being set (requiring a mix of methods to create the tree), but this will work when APTERYX_VALUE / APTERYX_NODE is used the same way multiple times to create a tree with duplicate nodes.
Commits on Nov 20, 2024
Commits on Nov 28, 2024
-
Apteryxd can end up storing a lot of duplicated strings. This commit adds a cache to remove this duplication - it is called the same way strdup / free would be and will take care of allocating, reference counting and releasing strings. This cache has no locking, so needs to be called under an external lock. At the moment this used solely in code inside the db_lock in database.c
-
Update apteryx query assertions
When showing an incomplete nat rule router would crash, due to an invalid memory access. This was caused by having a validity check on an entity that does not exist resulting in an apteryx node with null data. This has been remedied by updating the assertion in the apteryx query func to return NULL when the node data is NULL. Preventing an invalid memory access when destroying the node.
Commits on Nov 29, 2024
-
Update memuse for use of string-cache
Also use malloc_usable_size + malloc header of 8 bytes where we can.
Commits on Dec 10, 2024
-
Fix thread contention in string-cache
Both database.c and callback.c use the hashtree code, so the cache needs to be protected with a lock.
There are no files selected for viewing