7 Commits (feature/pgsql-native-backend-protocol)

Author	SHA1	Message	Date
Rene Cannao	9ed621edac	test: substantially expand plugin unit test coverage ... Plugin assertions: 63 → 230 (+167) across 4 binaries. Per-binary deltas (assertions / plan): plugin_registry_unit-t : 25 → 68 (+43) plugin_manager_unit-t : 26 → 83 (+57) plugin_config_unit-t : 12 → 47 (+35) plugin_dispatch_unit-t : - → 32 (new file) New coverage areas: Registry edge cases: - case-insensitive duplicate-table detection (lower vs upper vs mixed) - case-insensitive duplicate-command detection across canonical forms - tabs / newlines / multi-space canonicalisation - leading + trailing spaces, multiple trailing semicolons (incl. interleaved) - internal semicolons preserved (not collapsed) - ";;;" / whitespace-only / "\\t\\n \\r ;" canonicalise to empty → reject - register_table input validation (null name, empty name, null/empty def, out-of-range db_kind) - same-name tables across admin/config/stats kinds all accepted - tables() returns empty for invalid db_kind - insertion-order preservation for both tables and commands - 64-table stress: all stored pointers remain valid after deque growth - 2KB command roundtrip - dispatch result fields (error_code / rows_affected / message) propagated Manager state machine: - load on a missing path: false, non-empty err mentions dlopen, size stays 0 - empty manager: init/start/stop all succeed - idempotent init/start/stop (re-calls are no-ops) - destructor stops started plugins (verified via log) - destructor SKIPS plugins that were never started - destructor doesn't re-run stop if explicit stop_all already happened - register_command failure inside init_all aborts and names the operation - register_table-with-invalid-kind inside init_all aborts and names op Multi-plugin (new fake_plugin2.so): - init/start happen in registration order; stop runs in reverse order (verified via shared log) - second plugin's init failure prevents both starts (no destructor stops) - second plugin's start failure: first plugin (which started) IS stopped by the destructor, second is NOT (never successfully started) Config helpers: - multiple plugins parsed - empty plugins=() clears stale modules - non-string entries (int, bool) silently skipped between strings - missing path bubbles up with the path in err - active manager visibility before/after load and stop - reload replaces previous manager (full lifecycle for both old and new) - reload to empty list: previous plugin stopped, manager nulled - partial-load failure leaves no active manager behind New plugin_dispatch_unit-t: - global dispatcher returns false when no manager active - dispatcher canonicalises input the same way registry does - dispatcher tolerates context with all-null DB pointers - dispatcher returns false after stop, even for a previously-routed command - 8 threads x 250 dispatches concurrent: 2000/2000 succeed, 0 fail - stop helper / start helper safe on null inputs Infrastructure changes: - test/tap/test_helpers/fake_plugin.cpp: parameterised via FAKE_PLUGIN_NAME and FAKE_PLUGIN_ENV_PREFIX so the same source can be compiled into a second .so (libproxysql_fake_plugin2.so) used by multi-plugin tests. Log entries now prefixed with "<name>:" so a shared log file can distinguish events from each plugin. - test/tap/tests/unit/Makefile: builds the second .so, wires it into plugin_manager_unit-t and plugin_config_unit-t, and adds the new plugin_dispatch_unit-t target. All 57 unit-test binaries pass.	2 months ago
Rene Cannao	d34fb38162	refactor: remove PLUGIN prefix gate from plugin command registration ... Remove the 'PLUGIN ' prefix requirement from the plugin manager's command registration and dispatch. This allows plugins to register any canonical command string (e.g. 'LOAD MYSQLX USERS TO RUNTIME') instead of being forced into the 'PLUGIN <name> ...' namespace. Changes to lib/ProxySQL_PluginManager.cpp: - Delete kPluginCommandPrefix constant and has_plugin_command_prefix() - Remove prefix validation from register_command() - now only rejects empty canonical forms - Remove prefix check from dispatch_admin_command() - any registered command string can be dispatched Updated plugin_registry_unit-t.cpp to use canonical 'LOAD MYSQLX USERS TO RUNTIME' syntax instead of 'PLUGIN MYSQLX LOAD USERS TO RUNTIME'. The 'SELECT 1' test now succeeds (any string accepted) since the namespace restriction is gone.	3 months ago
Rene Cannao	5dd717a8f1	fix: fail fast on plugin table conflicts	3 months ago
Rene Cannao	d036ba8327	fix: canonicalize plugin command registration	3 months ago
Rene Cannao	2430516600	fix: reject conflicting plugin registrations	3 months ago
Rene Cannao	123bb7eaf0	fix: harden plugin registry ABI and ownership	3 months ago
Rene Cannao	cd15afdd1e	feat: support plugin-owned admin tables and commands	3 months ago