157 Commits (ca202ceba1ea08711c5b1803ddffdbfbb41fb9dd)

Author	SHA1	Message	Date
Rene Cannao	5916ef4d43	docs(genai): update plugin README for new features, add implementation plan ... Update plugins/genai/README.md to reflect the current state after Features A, C, D, E: - Replace all PROXYSQLGENAI references with PROXYSQL40 (flag was removed in PR #5741) - Document stats_genai_global table and its 21 SQL status variables in the Observability section - Document MCP query digest persistence in Observability - List all stats tables in the Configuration section (6 stats tables including the new stats_genai_global) - Remove stale 'Known gaps' about stats_mcp_* tables (now registered) and LLM Bridge cache TODO (now implemented) Also track the implementation plan used for this batch: docs/superpowers/plans/2026-05-03-genai-features-batch.md	2 months ago
Rene Cannao	1fe9c3a805	feat(genai): implement LLM Bridge semantic cache with vector similarity (Feature A) ... Complete the 4 TODO stubs in LLM_Bridge to make semantic caching functional, reducing LLM API costs for repeated similar queries. Config changes: - Add cache_enabled bool to LLM_Bridge config struct (default: true) - Update update_config() signature to accept cache_enabled parameter - Wire genai_llm_cache_enabled through AI_Features_Manager::init_llm_bridge() check_cache(): - Generate embedding for the prompt via GloGATH->embed_documents() - KNN search against llm_cache_vec (sqlite-vec) joined with llm_cache - Compute cosine similarity = 1 - distance; compare against threshold - On hit: update hit_count/last_hit, increment GloAI cache_hits counter - On miss: increment GloAI cache_misses counter - Track lookup timing via GloAI->add_llm_cache_lookup_time_ms() store_in_cache(): - Generate embedding, INSERT into llm_cache, INSERT into llm_cache_vec - Track store timing via GloAI->add_llm_cache_store_time_ms() clear_cache(): - DELETE FROM llm_cache_vec, DELETE FROM llm_cache get_cache_stats(): - SELECT COUNT(*), SUM(hit_count) FROM llm_cache - Read misses/lookups/stores from AI_Features_Manager::collect_status_variables() - Return JSON: {entries, hits, misses, lookups, stores}	2 months ago
Rene Cannao	827f292f2f	feat(genai): register stats_genai_global status table (Feature C) ... Add a new stats_db runtime view that aggregates all 21 GenAI/MCP/AI status variables into a single queryable admin table. New table: stats_genai_global (Variable_name VARCHAR, Value VARCHAR) Registered as a stats_db runtime view with refresh callback that: 1. Collects status variables from GloGATH, genai_context().mcp, GloAI 2. Wipes and repopulates the table within BEGIN/COMMIT Usage: SELECT * FROM stats_genai_global ORDER BY Variable_name; Includes GenAI_Thread.h and AI_Features_Manager.h includes for access to the extern GloGATH and GloAI globals from the refresh callback.	2 months ago
Rene Cannao	b4ff986f10	feat(genai): add collect_status_variables() to genai/mcp/ai handlers (Feature C) ... Expose internal counters from all three handler classes so operators can monitor plugin health from the admin interface. Each handler gets a collect_status_variables() method returning std::vector<std::pair<std::string, std::string>>: - GenAI_Threads_Handler: 4 counters (threads_initialized, active_requests, completed_requests, failed_requests) - MCP_Threads_Handler: 3 counters (total_requests, failed_requests, active_connections) - AI_Features_Manager: 14 counters (llm_total_requests, llm_cache_hits/misses/lookups/stores, llm_local/cloud_model_calls, llm_total_response_time_ms, llm_cache_total_lookup/store_time_ms, anomaly_total_checks/blocked_queries/flagged_queries, daily_cloud_spend_usd) Total: 21 status variables queryable via stats_genai_global table.	2 months ago
Rene Cannao	dca155aa97	feat(genai): persist MCP query digest stats to SQLite (Feature E) ... Add restart resilience for MCP query digest statistics by persisting the in-memory digest map to the catalog's own SQLite database. New methods on Discovery_Schema: - create_digest_persist_table(): creates mcp_query_digest_persist table in the catalog DB (idempotent, called during init_schema()) - flush_digest_to_sqlite(): serialises the in-memory digest map to SQLite under BEGIN IMMEDIATE / COMMIT, called every 100 updates via update_mcp_query_digest() - load_persisted_digests(): restores persisted digest stats from SQLite back into the in-memory map, called during init_schema() Reset handling: get_mcp_query_digest(reset=true) now also clears the persist table so reset stats don't reappear after restart. Thread safety: flush uses rdlock (read-only traversal), load uses wrlock (populates the map). Reset uses the existing wrlock path.	2 months ago
Rene Cannao	a9f2eaaa13	feat: remove PROXYSQLGENAI flag, genai now builds under PROXYSQL40 (issue #5722) ... PROXYSQLGENAI flag eliminated entirely. PROXYSQL40=1 now unconditionally builds the genai plugin alongside core. All #ifdef PROXYSQLGENAI replaced with #ifdef PROXYSQL40 in source, Makefiles, and 61 CI workflows.	2 months ago
Rene Cannao	a25e11cb75	fix: replace non-existent next_row() with rows[]->fields API (issue #5722) ... SQLite3_result has no next_row() method. The correct iteration API is: for (int i = 0; i < result->rows_count; i++) { char** row = result->rows[i]->fields; } This was the root cause of TSAN/genai-gcov CI failures — the genai plugin failed to compile inside Docker, producing no test binaries.	2 months ago
Rene Cannao	5fea82e617	fix: resolve remaining SonarCloud warnings for PR #5741 ... - Replace #define with constexpr for ABI version constants - Replace strdup/free with std::string in AI_Features_Manager - Change mkdir permissions from 0755 to 0700 - Reduce nesting from 4 to 3 levels via early continue	2 months ago
Rene Cannao	da089147ca	fix: set result=nullptr after delete to satisfy SonarCloud (issue #5741) ... - All 20 delete result; statements now followed by result = nullptr; - Eliminates 17 SonarCloud security hotspot warnings about delete without null - Part of SonarCloud quality gate fix for PR #5741	2 months ago
Rene Cannao	88db9fd10f	fix: resolve PR #5741 review findings (issue #5729) ... - Fix groups.json sort order (CI lint failure) - Add sqlite3_mprintf null check in Anomaly_Detector (Gemini high) - Add ROLLBACK on DELETE/COMMIT error in 5 plugin_tables.cpp callbacks - Fix TAP plan count in test_stats_mcp_tables-t (12→11) - Fix TAP plan count in genai_plugin_anomaly_unit-t (24→25) - Fix ABI.md: register_runtime_view is ABI 3+, not ABI 4 - Update plan/spec docs: db_kind at struct tail (ABI backward compat) - Add Darwin -force_load fallback in src/Makefile - Add mkdir() error handling in AI_Features_Manager Verified: build clean, 30/30 unit tests pass, lint passes.	2 months ago
Rene Cannao	b8cfa32c63	fix: address all critical and important code review findings (issue #5729) ... Critical fixes: - Move db_kind to end of ProxySQL_PluginRuntimeView (tail-append = free ABI-3 compat) - Add mutex to protect Anomaly_Detector::user_statistics from concurrent session threads - Replace system(mkdir -p) with pure mkdir() loop (command injection fix) - Replace 77 std::stoi/stoll with non-throwing parse_int_or_zero/parse_ll_or_zero - Fix proxysql.cnf filename: Mysqlx -> MySQLX - Add --whole-archive to non-ClickHouse link path (plugin symbol resolution) Important fixes: - Add default: case with warning to db_kind switch in ProxySQL_PluginManager.cpp - Add config_db dispatch and null-handle tests to plugin_runtime_views_unit-t (6 new tests) - Update CLAUDE.md with sqlite-vec PROXYSQLGENAI exception - Remove stale anomaly_detection-t Makefile target - Replace fixed 1024-byte buffer with dynamic sqlite3_mprintf in Anomaly_Detector Minor: update ABI.md, PLUGIN_API.md, stale comments, unit test plan counts	2 months ago
Rene Cannao	0d53c57db7	feat(genai): register stats_mcp_* tables with db_kind=stats_db and refresh callbacks (issue #5729)	2 months ago
Rene Cannao	7c4c9e81e0	feat(mysqlx): set db_kind on runtime views, drop statsdb back-channel (issue #5729)	2 months ago
René Cannaò	54844df0bb	Merge branch 'v3.0' into v3.0-genai-plugin	2 months ago
Rene Cannao	960eeab8d2	feat(mysqlx): wire per-route tls_mode through handler_capabilities_set entry path ... PR #5709 landed the splice mechanics for tls_mode='passthrough' plus the per-route tls_mode column on mysqlx_routes, but no code consulted the route's tls_mode to actually drive the session into X_PASSTHROUGH_FORWARD from real client traffic — a route configured as `tls_mode='passthrough'` behaved identically to `'inherit'`. This commit closes the gap (issue #5710): (1) Advertise gating in send_capabilities() ----------------------------------------- A new helper, MysqlxSession::effective_route_tls_mode(), reads listener_route_name_ + the thread's MysqlxConfigStore and returns the per-route tls_mode (or `inherit` when listener route is empty / config store is unwired — preserving historical behaviour for tests and unconfigured deployments). send_capabilities() now suppresses the `tls` capability from the advertised set when route_tls_mode == disabled. Other modes (inherit / preferred / required / passthrough) advertise TLS the same way as before — passthrough MUST advertise so the client even thinks to upgrade. (2) Symmetric refusal in handler_capabilities_set() ------------------------------------------------ A client that sends CapabilitiesSet(tls=true) on a disabled-mode route is refused with X-Protocol error 5052 ("TLS is not enabled on this route"). Symmetric with the advertise gate; we never told the client TLS was available, so accepting it would be a silent policy bypass. (3) Passthrough entry from handler_capabilities_set() -------------------------------------------------- When CapabilitiesSet(tls=true) lands on a passthrough-mode route: a. resolve_passthrough_backend_target() picks the endpoint from the route's destination_hostgroup (no identity_ — passthrough authenticates end-to-end); b. the original CapabilitiesSet frame bytes are buffered in passthrough_pending_capset_frame_ for verbatim forwarding; c. tls_mode_ flips to TLS_PASSTHROUGH and status_ to the new X_PASSTHROUGH_BACKEND_CONNECTING state. (4) X_PASSTHROUGH_BACKEND_CONNECTING handler ----------------------------------------- handler_passthrough_backend_connecting() drives the four-step sequence: 1. allocate / pick up backend_conn_, start non-blocking TCP connect; 2. poll check_connect() to completion; 3. write the buffered CapabilitiesSet bytes verbatim to the backend (with EAGAIN/partial-write retries); 4. read exactly one X-Protocol frame from the backend (the LAST X-Protocol parse this session does); on Ok forward to the client and transition to X_PASSTHROUGH_FORWARD; on Error propagate to the client and close. The handler is registered both in the early fast-path branch of MysqlxSession::handler() (so the X-Protocol read on client_ds_ is skipped — the client's next bytes are the TLS ClientHello which is not parseable as an X frame) and in the main dispatch switch. Tests (test/tap/tests/unit/mysqlx_session_unit-t.cpp, +11 ok()s): * test_route_tls_mode_disabled_no_tls_in_advertise: drives CapabilitiesGet on a tls_mode='disabled' route, parses the response, asserts the advertised set has no `tls` capability. * test_capabilities_set_tls_refused_on_disabled_route: drives CapabilitiesSet(tls=true) on the same route, asserts session becomes unhealthy and the client receives a Mysqlx::Error. * test_route_passthrough_full_entry_path: end-to-end exercise. Stands up a loopback TCP listener as the "backend"; configures a passthrough route pointing at the listener; drives CapabilitiesGet -> CapabilitiesSet(tls=true); asserts the proxy forwards the CapabilitiesSet bytes verbatim, transitions through X_PASSTHROUGH_BACKEND_CONNECTING into X_PASSTHROUGH_FORWARD on the backend's Ok response, and stays healthy. Plan count: 87 -> 98 (after Fix B) -> 109. Closes #5710.	2 months ago
Rene Cannao	b7c5a08b96	feat(mysqlx): backlog and resume EAGAIN'd writes in passthrough splice ... Previously, the passthrough splice loop's pump_one_direction helper treated EAGAIN/EWOULDBLOCK on write() as fatal — the session transitioned to X_SESSION_CLOSING and the connection was dropped. In practice this turned any transient destination back-pressure (slow client, slow backend, kernel send-buffer drained more slowly than the proxy reads from the source side) into a connection drop. End-to-end encrypted sessions (the whole point of tls_mode='passthrough') would RST mid-stream under bursty load. Fix: per-direction write backlog, drained before the next read. * Add MysqlxSession::passthrough_c2b_backlog_ and passthrough_b2c_backlog_ — std::vector<uint8_t> buffers for bytes that read() from the source side could not be immediately flushed to the destination side. Cleared on init() / reset(). * handler_passthrough_forward() now runs a drain pass over the backlog BEFORE the read pass. drain_backlog() loops on write(), EAGAIN means "not now, leave it for the next tick"; any other write error fails the session. * On a partial-write inside the read pass, the unwritten tail is appended to the backlog (append_to_backlog) and the read pass yields. A future pump tick (driven by either the libev EV_READ wakeup on the source or the buffered-work re-arm via `to_process = true`) drains the backlog and resumes reading. * Backlog cap: 1 MiB per direction (PASSTHROUGH_BACKLOG_CAP). Above the cap, append_to_backlog refuses and the session is killed (X_SESSION_CLOSING) — slow- consumer DoS protection. The drain-first contract limits the backlog organically to BURST_BYTES (64 KiB), so the cap is purely defensive against a future logic bug or a larger burst. * to_process = true on a non-empty backlog so the outer dispatch loop re-enters us on the next tick rather than waiting for a new fd-readable event on the source — without this, a long stall on the destination side could let the backlog grow on each new source byte before the previous backlog was drained. Tests (test/tap/tests/unit/mysqlx_session_unit-t.cpp): * test_passthrough_eagain_backlog_drains_across_ticks: drives 256 KiB through a backend leg with SO_SNDBUF/RCVBUF shrunk to 4 KiB so write() returns EAGAIN early and often. Asserts session stays healthy in X_PASSTHROUGH_FORWARD across the repeated EAGAIN events; all 256 KiB are eventually delivered; no byte-level corruption. * test_passthrough_backlog_cap_kills_session: uses the MYSQLX_TEST_BUILD-only seed_passthrough_c2b_backlog_for_test + try_append_to_passthrough_c2b_backlog_for_test helpers to drive the cap branch directly. Asserts within-cap append is accepted, over-cap append is refused, and the session transitions to X_SESSION_CLOSING with healthy=false. Plan count moves from 87 to 98.	2 months ago
Rene Cannao	e855ba37c8	fix(mysqlx): decode backend TLS error code before falling back to plaintext (preferred mode) ... Previously, when a session was in `preferred` TLS mode (try TLS, fall back to plaintext on Error) and the backend responded with ANY Mysqlx::Error to the client's CapabilitiesSet(tls=true), the code silently dropped TLS and proceeded plaintext. Non-TLS errors ("internal server error", "out of memory", "permission denied", etc.) were therefore swallowed and the auth flow proceeded against a backend that just told us it was unhealthy — masking real failures and encouraging brittle deployments. The upstream MySQL X client at plugin/x/client/xsession_impl.cc gates the same fallback exactly on `ER_X_CAPABILITIES_PREPARE_FAILED` (5001), which is the only code that the X plugin's `Capability_tls::set_impl` emits when the server has no SSL context (see plugin/x/src/capabilities/handler_tls.cc — code 5001 is what `ngs::Error(ER_X_CAPABILITIES_PREPARE_FAILED, ...)` produces). Mirror that policy here: * Parse the Mysqlx::Error body's `code` field. * Fall back to plaintext only when the code is exactly 5001. * Any other code (or a code-less / unparseable Error body) is fatal — transition to BACKEND_AUTH_ERROR and return -1. The existing `preferred-mode fallback` unit test (which already injects code 5001) still passes. Two new tests cover the new gate: * preferred-mode + non-TLS error code 1045 (Access denied): no fallback, auth_state=ER, rc=-1. * preferred-mode + Error frame with no decodable code (empty body): no fallback, auth_state=ER, rc=-1. The pre-existing required-mode test (no fallback regardless of code) is unchanged.	2 months ago
Rene Cannao	e4908a1df6	fix(mysqlx): address reviewer feedback (transaction safety, fail-closed TLS, hostgroup refresh, status string) ... Four reviewer-flagged issues from CodeRabbit + Gemini on the consolidated mysqlx stack (originally raised on PRs #5704, #5707, #5709), all addressed in one place since this is the consolidation branch: 1. SQLite transaction safety in stats projections. `MysqlxStatsStore::flush_to_sqlite` and `mysqlx_populate_stats_processlist` previously did a bare DELETE followed by a per-row INSERT loop. A transient SQLite error in any INSERT left the projection table empty, which is far more misleading to operators than stale-but-recent rows. Wrap both in BEGIN/DELETE/INSERT/COMMIT with explicit ROLLBACK on any failure; on rollback, the previous projection stays in place. 2. Fail-closed when backend TLS is required but no SSL_CTX is available. The earlier code in handler_connecting_server had `if (desired_backend_tls) { if (ctx) { /* set up TLS */ } }` — the inner `if (ctx)` was a guard with no else-branch, so a missing SSL_CTX silently produced a plaintext backend connection in TLS-required and AsClient-on-TLS-frontend scenarios. Reject the connect with X-Protocol error 2026 ("Backend TLS required but no SSL context configured on this worker") instead. Caught by CodeRabbit on PR #5707. 3. Stale destination_hostgroup in stats rows. `MysqlxStatsStore:: get_or_create` only set `destination_hostgroup` on first insert. If a route was rebound to a different hostgroup via LOAD MYSQLX ROUTES TO RUNTIME, subsequent traffic continued reporting the first-seen hostgroup forever. Refresh the hostgroup field on every lookup; counters are not reset (only metadata is updated). 4. session_status_to_string missing X_PASSTHROUGH_FORWARD case. `Mysqlx_Thread::session_status_to_string` enumerated every MysqlxSession::Status except the new X_PASSTHROUGH_FORWARD, which fell through to the "UNKNOWN" default. The new state was added in the TLS passthrough work (commit 4d8af5a02) but the stringifier was not updated. Add the explicit case. All four are local fixes with no test changes required — existing unit tests still pass under ASAN (mysqlx_session 87/87, mysqlx_message_dispatch 114/114, mysqlx_stats 26/26, mysqlx_thread 25/25, mysqlx_concurrent 6/6, mysqlx_admin_schema 25/25, mysqlx_backend_auth 58/58, mysqlx_tls 44/44, mysqlx_config_store 33/33). The pre-existing mysqlx_admin_commands_unit-t test 24 failure ("mysqlx_variables has 4 rows after save") predates this branch.	2 months ago
Rene Cannao	0ccf3615f1	fix(mysqlx): classify backend TLS handshake failures into specific error codes ... Closes #5698 (P3 — meaningful TLS error messages on backend failures). Background: prior to this commit, backend TLS handshake failures all emitted a single generic "Backend TLS handshake failed" with code 3152; frontend TLS failures all emitted "TLS handshake failed" code 3151. Operators debugging TLS issues had to attach with `openssl s_client` or run the proxy under SSL_TRACE / SSLKEYLOGFILE to figure out whether the underlying cause was an expired cert, a hostname mismatch, an untrusted CA, or a protocol-version mismatch. MySQL Router translates each of these into a distinct X-Protocol error code and shortcuts the debug session. Implementation: 1. New MysqlxTlsErrorClass enum (HANDSHAKE_FAILED, CERT_VERIFY_FAILED, CERT_EXPIRED, HOSTNAME_MISMATCH, PROTOCOL_MISMATCH, UNKNOWN_CA, NO_SSL_CTX, UNKNOWN). Lives in mysqlx_protocol.h alongside the per-class X-Protocol error code constants. 2. New mysqlx_classify_tls_error(SSL*, peek_err_queue) helper. Classification logic in order: a. SSL == nullptr -> NO_SSL_CTX b. SSL_get_verify_result() != X509_V_OK -> chain class (CERT_EXPIRED for expired/not-yet-valid; HOSTNAME_MISMATCH; UNKNOWN_CA for self-signed-in-chain or unable-to-get-issuer; CERT_VERIFY_FAILED for everything else cert-chain) c. ERR_get_error() walked for SSL_R_UNSUPPORTED_PROTOCOL / TLSV1_ALERT_PROTOCOL_VERSION / WRONG_VERSION_NUMBER / UNKNOWN_PROTOCOL -> PROTOCOL_MISMATCH d. Fallback -> HANDSHAKE_FAILED Cert-chain reasons take precedence because the OpenSSL queue sometimes carries both, and the chain reason is more actionable. 3. Backend codes (5+ distinct, 3152-3157): 3152 BACKEND_TLS_ERR_HANDSHAKE_FAILED (was: only existing code) 3153 BACKEND_TLS_ERR_CERT_VERIFY_FAILED (NEW) 3154 BACKEND_TLS_ERR_CERT_EXPIRED (NEW) 3155 BACKEND_TLS_ERR_HOSTNAME_MISMATCH (NEW) 3156 BACKEND_TLS_ERR_PROTOCOL_MISMATCH (NEW) 3157 BACKEND_TLS_ERR_UNKNOWN_CA (NEW) Backend messages name the OpenSSL reason in operationally-useful terms ("certificate hostname mismatch") but do NOT include the raw OpenSSL queue string in the wire frame — that is logged to stderr separately for the operator's benefit. 4. Frontend codes (3 distinct): 3150 FRONTEND_TLS_ERR_NOT_CONFIGURED (existing — NO_SSL_CTX) 3151 FRONTEND_TLS_ERR_HANDSHAKE_FAILED (existing — every other class) 3158 FRONTEND_TLS_ERR_PROTOCOL_MISMATCH (NEW — operationally useful for legitimate clients hitting a too-old/too-new TLS version) Frontend collapses CERT_*/HOSTNAME_MISMATCH/UNKNOWN_CA onto HANDSHAKE_FAILED specifically to avoid leaking attacker-supplied cert info into the response (asymmetric threat model — the frontend client may be the attacker). 5. New MysqlxConnection::tls_error_class_ member + accessors. Set by step_auth_tls_handshake() at the failure site (the OpenSSL error queue is thread-local FIFO and must be drained while fresh; we can't defer the classification to the session). Read by the session's BACKEND_AUTH_ERROR branch in handler_connecting_server when emitting send_error(). 6. Frontend handler_tls_accept_init() also runs the classifier and emits via the frontend code/message helpers. Logs the OpenSSL queue to stderr (operator-side only). Tests (in mysqlx_tls_unit-t): - test_classify_null_ssl: nullptr -> NO_SSL_CTX (1 ok) - test_classify_cert_expired_via_verify_result (1 ok) - test_classify_hostname_mismatch_via_verify_result (1 ok) - test_classify_unknown_ca_via_verify_result: 2 sub-cases (UNABLE_TO_GET_ISSUER + SELF_SIGNED_IN_CHAIN) (2 ok) - test_classify_generic_cert_verify_failed: SIGNATURE_FAILURE -> generic (1 ok) - test_classify_handshake_failed_default (1 ok) - test_classify_code_message_round_trip_backend: 7 classes * 2 (msg + code-range) + 1 distinct-codes assertion (15 ok) - test_classify_code_message_round_trip_frontend: collapse + 3 distinct-code asserts (4 ok) The cert-fixture-driven tests (real expired/hostname-mismatched/ self-signed cert chains) are explicitly punted per the issue's "don't generate cert fixtures in this PR" guidance. The classifier accepts SSL_set_verify_result-staged synthetic state, which exercises the same code paths as a real cert failure (since the only thing the classifier reads off SSL is the verify_result long); a follow-up issue covers wiring an end-to-end TAP test against a fixture-driven backend. Build: NOJEMALLOC=1 WITHASAN=1 PROXYSQLGENAI=1 plugin .so + build_tap_test_debug all green. Pre-existing RSA_new deprecation warnings unchanged. Tests: mysqlx_tls_unit-t (44/44, was 18 — +26 from 8 new tests). Other suites unchanged. ASAN-clean.	2 months ago
Rene Cannao	767f808309	fix(mysqlx): mark post-Session::Reset connections non-cacheable ... Closes #5697 (P3 — Session Reset pool invalidation). Background: X-Protocol's Session::Reset frame asks the server to wipe session state (current schema, isolation level, charset, prepared statements, session vars) without reauthenticating. ProxySQL forwards the frame and dispatches the backend Ok response correctly — but until this commit, the connection was returned to the per-thread cache as a regular IDLE entry afterwards. The cache key (hostgroup, user, schema, tls_active) doesn't capture session-state-version, so a subsequent reuse from the same identity could pick up the post-reset connection and silently inherit blank state instead of the per-identity defaults. The pool-leak scenario in practice: - Client A issues Session::Reset on backend B. - Backend B is returned to the per-thread cache as IDLE. - Client A2 (same user/schema/hostgroup/tls posture) next request pulls B from the cache. - Client A2 sees blank session state, even though it expected the per-identity defaults a fresh-auth connection would provide. MySQL Router treats Session::Reset as pool-terminating; this commit brings ProxySQL in line. Implementation: - New MysqlxConnection::needs_post_reset_rehandshake_ bool member + accessor pair (set/get). Default false. - is_reusable() returns false when the flag is set, so Mysqlx_Thread::return_connection_to_cache deletes the connection instead of pooling it. - reset() defensively clears the flag too. Production path doesn't rely on this (the cache deletes non-reusable conns instead of resetting them), but the invariant matters for any future code path that calls reset() directly (e.g. retry-on-error). - handler_session_reset_waiting() sets the flag on the Mysqlx.Ok branch — the only path where Session::Reset succeeded. The Error branch already marks the connection non-reusable for unrelated reasons (the Reset itself failed), so no change needed there. We do NOT implement the rehandshake path itself in this commit; the simpler "drop on reset" semantic matches Router's behaviour and is strictly safer than the pre-fix "silently pool blank state". Tests: - mysqlx_connection_unit-t::test_connection_post_reset_rehandshake_flag asserts the flag's defaults, set/get round-trip, is_reusable() interaction, and reset() defensive-clear behaviour (6 ok()). - mysqlx_message_dispatch_unit-t::test_dispatch_sess_reset_marks_non_cacheable drives the full SESS_RESET → backend-Ok → return-to-pool flow with a fake backend and asserts the connection is non-cacheable afterwards. Strengthens the existing test_dispatch_sess_reset which only proved the dispatch reached the backend (5 ok() — pre-flag/state, post-dispatch status, post-condition observed via backend_conn being deleted on the non-reusable path). - mysqlx_connection_unit-t plan switched to plan(0) so future additions don't require re-counting. Build: NOJEMALLOC=1 WITHASAN=1 PROXYSQLGENAI=1 plugin .so + build_tap_test_debug all green, no compile warnings. Tests: mysqlx_connection_unit-t (16/16, was 10), mysqlx_message_dispatch_unit-t (114/114, was 109). Other suites unchanged. ASAN-clean.	2 months ago
Rene Cannao	31a3934014	fix(mysqlx): explicit auth-phase notice forwarding policy in step_auth ... Closes #5695 part 2 (P2 — explicit per-state policy for backend NOTICE frames during the backend auth handshake; data-plane validation landed in the previous commit). Background: prior to this commit, MysqlxConnection::read_auth_frame() silently drained ALL backend NOTICE frames during the BACKEND_AUTH_* state machine. The drain was needed (MySQL backends commonly emit a SESSION_STATE_CHANGED notice before AuthenticateContinue / Ok, and returning nullopt on a NOTICE caused auth to spin until the 10s handshake timeout) but it was indiscriminate — a malformed/unknown-type NOTICE would be drained as silently as a legitimate one. This commit adds an explicit per-state policy: - New helper auth_phase_notice_is_drainable(body, body_len). Parses the NOTICE as a Mysqlx::Notice::Frame and applies a per-type decision: WARNING / SESSION_VARIABLE_CHANGED / SESSION_STATE_CHANGED / GROUP_REPLICATION_STATE_CHANGED / SERVER_HELLO -> drain malformed / empty / unknown enum value -> fail auth WARNING and GROUP_REPLICATION_STATE_CHANGED are also logged on the drain path: WARNING during auth typically signals a server-side misconfig (deprecated auth method, etc.); GROUP_REPLICATION_* during auth is operationally out-of-place (cluster-membership notices belong on data-plane connections, not handshakes). Both are drained to preserve compatibility with over-eager backends but surfaced for operator triage. - Wired into read_auth_frame() between the MAX_LEADING_NOTICES cap and the type==NOTICE branch. On a fail-auth result, the helper has already set auth_state_=BACKEND_AUTH_ERROR and logged; the caller propagates std::nullopt and step_auth fails the connection. - Frontend forwarding during auth: explicitly NEVER allowed. The frontend client has no context to interpret a backend NOTICE before it sees AuthenticateOk; surfacing them would also leak server-side state mid-handshake (the load-bearing concern in issue #5695). The proxy terminates ALL backend NOTICEs on the backend leg during the backend auth phase — even legitimate ones. - Test-only hooks under MYSQLX_TEST_BUILD: a public auth_phase_notice_is_drainable_for_test() and a set_auth_state_for_test() so unit tests can drive the per-type decision matrix directly without running the full step_auth state machine. Tests (in mysqlx_backend_auth_unit-t): - test_auth_phase_notice_known_types_drained: all five spec types are drainable and don't perturb auth_state_ (5*2=10 ok()). - test_auth_phase_notice_unknown_type_fails_auth: type=99 fails auth and sets BACKEND_AUTH_ERROR (2 ok()). - test_auth_phase_notice_malformed_fails_auth: garbage protobuf bytes fail auth (2 ok()). - test_auth_phase_notice_empty_fails_auth: empty body fails auth (2 ok()). - Switched the existing plan(42) to plan(0) so future additions don't require re-counting; existing test_backend_auth_notice_skip (which uses Frame_Type_WARNING) continues to pass under the new policy because WARNING is still on the drainable list. Build: NOJEMALLOC=1 WITHASAN=1 PROXYSQLGENAI=1 plugin .so + build_tap_test_debug all green, no compile warnings. Tests: mysqlx_backend_auth_unit-t (58/58, was 42 — +16 from the 4 new tests), other suites unchanged. ASAN-clean.	2 months ago
Rene Cannao	6ef8cf3f1f	fix(mysqlx): validate Mysqlx::Notice::Frame::type before forwarding ... Closes #5695 part 1 (P2 — explicit notice forwarding awareness; auth-phase per-state policy is the next commit; per-notice metric remains under #5691 follow-up scope). Background: X-Protocol NOTICE frames are non-terminal informational messages a backend can emit at essentially any state transition (warnings, session-state changes, generated-doc-id notifications, group-replication membership changes, etc.). Prior to this commit the proxy forwarded NOTICE frames uncritically; a buggy or hostile backend (or an MITM that bypassed TLS) could ship a NOTICE frame with an unknown `type` enum value and confuse a strict client. The risk is bounded — info disclosure, no auth bypass — but it's unhardened territory MySQL Router handles explicitly. This commit adds a per-frame validation hook on the data plane: - New MysqlxSession::is_notice_frame_valid(body, body_len). Pure query — no session-state mutation. Parses the body as a Mysqlx::Notice::Frame and verifies the outer `type` field is in the spec range (1..5: WARNING, SESSION_VARIABLE_CHANGED, SESSION_STATE_CHANGED, GROUP_REPLICATION_STATE_CHANGED, SERVER_HELLO) using the protoc-generated Frame_Type_IsValid(). Empty body, malformed protobuf, missing required `type` field, and out-of-range enum values all return false. - Wired into handler_waiting_server_msg() and handler_session_reset_waiting() before forwarding NOTICE to the client. Invalid notices are dropped with a stderr log line tagged with route+hostgroup+body_len for operator triage; the response sequence continues (NOTICE is non-terminal in every state). - Test-only hook is_notice_frame_valid_for_test() exposed under MYSQLX_TEST_BUILD so unit tests can drive the predicate directly with synthetic bodies without setting up a fake backend. Tests (in mysqlx_message_dispatch_unit-t): - test_notice_validation_known_types_accepted: all five spec types pass (5 ok()). - test_notice_validation_unknown_type_rejected: types 0, 99, and 100000 are rejected (3 ok()). - test_notice_validation_empty_body_rejected: nullptr body and zero-length body both rejected (2 ok()). - test_notice_validation_malformed_protobuf_rejected: garbage bytes that fail ParseFromArray are rejected (1 ok()). We deliberately do NOT validate the inner `payload` bytes — that's a type-specific protobuf the client parses, and the proxy has no business reaching into it. The outer `type` is the load-bearing field for client-side branching, so that's where we focus. Build: NOJEMALLOC=1 WITHASAN=1 PROXYSQLGENAI=1 make build_lib + plugin .so + build_tap_test_debug all green. Tests: mysqlx_session_unit-t (87/87), mysqlx_message_dispatch_unit-t (109/109, was 98 — +11 from the 4 new tests), mysqlx_compression_unit-t (64/64), mysqlx_tls_unit-t (18/18), mysqlx_thread_unit-t (25/25), mysqlx_concurrent_unit-t (6/6). All ASAN-clean.	2 months ago
Rene Cannao	9ea2b31bf5	fix(mysqlx): match upstream X-Protocol error code for compression-without-algorithm ... Closes #5696 (P2 — MySQL Router parity gap, compression error code). Before this change the mysqlx plugin emitted error code 5008 for every compression-related failure ("Compression frame received without negotiated algorithm", parse failures, decompression failures, etc.). That code does not match upstream MySQL X plugin: 5008 in plugin/x/src/xpl_error.h is ER_X_BAD_CONNECTION_SESSION_ATTRIBUTE_TYPE — a totally unrelated session-attribute error. A client written specifically against the upstream X-Protocol error contract would see the wrong error class entirely. Authoritative source for the canonical codes: plugin/x/src/xpl_error.h in the MySQL 8.4 source tree (deps/mysql-connector-c-8.4.0/.../xpl_error.h in this repo). Mapping: - 5170 ER_X_FRAME_COMPRESSION_DISABLED — compression frame received when the client did not negotiate compression. Upstream message: "Client didn't enable the compression." Reproduced verbatim. - 5174 ER_X_BAD_COMPRESSED_FRAME — structural problem with the Compression envelope (empty body, malformed protobuf, bogus uncompressed_size hint, decompressed payload that doesn't reframe into valid X messages). - 5171 ER_X_DECOMPRESSION_FAILED — algorithm-side failure (lz4/zstd error, OOM, stall, decompressed payload exceeds cap). - 5000 ER_X_BAD_MESSAGE — server-direction compression on the client→server path; treated as a wrong-direction message rather than a compression-specific failure. The pre-auth capability replay-cap error (also previously emitting 5008) is a ProxySQL-specific guardrail with no upstream analogue and is left untouched in this commit — the parity fix is scoped to the genuine compression error sites. Test updates: - mysqlx_message_dispatch_unit-t::test_dispatch_compression_rejected: asserts 5170 (was 5008). - mysqlx_compression_unit-t: oversize/garbage rejections assert 5171, no-negotiation rejection asserts 5170, "no error frame" tests accept any of 5170/5171/5174 (renamed got_5008 → got_compression_err). - test_compression_without_negotiation_still_5008 → test_compression_without_negotiation_still_5170. Build: NOJEMALLOC=1 WITHASAN=1 PROXYSQLGENAI=1 make build_lib + plugin .so + build_tap_test_debug all green. Tests: mysqlx_session_unit-t (87/87), mysqlx_message_dispatch_unit-t (98/98), mysqlx_compression_unit-t (64/64), mysqlx_tls_unit-t (18/18), mysqlx_thread_unit-t (25/25), mysqlx_concurrent_unit-t (6/6) — all ASAN-clean.	2 months ago
Rene Cannao	2e7e532050	feat(mysqlx): X_PASSTHROUGH_FORWARD session state — raw byte splice ... Implements the data-plane half of the per-route TLS passthrough mode (issue #5692). Once a session enters this state the proxy stops parsing X-Protocol frames; bytes are read from one side and written to the other verbatim using read(2)/write(2). Used cases: end-to-end TLS where operator policy forbids proxy MITM (compliance, original cert/SNI/ALPN preservation). Session machine additions: * `MysqlxSession::Status::X_PASSTHROUGH_FORWARD` — terminal state for passthrough sessions. Once entered, the session never returns to any X-Protocol parsing state. handler() short-circuits past client_ds_.read_from_net()/parse so the bytes are not interpreted as frames. * `MysqlxTlsMode::TLS_PASSTHROUGH` — added back to the per-session TLS posture enum. The previous prototype carried the value without a real implementation; that was removed in the asymmetric-TLS series and is now reintroduced as a real feature. * `handler_passthrough_forward()` — pumps both directions per call up to a 64 KiB burst, then yields so other sessions on the thread are not starved. EAGAIN/EWOULDBLOCK is the normal "no more bytes right now" exit. EOF on either side or any other I/O error transitions to X_SESSION_CLOSING. Listener-route propagation: * New `MysqlxSession::init(fd, thread, listener_route)` overload. The base init(fd, thread) overload calls it with an empty route name, preserving existing test harnesses that construct sessions directly. * `Mysqlx_Thread::accept_new_connection()` looks up the listener fd in the parallel `listener_route_names_` vector and propagates the route name to the new session at accept time. This is the hook that lets per-route policies fire before any X-Protocol message is received — most importantly, before any local handshake decision so a passthrough route can splice from the very first byte. Test-only entry path: * `enter_passthrough_for_test(int backend_fd)` (gated behind MYSQLX_TEST_BUILD) drops a session straight into the splice state with a caller-supplied backend fd. Bypasses CapabilitiesSet / auth / resolve_backend_target so the splice mechanics can be asserted in isolation. set_reusable(false) is set on the stub backend connection to mirror the production-side invariant that a passthrough connection never returns to the pool. Tests (mysqlx_session_unit-t, +22 assertions, plan 65 -> 87): * listener_route_propagation: init() captures, reset() clears * forward_client_to_backend / forward_backend_to_client: bytes survive the splice unchanged in both directions * close_on_client_eof / close_on_backend_eof: peer half-close transitions to X_SESSION_CLOSING and marks the session unhealthy * disables_backend_reuse: backend_conn() exists but is_reusable()=false, and tls_mode_ reports TLS_PASSTHROUGH * handler_dispatch_skips_xprotocol: bytes that would parse as a fake X frame in a non-passthrough state survive the splice unconsumed — proves handler() takes the fast path before the X frame parser sees the buffer Production wiring of the entry path (CapabilitiesSet on a passthrough route -> connect to backend, splice the TLS handshake) lands in a follow-up. This commit is intentionally limited to the splice mechanics plus the listener-route plumbing the entry path will need.	2 months ago
Rene Cannao	09c42d4fba	feat(mysqlx): add per-route tls_mode column with passthrough enum value ... Adds the schema and config-store plumbing needed to express MySQL Router's TLS-Passthrough mode as a per-route attribute on `mysqlx_routes`. The actual data-plane forwarding is implemented in a follow-up commit; this commit only wires the config so the operator can set / round-trip / project the value. Schema additions: * `mysqlx_routes.tls_mode VARCHAR NOT NULL DEFAULT 'inherit'` plus a CHECK constraint on the canonical lowercase spellings ('inherit', 'disabled', 'preferred', 'required', 'passthrough'). Same column added to `runtime_mysqlx_routes`. * Default 'inherit' means "use the deployment-wide `mysqlx_tls_mode`", preserving existing behaviour for every row that was loaded before this commit. Operators wanting compliance-pinned passthrough can override one route at a time without flipping the global mode. Config store additions: * `MysqlxRouteTlsMode` enum class (inherit / disabled / preferred / required / passthrough) plus matching `mysqlx_route_tls_mode_from_string` (case- insensitive, empty/NULL -> inherit, unknown -> nullopt) and `_to_string` helpers — same shape as the existing `mysqlx_backend_tls_mode_*` family. * `MysqlxRoute::tls_mode` field with a default of `inherit`. * `MysqlxConfigStore::route_tls_mode(name)` accessor; unknown routes report `inherit` (matching how `route_hostgroup` returns 0 for unknown routes). LOAD / SAVE / project paths: * `install_routes_from_admin` first PRAGMA-probes for the `tls_mode` column so a pre-upgrade admin DB without it loads cleanly (column treated as NULL -> inherit). A malformed value (e.g. operator typo) fails the install with a descriptive `err` rather than silently coercing. * `save_routes_to_admin_table` and `project_routes_to_runtime_view` write the column using `mysqlx_route_tls_mode_to_string`. Tests: * mysqlx_config_store_unit-t gains nine assertions covering parser case-insensitivity, the empty-string / unknown-value contract, the canonical render, the legacy-schema fallback path, and a full install / parse-failure round-trip against a schema that has the column. * All previously-passing unit tests remain green. This commit is intentionally behaviour-preserving: passthrough is parsed and stored but not yet enforced. Resolving a route with tls_mode=passthrough still goes through the proxy-terminated path until the follow-up commit adds the X_PASSTHROUGH_FORWARD session state.	2 months ago
Rene Cannao	744fb0bf36	feat(mysqlx): preferred-mode backend TLS fallback to plaintext + docs ... Closes the loop on issue #5693 (P1: asymmetric TLS / AsClient mode). The previous two commits added the runtime variable and the per- session decision; this commit wires the actual `preferred`-mode fallback path in the backend auth state machine and updates the operator-facing documentation to describe the full four-mode model. What changes: * MysqlxConnection::step_auth_capabilities_set_sent now treats a Mysqlx::Error response from CapabilitiesSet(tls=true) as a fall-back trigger when backend_tls_fallback_allowed_ is set (set by the session for mode=preferred). The connection silently downgrades on the same TCP socket — backend_tls_required_ is cleared so subsequent steps don't re-attempt TLS, tls_active_ stays false (which keeps the connection out of the encrypted half of the per-thread pool), and step_auth proceeds straight into AuthenticateStart on plaintext. Under mode=required (and AsClient + frontend-TLS), the same Error is still fatal — the operator's policy demands encryption. The pre-existing `Frame is not OK` rejection branch is preserved for malformed / unexpected response types. * doc/mysqlx/README.md §8 rewritten: * §8.1 frontend TLS modes (mysqlx_tls_mode, unchanged behaviour). * §8.2 backend TLS modes (mysqlx_tls_backend_mode, full table of the four documented values, migration notes from MySQL Router AsClient, and a connection-pool partitioning note explaining that the cache key now includes tls_active. * §8.3 configuration example showing the asymmetric pattern (frontend REQUIRED + backend preferred). The §3.1 variables table gains the new variable. * doc/mysqlx/MYSQL_ROUTER_PARITY.md updated: * Feature comparison row 4 (TLS modes) now lists 4-frontend x 4-backend with explicit mention of mysqlx_tls_backend_mode and issue #5693. * Asymmetric TLS row marked implemented. * Architecture section §TLS rewritten to describe both the frontend mysqlx_tls_mode and backend mysqlx_tls_backend_mode models; identifies passthrough as the remaining gap. * Feature gaps table marks "Asymmetric TLS / AsClient" and "Per-message response state machines" as implemented. * Summary updated. Tests: test/tap/tests/unit/mysqlx_backend_auth_unit-t (was 34): adds 7 new ok lines covering both branches of the fallback decision: * test_backend_auth_preferred_mode_fallback_to_plaintext: 5 assertions across the wire-level transition (state on entry, rc=1 on Error, AUTHENTICATE_START_SENT after fallback, backend_tls_required_ cleared, tls_active false, AuthStart frame on the wire). * test_backend_auth_required_mode_no_fallback_on_error: 2 assertions confirming the same Error is fatal when fallback_allowed=false. Total 42. All 8 (mode x frontend_tls) acceptance combinations from the issue are now covered: 6 by the decision-function tests in mysqlx_message_dispatch_unit-t (commit 2), and the remaining 2 mode=preferred + backend Error cases by the wire-level fallback tests added here. Tested under NOJEMALLOC=1 WITHASAN=1 PROXYSQLGENAI=1. Closes #5693. Stacks on #5706 / #5704.	2 months ago
Rene Cannao	167a896c16	feat(mysqlx): mode-driven backend TLS decision + tls_active conn-cache key ... Implements the per-session decision side of issue #5693 (P1: asymmetric TLS / AsClient mode parity gap with MySQL Router 8.0). The previous commit added the runtime variable; this commit wires it through to the actual backend-connect path and partitions the connection pool by encryption posture so AsClient/required-TLS sessions never reuse plaintext-pooled backends (and vice versa). What changes: * Lifts the per-session decision out of an inline switch in MysqlxSession::handler_connecting_server() into a pure helper `mysqlx_resolve_backend_tls_decision(mode, endpoint_override, frontend_is_encrypted) -> {require_tls, fallback_allowed}`. The helper lives at file scope (not in the anonymous namespace) so the unit test can exercise the 8 (mode x frontend_tls) combinations called out in the issue acceptance criteria directly, without driving the full session state machine. The decision itself replaces the legacy `target_use_ssl_ || client_ds_.is_encrypted()` expression. * Adds two fields on MysqlxConnection: * tls_active_ -- set true by step_auth_tls_handshake() once the OpenSSL handshake completes successfully on the backend leg. Read by the connection cache to partition encrypted-pooled connections from plaintext-pooled ones. * backend_tls_fallback_allowed_ -- carries the `preferred` mode's "downgrade-to-plaintext-on-error" intent through to the auth state machine. Read-only metadata in this commit; the actual fallback path lands in the next commit. * Mysqlx_Thread::get_connection_from_cache now takes a `bool tls_active` parameter and matches on it. Without this, an AsClient TLS session pulling a pooled connection could land on a plaintext backend (or vice versa), corrupting the wire protocol when the next dispatched frame goes out over a socket in the wrong encryption posture. * Endpoint-override semantics (mysqlx_backend_endpoints.use_ssl=1) preserved exactly: it can promote plaintext to TLS regardless of mode, but cannot demote a TLS-required mode to plaintext. Under mode=preferred, the override leaves fallback_allowed=true so the operator's best-effort "preferred" intent isn't silently upgraded to "required". Tests: test/tap/tests/unit/mysqlx_message_dispatch_unit-t (was 86): adds 8 combinations (mode x frontend_tls) + 4 endpoint-override cases = 12 new ok lines, total 98. test/tap/tests/unit/mysqlx_thread_unit-t (was 22): adds test_connection_cache_tls_partition with 3 ok lines exercising the (plaintext, encrypted) hard partition, total 25. Known limitation (TODO addressed in next commit): mode=preferred currently fails the backend connect on Mysqlx::Error from CapabilitiesSet(tls=true) because the fallback-to-plaintext branch isn't wired in step_auth_capabilities_set_sent yet. The metadata (`backend_tls_fallback_allowed_` on the connection) is in place; only the state machine branch is missing. Two of the 8 documented test combinations (mode=preferred + Error from backend) are validated against the metadata only in this commit; the end-to-end fallback behaviour follows. Tested under NOJEMALLOC=1 WITHASAN=1 PROXYSQLGENAI=1. Refs #5693. Stacks on #5706 / #5704.	2 months ago
Rene Cannao	0246237426	feat(mysqlx): add MysqlxBackendTlsMode enum + mysqlx_tls_backend_mode runtime variable ... Wires up the configuration plumbing for issue #5693 (P1: asymmetric TLS / AsClient mode parity gap with MySQL Router 8.0). This commit is behaviour-neutral: the new variable is parsed, validated, persisted and exposed via MysqlxConfigStore::get_backend_tls_mode(), but the per- session backend-TLS decision still uses the legacy target_use_ssl_ || client_ds_.is_encrypted() expression. The decision site is rewritten in the next commit. What this adds: * MysqlxBackendTlsMode enum with four values matching MySQL Router's client_ssl_mode / server_ssl_mode taxonomy: disabled, preferred, required, as_client. Default is as_client because that most closely matches the legacy implicit behaviour where the backend leg encryption was tied to the frontend leg's encryption. * mysqlx_backend_tls_mode_from_string() / mysqlx_backend_tls_mode_to_string() for case-insensitive parsing and canonical lower-case rendering. The parser returns std::optional so the install path can surface a useful error to the operator on a typo instead of silently coercing to a default. * MysqlxConfigStore now reads the mysqlx_tls_backend_mode key from mysqlx_variables in install_variables_from_admin(), persists it via save_variables_to_admin_table(), and projects it in project_variables_to_runtime_view(). install fails atomically with a descriptive error when the value is unrecognised; an absent row leaves the cached mode untouched (matches how the other tunables already behave). * MysqlxBackendEndpoint.use_ssl=1 remains an operator-controlled override that forces TLS regardless of the mode (per existing comment at handler_connecting_server). The mode interacts with that flag in the next commit. Tests: test/tap/tests/unit/mysqlx_config_store_unit-t: 16 -> 24 assertions. New coverage: parser accepts all four documented values case-insensitively, parser rejects unknown values, default mode is as_client, LOAD round-trip caches the parsed mode, invalid value fails install with descriptive error, store retains last-good mode after rejected install, absent row leaves cached mode untouched. Tested under NOJEMALLOC=1 WITHASAN=1 PROXYSQLGENAI=1. Stacks on PR #5706 (mysqlx response state machines), which itself stacks on PR #5704 (mysqlx observability P0). Refs #5693.	2 months ago
Rene Cannao	170ead7ad6	fix(mysqlx): reject backend frames disallowed in current response state ... Tightens MysqlxSession::is_frame_allowed: a backend frame outside the per-state allowed set now closes the session with X-Protocol Error 4006 ("Backend sent an unexpected message in the current response state") instead of being forwarded blindly to the client. This guards against a buggy or hostile backend pushing a frame whose shape the client cannot parse, which would silently desync the wire and (in the worst case) amplify a backend protocol bug into a client-side fault. Adds a new private sub-state field MysqlxSession::seen_column_metadata_ that gates RESULTSET_ROW frames in the four response states where the X-Protocol requires ColumnMetaData to precede any Row: - RESP_WAITING_STMT_EXECUTE - RESP_WAITING_CRUD - RESP_WAITING_PREPARE_EXECUTE - RESP_WAITING_CURSOR_OPEN The flag is set when ColumnMetaData is forwarded in the validation loop, cleared at every transition into a state that begins a new column- metadata sequence (in dispatch_client_message), at terminal-frame flush in handler_waiting_server_msg, and at init() / reset(). It is deliberately NOT cleared on transition into RESP_WAITING_CURSOR_FETCH — per the X-Protocol spec, ColumnMetaData is sent at Cursor::Open and not re-sent at Cursor::Fetch, so CURSOR_FETCH's allowed-set unconditionally accepts RESULTSET_ROW. The dispatch handler skips the clear at CURSOR_FETCH entry to make this carry-across explicit. The rejection action in handler_waiting_server_msg pops the offending frame from the backend queue (does not forward), emits a fatal error frame to the client, marks backend_conn_ non-reusable so return_backend_to_pool deletes it instead of caching a poisoned connection, sets healthy=false and status_=X_SESSION_CLOSING, and short-circuits before the bytes_recv accounting (so the disallowed frame is not double-counted as forwarded traffic). Tests in mysqlx_message_dispatch_unit-t (20 new assertions across 8 cases) cover: 1. StmtExecute → Row-without-metadata (canonical hostile-backend case the issue called out): expect Error frame, !healthy, X_SESSION_CLOSING. 2. StmtExecute → ColumnMetaData → Row → SQL_STMT_EXECUTE_OK: happy path, terminal flushes seen_column_metadata_. 3. CursorOpen → ColumnMetaData → Row → FETCH_SUSPENDED: terminal, response_state_ resets, flag cleared at boundary (Cursor::Fetch does not consult it). 4. CursorOpen → ColumnMetaData → FETCH_DONE: terminal, clean exit. 5. PreparePrepare → SQL_STMT_EXECUTE_OK (allowed-set is OK-only): expect rejection, X_SESSION_CLOSING. 6. PreparePrepare → OK: happy path. 7. STMT_EXECUTE → ColumnMetaData → NOTICE → Row → SQL_STMT_EXECUTE_OK: NOTICE is universal-allowed, non-terminal, doesn't consume the response. 8. CURSOR_FETCH → Row (with seen_column_metadata_=false): forwarded, NOT rejected (the per-state-pair carve-out for Cursor::Fetch). Tests use a new setup_session_for_validation helper that drives the auth flow, attaches a fake backend over a socketpair, parks the session in WAITING_SERVER_XMSG with the desired response_state_ via the test-only set_response_state_for_test (gated behind MYSQLX_TEST_BUILD), and lets the test write synthetic server frames to the backend half of the pair. Drains pending client-side bytes via a non-blocking helper so the validation hook's output is unambiguous. Closes the rejection-side acceptance criterion of #5694. Refs: #5694.	2 months ago
Rene Cannao	34968e08d4	fix(mysqlx): split CURSOR/PREPARE response states for accurate terminal detection ... CURSOR_OPEN's terminal set was conflated with CURSOR_FETCH's, and PREPARE_PREPARE/DEALLOCATE were lumped with PREPARE_EXECUTE which has a wider terminal set. Adds RESP_WAITING_CURSOR_{OPEN,FETCH,CLOSE} and RESP_WAITING_PREPARE_{PREPARE,EXECUTE,DEALLOCATE} so each X-Protocol response shape gets its own per-state contract. The two old conflated states accepted the union of all terminal frames that any sub-shape could legitimately emit, so the proxy would advance to RESP_IDLE on the first such frame regardless of whether the backend's response was actually shaped right. With the split: - PREPARE_PREPARE / PREPARE_DEALLOCATE / CURSOR_CLOSE accept Mysqlx.Ok only - PREPARE_EXECUTE accepts Ok / SQL_STMT_EXECUTE_OK / FETCH_DONE / FETCH_SUSPENDED (it inherits the response shape of whichever request was prepared) - CURSOR_OPEN / CURSOR_FETCH accept FETCH_DONE / FETCH_SUSPENDED This is purely a tightening of is_terminal_frame; the validation hook that uses these per-state contracts to actively reject out-of-shape backend frames is added in the next commit. No behaviour change for well-behaved backends; existing dispatch tests (66 assertions in mysqlx_message_dispatch_unit-t, 65 in mysqlx_session_unit-t) continue to pass unmodified. Refs: #5694.	2 months ago
Rene Cannao	b812045e11	refactor(mysqlx): split is_terminal_for_state into is_frame_allowed + is_terminal_frame ... No behavior change. Permissive is_frame_allowed (allows everything the existing code implicitly allows by forwarding) sets up the validation hook for the follow-on commits. Removes the now-unused is_terminal_server_frame_generic helper since every state has an explicit case. The new is_terminal_frame variant centralises the universal-NOTICE rule: NOTICE is never terminal in any state, ERROR always is. Previously the caller in handler_waiting_server_msg() short-circuited NOTICE at the call site; folding that into the predicate keeps the validation hook (added in the next commits) from having to know about NOTICE separately. Refs: #5694.	2 months ago
Rene Cannao	ff9a41871d	feat(mysqlx): project per-session state into stats_mysqlx_processlist ... stats_mysqlx_processlist had its DDL registered (mysqlx_admin_schema.cpp: kStatsMysqlxProcesslistTable) since the plugin-chassis merge but no writer ever existed. Operators querying the table got nothing back, no matter how many concurrent X-Protocol clients were connected. Issue #5691 caught this as the second of the two fully-empty stats surfaces. Wires the projection through the chassis ABI 3 register_runtime_view hook landed in PR #5688 — same pattern as the four runtime_mysqlx_<X> views, mirroring how core's stats___mysql_processlist re-projects on every admin SELECT. Three new pieces: 1. MysqlxSessionSnapshot struct (plugins/mysqlx/include/mysqlx_thread.h) captures one row's worth of session state: username, route name, worker_id, backend host:port, auth_mode, connection_state, and session_age_ms. bytes_in/bytes_out are reserved at 0; the per-route counters in MysqlxStatsStore aggregate by route, not by session, so filling these is P1 work. 2. Mysqlx_Thread::snapshot_sessions_for_stats(out, now_ms) walks sessions_ under sessions_mutex_ and appends one snapshot per active session. Lock scope is bounded (string copy + a few struct field reads); no I/O under the mutex; no cross-thread lock is held. Safe to call from any thread and intended for the chassis runtime-view refresh callback path. 3. mysqlx_populate_stats_processlist(SQLite3DB&) iterates over mysqlx_context().threads, gathers all snapshots into a flat vector under each thread's lock, then DELETEs and re-INSERTs the rows on statsdb. DELETE always runs (empty thread pool / empty session list both mean "no active sessions" — operators must see that, not stale state from the previous refresh). Wires three small read-only state observers on MysqlxSession (username_for_stats, route_name_for_stats, identity_for_stats, start_time_for_stats) so the snapshotter can read what it needs without needing friend-of-thread or exposing the session's mutable fields. These getters return values a debugger could already observe and cannot mutate the session. The refresh callback (refresh_stats_processlist_view in mysqlx_admin_schema.cpp) follows the same get_statsdb-via-services pattern as refresh_stats_routes_view, with one extra null-pointer guard: mysqlx_populate_stats_processlist is __attribute__((weak)) so the admin-schema unit test (which compiles mysqlx_admin_schema.cpp but doesn't link mysqlx_plugin.cpp) still links — the runtime null check is the safety net for that test build. Tests: relies on the existing fixture coverage in mysqlx_thread_unit-t and mysqlx_concurrent_unit-t to exercise the session-walk path without regression. A standalone snapshotter test would need to construct fake Mysqlx_Thread + MysqlxSession instances, which was scope-creep for P0 — defer to integration TAP coverage as the issue body recommends. All 9 affected unit suites stay green (255 assertions, no leaks introduced under ASAN). Known limitation, same as the runtime-view callback in the previous commit: SELECTs against the stats port (port 6032 with stats credentials) bypass the chassis dispatcher and return whatever's in the table since last refresh. Filing as a separate chassis-side follow-up. Refs: #5691.	2 months ago
Rene Cannao	796f336cb0	feat(mysqlx): expose stats_mysqlx_routes via runtime-view refresh ... stats_mysqlx_routes was previously declared (DDL registered in mysqlx_admin_schema.cpp:466) but had no writer — the MysqlxStatsStore::flush_to_sqlite() projection method was implemented months ago and has never been called. Issue #5691 caught that "no caller" gap; admin SELECTs against the table return whatever's left from previous test fixture data, which is empty in normal operation. Wires the projection through the chassis ABI 3 register_runtime_view hook landed in PR #5688. Identical shape to the four runtime_mysqlx_<X> projections already in place for users / routes / endpoints / variables: chassis fires the refresh callback before any admin SELECT references the registered table name, callback rebuilds the projected rows from the canonical in-memory state. The wrinkle vs the existing four: stats data lives in statsdb, not admindb. The chassis-supplied admindb argument is ignored; refresh_stats_routes_view fetches the statsdb handle through the plugin's cached services pointer (mysqlx_context().services-> get_statsdb()) and hands that to flush_to_sqlite. flush_to_sqlite writes a bare table name, so reaching stats_mysqlx_routes via the admindb-attached `stats.` schema would work but adds an unnecessary detour and forces a write through the wrong handle. Mirrors how core's stats___mysql_processlist() writes statsdb directly (lib/ProxySQL_Admin_Stats.cpp:885). Known limitation, flagged in the issue: the chassis only fires plugin runtime views on admin-session SELECTs (gated on `if (admin)` in lib/ProxySQL_Admin.cpp). SELECTs against the stats port (port 6032 with stats credentials) bypass the dispatcher entirely → return the table contents at last refresh. Filing as a separate chassis- side follow-up. Plumbing additions: - mysqlx_admin_schema.cpp pulls in mysqlx_stats.h and defines refresh_stats_routes_view (~10 lines). - The single-line register_runtime_view call goes inside the existing tail-extension guard (`if (services.register_runtime_view != nullptr)`) — same pattern as the other four. - test/tap/tests/unit/Makefile: three test build rules (mysqlx_admin_schema_unit-t, mysqlx_admin_commands_unit-t, mysqlx_admin_disk_commands_unit-t) now also need to compile mysqlx_stats.cpp into the test binary, since mysqlx_admin_schema.cpp now depends on it. Single-line addition per rule. Refs: #5691.	2 months ago
Rene Cannao	04f771d2a3	feat(mysqlx): account per-route bytes_sent / bytes_recv on the data plane ... The MysqlxRouteStats struct already had bytes_sent / bytes_recv atomic fields, but no setter and no caller — issue #5691 caught the third and fourth dead counters on the same dead-code list as conn_ok / conn_used. Adds two setters to MysqlxStatsStore: - record_bytes_sent(route, hg, n) - record_bytes_recv(route, hg, n) Both fetch_add(n, relaxed) under the existing mutex, with a 0-arg fast path that skips the lock entirely (frame.size() == 5 means a header- only frame and is the natural early-out — no underflow guard needed on `size - 5` because the call sites only fire when size > 5). Wires them at the two natural data-plane sites: - MysqlxSession::forward_to_backend — bytes_sent is "X-Protocol payload bytes the proxy forwarded from client to backend", measured as `frame.size() - 5` (strip the 5-byte X-Protocol frame header so the counter tracks operator-meaningful query/CRUD bytes rather than wire-level total). - MysqlxSession::handler_waiting_server_msg — bytes_recv is the same payload-only measure for the backend → proxy → client leg. NOTICE frames are counted (they're part of the data plane the operator paid for forwarding); the OK / EXECUTE_OK / FETCH_DONE family contributes 0 because their payloads are typically empty. Both counters need the route name available at the I/O site without re-reading identity_->default_route per frame. Caches that string into MysqlxSession::route_name_ at the resolve_backend_target() success tail and clears it in init() and reset(). The dataplane sites read route_name_ directly. (identity_ would not be null at those sites in practice — resolve_backend_target() is gated on a populated identity — but route_name_ avoids the optional-deref entirely and side-steps any future code path that nulls identity_ in-flight without a corresponding state transition.) Tests: extends mysqlx_stats_unit-t.cpp to plan(26), adding 4 assertions covering the new bytes_sent / bytes_recv accumulation through flush_to_sqlite, including the 0-arg no-op invariant and that destination_hostgroup is carried through. Refs: #5691.	2 months ago
Rene Cannao	b1f973b163	feat(mysqlx): record conn_ok and conn_used in handler_connecting_server ... The mysqlx plugin's MysqlxStatsStore declares five per-route counters (conn_ok, conn_err, conn_used, bytes_sent, bytes_recv) but issue #5691 caught that, of those five, only conn_err is incremented from production code (from the failure modes of resolve_backend_target). The other four are dead. Wires the two connection-success counters at their natural sites in MysqlxSession::handler_connecting_server: - record_conn_used at the cache-hit early-return branch (took an existing pre-warmed backend connection out of the per-thread pool in Mysqlx_Thread::get_connection_from_cache; no fresh TCP, no fresh TLS, no fresh backend auth performed). - record_conn_ok at the fresh-connection success tail (TCP connect, optional TLS handshake, and step_auth() all returned non-error, no early return fired). This is the "established a brand new backend connection from scratch" counter — strictly disjoint from conn_used. Both call sites already have target_hostgroup_ in scope (set by resolve_backend_target before this state is reachable) and identity_ (set during the frontend auth flow). The route-name argument falls back to the empty string when identity_ is null, defensively — in practice it cannot be null past resolve_backend_target's success path, but the ternary avoids a crash if the state machine ever ends up here without one and reads identity_->default_route via the dot operator on a moved-from optional. The bytes_sent/bytes_recv counters and the runtime-view projection that surfaces these in stats_mysqlx_routes are wired in follow-on commits per the P0 plan on #5691. Refs: #5691.	2 months ago
Rene Cannao	29402b4a42	docs(plugins/genai): close 3 small gaps from review ... # 1. Real prepared statement in local_proxy_endpoint.cpp read_global_variable's comment claimed "Parameterised lookup via prepared statement" but the implementation did string concatenation: "SELECT variable_value FROM main.global_variables WHERE variable_name='" + std::string(var_name) + "'" Today's only callers pass hardcoded "mysql-interfaces" / "pgsql-interfaces", so the SQL injection surface is zero — but the comment is a footgun for the next caller and we have a working prepare_v2 / bind_text pattern right next door in plugin_main.cpp. Replace with the proper prepared-statement form using proxy_sqlite3_step + proxy_sqlite3_column_text; semantics (empty-string on miss / error) are unchanged. # 2. Concurrency contract on genai_refresh_runtime_components The function (introduced in 9e4244321) tears down and reinitialises GloGATH and GloAI in place — with no rwlock around either global. Today it's called from: 1. genai_start (single-threaded, no traffic) — always safe 2. LOAD GENAI VARIABLES TO RUNTIME admin command — racy with active MCP listener / tool-handler threads 3. LOAD GENAI VARIABLES FROM CONFIG — same as (2) Document the contract honestly in the header: caller must guarantee no concurrent dereference of GloGATH/GloAI; today the practical mitigation is operator-side (quiesce traffic before issuing the LOAD). A proper rwlock around all GloGATH/GloAI consumers is a follow-up — too large for this PR's scope, requires audit of every listener/tool-handler reader. # 3. New plugins/genai/README.md The plugin is now ~30 source files spanning the MCP listener, GenAI worker pool, AI features manager, anomaly detector, schema discovery cache, FTS index, vector storage, LLM bridge, and 9 tool handlers — with no single doc explaining what it owns or how to operate it. Mirror the plugins/mysqlx pattern: a one-page README covering scope, building, loading, lifecycle, admin SQL surface, configuration pointer, observability, concurrency contract, testing, and known gaps. Cross-references doc/MCP/ and doc/LLM_Bridge/ for the deeper subsystem-internal documentation that already exists.	2 months ago
Rene Cannao	b501fa17dc	genai: harden real persistence and connection bugs ... Fix the remaining concrete defects from the GenAI review pass: - serialize MCP variable access with the handler rwlock instead of relying on external callers - make GenAI/MCP save-to-admin operations transactional and fail closed on SQLite errors - clamp genai_log() newline handling after truncation - use PQconnectdbParams() for libpq connections so credentials and schema values are passed safely - install and uninstall the GenAI plugin alongside mysqlx in the top-level Makefile - add a regression test for stale output handling in get_variable_string() These changes keep the runtime and persistence paths consistent with the branch's plugin lifecycle fixes.	2 months ago
Rene Cannao	967ab401ec	genai: fix lifecycle reloads and drop dead MCP stats registration ... Load GenAI and MCP variables before runtime initialization, reinitialize the AI stack on reload, and make the anomaly hook obey the configured enable flags. Restore the missing MCP command aliases and the disk/config reload verbs, fix the MySQLX packaging filename mismatch, and update the unit harness for the plugin ABI. Also remove the stale stats_mcp_* registrations from the plugin surface. There is no plugin-side writer for those tables in this branch, so advertising them as live schema was misleading. Update the unit test and bootstrap comments to match the actual exposed surface.	2 months ago
Rene Cannao	978a6f8d2f	fix(plugins/genai): blocker fixes from PR-#5701 review ... Five concrete fixes uncovered by the four-agent review (architecture, code-safety, build/packaging, test coverage): # Blocker 1: per-row INSERT failure leaves transaction half-done In the six MCP_Threads_Handler methods that follow the BEGIN -> DELETE -> prepare -> loop INSERT -> COMMIT pattern, the SAFE_SQLITE3_STEP2 macro inside the loop didn't check `rc`. If a single INSERT failed (constraint violation, disk full, ...), the loop would continue binding subsequent rows and the COMMIT below would either succeed with partial data or be left dangling. Hoist the per-row binding into three insert_*_row(stmt, row) -> bool helpers in an anonymous namespace, then have every caller check the return value and `ROLLBACK; return false/return;` on failure. Affects save_auth_profiles_to_admin_table, save_target_profiles_to _admin_table, save_query_rules_to_admin_table, project_auth_profiles _to_runtime_view, project_target_profiles_to_runtime_view, project_query_rules_to_runtime_view, plus the new save_profiles_to _admin_table below. # Blocker 2: cross-table install/save atomicity for profiles target.auth_profile_id is an FK to auth.auth_profile_id. The old mcp_load_target_auth_map_from_admindb / mcp_save_target_auth_map_to_ admindb wrappers ran the two installs / two saves as separate operations under separate locks / transactions. A failure of the second step (or a concurrent LOAD/SAVE racing in between) left target_auth_map rebuilt from a (auth_v2, target_v1) mismatch, or left main.mcp_target_profiles referencing auth_profile_ids that didn't exist in main.mcp_auth_profiles. Add atomic combined methods: - install_profiles_from_admin: reads both source tables BEFORE taking the wrlock (so a slow admindb doesn't gate the listener), then under one wrlock swaps both vectors and rebuilds the joined target_auth_map exactly once. - save_profiles_to_admin_table: snapshots both vectors under one rdlock, then deletes target before auth (FK direction-aware) and inserts auth before target inside a single BEGIN ... COMMIT. The plugin_main.cpp wrappers now route through these atomic methods. The per-table install_*_from_admin / save_*_to_admin_table public methods stay (used by future granular admin verbs and unit tests), but each rebuilds target_auth_map after its own swap as before. # Blocker 3: unit-test PROXYSQLGENAI autodetect probe test/tap/tests/unit/Makefile had been probing libproxysql.a for the `MCP_Tool_Handler` symbol to autodetect PROXYSQLGENAI. Step 7 of the carve-out moved that symbol to plugins/genai/, so the probe silently found 0 matches and stopped defining -DPROXYSQLGENAI for unit-test builds. Switch the probe to invoke_register_schemas_phase (the chassis-exclusive symbol that's already used to probe PROXYSQL40) so the macro auto-defines whenever the chassis is available — which matches when genai-plugin unit tests can actually be built and exercised. # Blocker 4: triplet coverage through dispatch surface Extend genai_plugin_load_unit-t to drive the full install/save/project flow through the production dispatch path: - seed mcp_<X> tables in admindb - dispatch "LOAD MCP PROFILES TO RUNTIME" via mgr.dispatch_admin_ command (same path admin SQL takes) - manually invoke refresh_runtime_views_for_query (the chassis hook that fires on every admin SELECT) and assert runtime_mcp_<X> rows match the source - mutate main.<X> directly to simulate an operator stomp, then dispatch "SAVE MCP PROFILES TO MEMORY" and assert the in-memory snapshot was written back over the stomp This is the end-to-end coverage that plugin_runtime_views_unit-t can't provide (it uses synthetic callbacks) and that a standalone genai_mcp_config_store unit test would have provided at the cost of stubbing 9 plugin classes. Plan bumped from 25 -> 36. # Should-fix: atomic fence + GloGATH lifetime genai_anomaly_embed_fn was std::atomic with relaxed ordering; the comment claimed nothing else needed synchronising. Wrong: the embed function dereferences GloGATH, a non-atomic global mutated by genai_init / genai_stop. Pair the load with acquire and the stores with release so a reader that observes the non-null pointer also observes the prior `GloGATH = new ...`, and a reader that observes the post-stop nullptr short-circuits before GloGATH's deletion is visible. Acquire/release is free on x86_64 / arm64 — no perf regression. # Should-fix: harden DEB awk substitution + RPM glob DEB entrypoint: replace the inline awk -v repl="$..." substitution (which mangled paths containing awk metacharacters via gsub) with a file-backed awk getline that reads the replacement block verbatim. Add a fail-fast grep for the placeholder after substitution so a silently-broken substitution can't ship a malformed package. RPM (rhel + suse): the spec gates `/usr/lib/proxysql/*.so` on %if 0%{?with_plugins}, but rpmbuild aborts on "File not found by glob" if the directory is empty (silent plugin build failure + with_plugins=1). Have the entrypoint compgen-check whether any .so actually got staged and only pass --define "with_plugins 1" when true.	2 months ago
Rene Cannao	482ce386cd	chore(genai-carveout): strip dead PROXYSQLGENAI blocks + plugin hardening ... Two threads of cleanup that had to land together because they touch both core (where the dead code lives) and the plugin (where the new module-owned state lives): # Strip dead #ifdef PROXYSQLGENAI from lib/, include/, src/ (Spec Q9) Step 7 of the carve-out left PSQLGA empty in lib/Makefile, so every #ifdef PROXYSQLGENAI block in the core build became permanently false. The blocks remained as dead text (visible to readers, irrelevant to the binary). Strip ~600 lines spanning: src/main.cpp 2 sites lib/Admin_Bootstrap.cpp 3 sites (mcpdb open + attach + flush_*_variables hooks) lib/Admin_FlushVariables.cpp 2 sites (~250 lines of stub + #if 0 _ORIGINAL bodies) lib/Admin_Handler.cpp 5 sites (LOAD/SAVE GENAI VARIABLES dispatch + LOAD/SAVE MCP VARIABLES dispatch + MCP profiles handler) lib/ProxySQL_Admin.cpp ~10 sites (init_*_variables + load_mcp_server + load/save_mcp_query_rules + GenericRefreshStatistics MCP probes) lib/ProxySQL_Admin_Stats.cpp 3 stats___mcp_* function bodies include/proxysql_admin.h 7 method declarations include/ProxySQL_Admin_Tables_Definitions.h 4 unused DDL macros The plugin now owns LOAD/SAVE MCP VARIABLES, LOAD/SAVE GENAI VARIABLES, LOAD/SAVE MCP PROFILES, LOAD/SAVE MCP QUERY RULES via its admin command registry; the chassis dispatcher catches them before any of the dead core branches would have fired. Caveat — known break: the `stats_mcp_*` tables are still registered (empty schema) but the populator path that fed them is gone. SELECTs return empty result sets until a follow-up wires the populator into the plugin (likely via register_runtime_view or a parallel stats-side ABI). Tracked separately. # Plugin hardening (review-feedback items 4-7) * plugins/genai/Makefile: add -fvisibility=hidden, -fvisibility-inlines-hidden, -fstack-protector-strong (mirrors plugins/mysqlx). Tag proxysql_plugin_descriptor_v1 with __attribute__((visibility("default"))) so the loader's dlsym still resolves; without the explicit annotation the hidden-default flag would also hide the entry point. * plugins/genai/src/Anomaly_Detector.cpp: convert genai_anomaly_embed_fn to std::atomic<…> with relaxed loads/stores. The detector hot-path read races with init/stop on the lifecycle thread; atomics give us a portable no-torn-pointer guarantee without paying for acquire/release fences we don't need (no other state synchronises with the pointer). Replaces the hand-wavy "no mutex needed" comment with the actual reasoning. * plugins/genai/src/{plugin_main,plugin_commands,plugin_tables}.cpp: route the 17 fprintf(stderr, "genai plugin: ...") log lines through a new genai_log() helper that calls services->log_message when the chassis is wired up and falls back to stderr for early-init / unit-test paths. Lines now land in proxysql.log alongside the rest of the runtime. * plugins/genai/include/MCP_Thread.h: add get_variable_string(name, std::string&) returning a std::string, and switch mcp_save_variables_to_admindb (plugin_main.cpp) to it. The legacy get_variable(name, char*) had an unbounded sprintf into a stack buffer and the SAVE path was passing char val[256] — a real overflow risk for *_endpoint_auth bearer tokens. * plugins/genai/src/plugin_main.cpp: rewrite the genai_init / genai_stop comment block with the actual teardown ordering rationale (consumers before producers; atomic-clear-then-delete for the embed hook; why Prometheus counters stay registered).	2 months ago
Rene Cannao	16253a1c99	feat(plugins/genai): adopt ABI-3 separation-of-duties for runtime_mcp_* ... Wire the genai plugin into the plugin-chassis ABI-3 contract added in PR #5688: Admin owns the editable mcp_<X> tables, the module owns the in-memory snapshot under its own mutex, and runtime_mcp_<X> is a chassis-projected view repopulated lazily by a refresh callback before each admin SELECT. No more triple-storage of profile / query-rule data. # MCP_Threads_Handler: per-table install / save / project triplets Replaces the monolithic load_target_auth_map(joined-resultset) with three independent triplets, mirroring the mysqlx config_store rework: install_<X>_from_admin reads main.mcp_<X>, replaces the in-memory snapshot vector under wrlock, rebuilds the joined target_auth_map (for profiles) save_<X>_to_admin_table REPLACEs main.mcp_<X> from the snapshot project_<X>_to_runtime_view DELETE+INSERT runtime_mcp_<X> from the snapshot — invoked only by the chassis runtime-view refresh dispatcher Two new vectors (auth_profiles_, target_profiles_, query_rules_) hold the authoritative per-table rows; target_auth_map remains as the joined view consumed by the listener and is rebuilt whenever either profile vector changes. # Plugin descriptor + table registration genai_register_admin_tables now also calls services->register_runtime_view for runtime_mcp_auth_profiles, runtime_mcp_target_profiles, and runtime_mcp_query_rules, pointing each at a project_*_to_runtime_view trampoline. The runtime_mcp_<X> tables stay registered for their CREATE TABLE schema only — their rows are entirely owned by the per-SELECT projection callbacks from this point on. # Admin SQL surface LOAD MCP PROFILES TO RUNTIME / LOAD MCP QUERY RULES TO RUNTIME now go through install_<X>_from_admin and never touch runtime_mcp_<X>. SAVE MCP PROFILES TO MEMORY is added (was missing pre-ABI-3) and joins SAVE MCP QUERY RULES TO MEMORY in pulling from the module snapshot instead of the runtime view. Query_Tool_Handler::refresh_target_registry no longer re-reads runtime_mcp_<X> via SQLite on every refresh; it consumes the in-memory snapshot directly through GloMCPH->get_all_target_auth_contexts(). # Drop stale GENAI:-prefix tests test/tap/tests/genai_{async,embedding_rerank,live_validation}-t.cpp were left over from before Step 4.A removed the GENAI: query-prefix escape hatch — they would all have failed under ai-g1. Delete them and their groups.json / ai/README.md references; equivalent coverage now flows through the MCP /mcp/ai endpoint and the genai_module-t suite that already drives the plugin surface.	2 months ago
Rene Cannao	7f148940e3	feat(plugins/genai): Step 8b — wire LOAD/SAVE GENAI VARIABLES via plugin ... Mirror of the 4.F MCP variable wiring, applied to genai-* variables. Closes the gap that surfaced in the ai-g1 spot-check where genai_module-t saw "default value" / "CHECKSUM GENAI VARIABLES" empty because flush_genai_variables___* was stubbed in Step 5 and there was no plugin-side replacement. New helpers (plugins/genai/src/plugin_main.cpp + genai_plugin.h): genai_load_variables_from_admindb(ctx) SELECT variable_name, variable_value FROM main.global_variables WHERE variable_name LIKE 'genai-%' → GloGATH->set_variable(name + 6, value) for each row. Mirrors the pre-Step-5 flush_genai_variables___database_to_runtime. genai_save_variables_to_admindb(ctx) Walks GloGATH->get_variables_list() and REPLACEs the `genai-<name>` rows in main.global_variables. Mirrors the pre-Step-5 flush_genai_variables___runtime_to_database. Both helpers wrlock GloGATH for the duration of their walk, so callers don't have to coordinate. New plugin-registered admin SQL verbs (plugin_commands.cpp): LOAD GENAI VARIABLES TO RUNTIME aliases: FROM MEMORY, FROM MEM, TO RUN SAVE GENAI VARIABLES TO MEMORY aliases: TO MEM, FROM RUNTIME, FROM RUN End-to-end verified against the running ai-g1 infra: SET genai-llm_provider='openai'; LOAD GENAI VARIABLES TO RUNTIME; SAVE GENAI VARIABLES TO MEMORY; SELECT * FROM main.global_variables WHERE variable_name LIKE 'genai-%'; -> 32 rows (the full GenAI_Threads_Handler variable set, populated by the SAVE FROM RUNTIME walk, including the value just SET). This is a follow-up gap the unit tests don't catch (the pattern mirrors LOAD/SAVE MCP VARIABLES which 4.F already validated end-to-end); ai-g1 caught it because genai_module-t hits CHECKSUM GENAI VARIABLES which evaluates over the populated row set. The pre-Step-5 ProxySQL_Admin::flush_genai_variables___* functions remain stubbed in core (no callers reach them now — the chassis dispatcher routes LOAD/SAVE GENAI VARIABLES verbs to the plugin before admin_handler_command_load_or_save can fire). Step 7's dead- code cleanup pass can drop them entirely.	2 months ago
Rene Cannao	25e4388ae6	feat(plugins/genai): Step 7 follow-up — Phase B schema registration + ai-g1 plugin wiring ... Three small but related fixes that unblock running the ai-g1 integration suite against the post-Step-7 build. 1. Plugin: register MCP admin tables in Phase B (register_schemas), not Phase C (init). The chassis ABI guarantees that tables registered via `services->register_table` from a plugin's `register_schemas` callback are merged into the admin SQLite schema BEFORE admin bootstrap creates the DB. Registering at init() instead is too late: by then the DB is already created without the plugin-declared tables. The genai plugin was registering at init(). Symptom: at plugin start (Phase D) the call into `mcp_load_target_auth_map_from_admindb` errored out with "no such table: runtime_mcp_auth_profiles" because the table never got created. plugins/genai/src/plugin_main.cpp: - New `genai_register_schemas` callback that invokes `genai_register_admin_tables(services)`. - Plugin descriptor gains the `register_schemas` field (mirrors plugins/mysqlx pattern; Step 0's PROXYSQL_PLUGIN_ABI_VERSION = 2 supports it). - Removed the redundant call in genai_init(). 2. Core: drop PROXYSQLGENAI gates that Step 7 left around the SET mcp-X / SET genai-X validation paths. Step 7 stopped passing -DPROXYSQLGENAI to core compiles, so the `#ifdef PROXYSQLGENAI` blocks I'd left in lib/Admin_Handler.cpp for the loose-prefix `mcp-*` / `genai-*` validators (Step 4.F / Step 5) silently expanded to nothing. Symptom: `SET mcp-port=6071` admin SQL got "Unknown global variable". lib/Admin_Handler.cpp: - Loose mcp-* / genai-* prefix checks in is_valid_global_variable are now unconditional (no #ifdef). - The MCP VARIABLES disk-side handler (LOAD MCP VARIABLES FROM DISK / SAVE MCP VARIABLES TO DISK) is also un-gated; it's pure SQL DML against global_variables LIKE 'mcp-%' with no plugin runtime dependency. The plugin-side runtime push remains the LOAD MCP VARIABLES TO RUNTIME verb registered in plugin_commands.cpp. 3. Test infra: bind-mount the genai plugin .so + per-group cnf for ai-g1, mirroring the mysqlx-soak / mysqlx-e2e pattern. Without these, the ai-g1 setup hook fails at SET mcp-port because the proxysql container's cnf doesn't list `plugins=("...")` and the .so isn't visible inside the container. test/infra/control/start-proxysql-isolated.bash: - PROXYSQL_LOAD_GENAI_PLUGIN=1 (set by the group's env.sh) now triggers a `-v ${WORKSPACE}/plugins/genai/...so:/usr/lib/proxysql/...so:ro` bind into the proxysql container. test/tap/groups/ai/proxysql-ci.cnf (new): - Identical to test/infra/control/proxysql-ci.cnf except for the trailing `plugins=("/usr/lib/proxysql/ProxySQL_GenAI_Plugin.so")` line that asks the chassis to load the genai plugin at Phase A. test/tap/groups/ai-g1/env.sh: - PROXYSQL_LOAD_GENAI_PLUGIN=1 - PROXYSQL_CONFIG_OVERRIDE=…/test/tap/groups/ai/proxysql-ci.cnf After this commit the ai post-infras hook (`MCP configuration completed` … `MySQL test data seeding completed` … `PostgreSQL test data seeding completed` … `AI post-infras hook completed`) runs to completion against the post-carve-out proxysql + genai plugin .so.	2 months ago
Rene Cannao	62ed0660a7	feat(plugins/genai): Step 7 — drop PROXYSQLGENAI from core build ... Last step of the GenAI plugin carve-out. The PROXYSQLGENAI macro is no longer compiled into core: lib/libproxysql.a and the proxysql binary build identically with or without `PROXYSQLGENAI=1`. The user-facing build flag stays — it now means "also build the genai plugin alongside proxysql" rather than "compile GenAI into the binary". What this commit does plugins/genai/Makefile Hardcodes -DPROXYSQLGENAI in the plugin's CXXFLAGS — the plugin IS the GenAI surface, so the macro is always-on inside its own compilation unit. Lets the existing `#ifdef PROXYSQLGENAI` guards inside plugin source files keep working without a massive guard-stripping pass (deferred follow-up cleanup). lib/Makefile, src/Makefile Drop the `PSQLGA := -DPROXYSQLGENAI` setter — core no longer sees the macro on the compile line. PSQLGA stays as an empty variable so the existing `$(PSQLGA)` cascades in CXXFLAGS references stay structurally identical (cleanup deferred). Also: lib/Makefile no longer compiles MCP_Tool_Handler.oo (the base class moved to the plugin in this commit too — see below). Top-level Makefile Plugin-build rules switched from `$(if $(filter 1,$(PROXYSQL40))` to `$(if $(filter 1,$(PROXYSQLGENAI))` — `PROXYSQL40=1` enables the chassis but not necessarily the genai plugin. Skip-message text updated to match. `PROXYSQLGENAI=1` still implies `PROXYSQL40=1` per the existing tier hierarchy, so the user invocation stays unchanged. MCP_Tool_Handler base class Moved to plugins/genai/ (`include/MCP_Tool_Handler.h` -> plugin include, `lib/MCP_Tool_Handler.cpp` -> plugin tool_handlers). Was the last GenAI-tier .cpp in core — only existed there because AI/RAG_Tool_Handler still inherited from it before Step 5. Now plugin-internal. lib/proxy_sqlite3_symbols.cpp Hook for sqlite_vec_init unconditional now (sqlite-vec extension object always linked into proxysql, since the genai plugin relies on it via this hook at dlopen time). lib/MySQL_Thread.cpp / lib/PgSQL_Thread.cpp `variables.ffto_enabled` default no longer flips to true when PROXYSQLGENAI is set (the gate was: PROXYSQLFFTO + PROXYSQLGENAI -> true; else false). With PROXYSQLGENAI removed, the default is false everywhere; operators opt in via admin SQL. CLAUDE.md Feature-tier table now lists Plugin Chassis (PROXYSQL40) and AI/MCP (PROXYSQLGENAI) separately, with a note explaining what PROXYSQLGENAI=1 means post-carve-out. Conditional Components section calls out that GenAI/MCP/RAG/LLM lives in plugins/genai/. What this commit does NOT do (deferred) - Stripping the `#ifdef PROXYSQLGENAI` guards from the plugin's own source files (LLM_Bridge.cpp, GenAI_Thread.cpp, etc. all have `#ifdef PROXYSQLGENAI ... #endif` brackets). The plugin's Makefile force-defines the macro so they keep expanding. A guard-stripping pass would be ~50 file edits with zero behaviour change. - Stripping the dead `#ifdef PROXYSQLGENAI` blocks from core (lib/Admin_Handler.cpp etc. still have FIXME-stubbed blocks inside `#ifdef PROXYSQLGENAI` from the 4.C / 5 surgery). With the macro undefined in core's compile line, those blocks silently expand to nothing — they're dead code already. Cleanup follow-up. - Removing the now-no-op stubs themselves (init_mcp_variables, flush_mcp_variables___*, load_mcp_server, stats___mcp_*, etc.). They linger as dead wrappers around early-returns; their declarations in proxysql_admin.h linger too. Cleanup follow-up. All three plugin unit tests pass: genai_plugin_load_unit-t : 25/25 genai_plugin_anomaly_unit-t : 6/6 genai_plugin_backend_client : 27/27 Carve-out scoreboard, FINAL: Step 0-2 : infrastructure (chassis cherry-pick, ABI extensions) Step 3 : ~1.1 K LOC Anomaly_Detector Step 4.A : ~1 K LOC GENAI: / LLM: query-prefix removal Step 4.B : ~0.5 K LOC backend_client helper Step 4.C : ~13 K LOC MCP subsystem + 8 tool handlers + MySQL_FTS Step 4.F : n/a runtime-reconfig admin SQL via plugin commands Step 4.G : n/a MCP admin/config/stats tables via plugin ABI Step 5 : ~7 K LOC GenAI/LLM/AI/RAG Step 6 : ~5 K LOC AI_Vector_Storage / Discovery_Schema / MySQL_Catalog / Static_Harvester pair Step 7 : n/a PROXYSQLGENAI macro dropped from core build Net core surface change: ~28 K LOC moved out of lib/ + include/ into plugins/genai/. src/proxysql binary shrunk from 197 MB (pre-carve-out, all PROXYSQLGENAI compiled in) to 152 MB; plugin .so grew to ~56 MB. Net process RSS roughly neutral.	2 months ago
Rene Cannao	4552236700	feat(plugins/genai): Step 6 — move RAG/Vector/discovery surface into the plugin ... Final file-move step in the carve-out: relocates the leaf modules that didn't depend on (and weren't depended on by) anything still in core. After this commit, the only PROXYSQLGENAI-guarded code remaining in libproxysql.a is `MCP_Tool_Handler` — the base class the plugin's AI_Tool_Handler / RAG_Tool_Handler inherit from — and that lives in core only because removing it would force pulling those derived plugin classes into core just for the inheritance edge. Step 7 deletes the PROXYSQLGENAI macro entirely. Files moved (git mv preserves history) Headers (include/ -> plugins/genai/include/) AI_Vector_Storage.h MySQL_Catalog.h Discovery_Schema.h Static_Harvester.h PgSQL_Static_Harvester.h Sources (lib/ -> plugins/genai/src/) AI_Vector_Storage.cpp MySQL_Catalog.cpp Discovery_Schema.cpp Static_Harvester.cpp PgSQL_Static_Harvester.cpp Core changes were minimal because these are leaf modules: include/proxysql.h: drop the AI_Vector_Storage.h include. include/cpp.h: drop AI_Vector_Storage / Discovery_Schema / MySQL_Catalog / Static_Harvester / PgSQL_Static_Harvester includes (replaced with a single carve-out history comment that Step 7 will delete entirely). lib/ProxySQL_Admin.cpp: drop the Discovery_Schema.h include — its only live use was via the now-#if-0'd save/load_mcp_query_rules blocks 4.C left behind for follow-up reference. lib/Makefile: drop AI_Vector_Storage.oo, MySQL_Catalog.oo, Discovery_Schema.oo, Static_Harvester.oo, PgSQL_Static_Harvester.oo from the GenAI _OBJ_CXX list. Test fixture probe update test/tap/tests/unit/Makefile: PROXYSQLGENAI auto-detect probe switched from `AI_Vector_Storage` (now plugin-only) to `MCP_Tool_Handler` (the last GenAI-tier symbol still in core). Step 7 deletes the auto-detect entirely. All three plugin unit tests pass: genai_plugin_load_unit-t : 25/25 genai_plugin_anomaly_unit-t : 6/6 genai_plugin_backend_client : 27/27 Binary size deltas: src/proxysql 159 MB -> 152 MB (-7 MB) plugins/genai/...Plugin.so 47 MB -> 55 MB (+8 MB) Net process RSS roughly neutral, as the design predicted. Carve-out scoreboard: Step 0-2 : infrastructure Step 3 : ~1.1 K LOC (Anomaly_Detector) Step 4.B : ~0.5 K LOC (backend_client) Step 4.C : ~13 K LOC (MCP subsystem + tool handlers + MySQL_FTS) Step 4.F : runtime-reconfig admin SQL routed through plugin commands Step 4.G : MCP admin / config / stats tables registered via plugin ABI Step 5 : ~7 K LOC (GenAI/LLM/AI/RAG) Step 6 : ~5 K LOC (AI_Vector_Storage + Discovery_Schema + MySQL_Catalog + Static_Harvester pair) Pending: Step 7 — delete the PROXYSQLGENAI macro from build flags and any remaining `#ifdef PROXYSQLGENAI` blocks.	2 months ago
Rene Cannao	28cf4ceab8	feat(plugins/genai): Step 5 — move GenAI/LLM/AI surface into the plugin ... Largest carve-out commit since 4.C: moves the GenAI / LLM / AI / RAG surface (about 7K LOC) out of core into plugins/genai/. After this commit, libproxysql.a no longer publishes any of GenAI_Threads_Handler, LLM_Bridge, LLM_Clients, AI_Features_Manager, AI_Tool_Handler, RAG_Tool_Handler, or the legacy GloGATH/GloAI globals. Files moved (git mv preserves history) Headers (include/ -> plugins/genai/include/) GenAI_Thread.h LLM_Bridge.h AI_Features_Manager.h AI_Tool_Handler.h RAG_Tool_Handler.h (promoted from Step 6 — bidirectional dep with AI_Features_Manager forced the early move, same way MySQL_FTS.h got promoted in 4.C) Sources (lib/ -> plugins/genai/src/ or …/tool_handlers/) GenAI_Thread.cpp LLM_Bridge.cpp LLM_Clients.cpp (.cpp without a header; class decls live in LLM_Bridge.h) AI_Features_Manager.cpp AI_Tool_Handler.cpp -> tool_handlers/ RAG_Tool_Handler.cpp -> tool_handlers/ Plugin lifecycle takeover plugins/genai/src/plugin_main.cpp: - Defines plugin-local GloGATH and GloAI alongside the existing GloMCPH (all three are .so-internal aliases for the contexts genai_init constructs). - genai_init now also constructs GenAI_Threads_Handler (calls ->init()) and AI_Features_Manager (calls ->init()), mirroring the pre-Step-5 ProxySQL_Main_init_GenAI_module sequence. - genai_stop tears them down in reverse order. plugins/genai/src/ProxySQL_MCP_Server.cpp: - AI / RAG endpoint registration re-enabled (was #if 0'd in 4.C because the constructors weren't reachable from the plugin then). /mcp/ai and /mcp/rag are once again live when the operator has GloAI configured. plugins/genai/src/Anomaly_Detector.cpp + plugin_main.cpp: - get_query_embedding() now calls through a function-pointer hook (genai_anomaly_embed_fn) that plugin_main installs in genai_init and clears in genai_stop. This indirection lets the test fixture (which compiles Anomaly_Detector.cpp directly into a test binary, without GenAI_Thread.cpp) link cleanly — the pointer stays null, the detector short-circuits to empty embedding, and the existing nullptr-guard in check_embedding_similarity does the rest. Core neutralization src/main.cpp: - Drops the GloGATH / GloAI globals, ProxySQL_Main_init_GenAI_module definition + call, the shutdown branch, and the `if (GloGATH) init_genai_variables()` call. lib/Admin_Bootstrap.cpp / Admin_FlushVariables.cpp / Admin_Handler.cpp / ProxySQL_Admin.cpp / MySQL_Session.cpp / ProxySQL_Admin_Stats.cpp: - Drop the corresponding `#include` and `extern` declarations. - flush_genai_variables___* stubbed (mirrors the 4.C MCP pattern; original bodies preserved in `#if 0` for the upcoming LOAD/SAVE GENAI verbs port). lib/Admin_Handler.cpp: - Loose-prefix accept for `genai-*` in is_valid_global_variable (mirrors the 4.F mcp-* fix), so SET genai-X=Y still reaches main.global_variables; runtime push happens via the plugin's LOAD GENAI VARIABLES TO RUNTIME (registered in a follow-up). include/cpp.h, include/proxysql.h: - Removed the AI_Features_Manager.h / LLM_Bridge.h / GenAI_Thread.h / AI_Tool_Handler.h / RAG_Tool_Handler.h includes that no longer resolve from core's -I path. test/tap/test_helpers/test_globals.cpp: - Drops the matching `#include`s and stub globals. Test fixture test/tap/tests/unit/Makefile: - PROXYSQLGENAI auto-detect probe switched from `GenAI_Thread` (now plugin-only) to `AI_Vector_Storage` (still in core for Step 6). After Step 6 the probe will switch again, and Step 7 deletes the auto-detect entirely. All three plugin unit tests pass: - genai_plugin_load_unit-t : 25/25 - genai_plugin_anomaly_unit-t : 6/6 - genai_plugin_backend_client : 27/27 Carve-out scoreboard: Step 0-2 : infrastructure Step 3 : ~1.1 K LOC (Anomaly_Detector) Step 4.B : ~0.5 K LOC (backend_client) Step 4.C : ~13 K LOC (MCP subsystem) Step 4.F : runtime-reconfig admin SQL routed through plugin commands Step 4.G : MCP admin / config / stats tables registered via plugin ABI Step 5 : ~7 K LOC (GenAI/LLM/AI/RAG) Pending: Step 6 (AI_Vector_Storage, MySQL_Catalog, Discovery_Schema, Static_Harvester, PgSQL_Static_Harvester) and Step 7 (delete the PROXYSQLGENAI macro entirely).	2 months ago
Rene Cannao	9078fd3150	feat(plugins/genai): Step 4.G — register MCP admin tables via plugin ABI ... Moves MCP admin / config / stats SQLite table registration from lib/Admin_Bootstrap.cpp into the plugin via the chassis register_table service. Pure refactor: the same DDL strings produce the same SQLite tables in the same SQLite DBs; only the registration path changes. New plugins/genai/src/plugin_tables.cpp registers 14 tables: admin DB (6) mcp_query_rules, runtime_mcp_query_rules mcp_auth_profiles, runtime_mcp_auth_profiles mcp_target_profiles, runtime_mcp_target_profiles config DB (3) mcp_query_rules, mcp_auth_profiles, mcp_target_profiles (no runtime_* in config; matches the pre-4.G layout) stats DB (5) stats_mcp_query_tools_counters, _reset stats_mcp_query_digest, _reset stats_mcp_query_rules The DDL strings still live in include/ProxySQL_Admin_Tables_Definitions.h for now — plugin_tables.cpp #includes the header. Migrating the macros into a plugin-private header is a follow-up gated on no remaining core caller of the MCP table macros (Step 7 territory). Core changes (lib/Admin_Bootstrap.cpp): - Deleted the 9 admin/config insert_into_tables_defs() entries inside `#ifdef PROXYSQLGENAI` for mcp_*. - Deleted the 5 stats insert_into_tables_defs() entries similarly. - Replaced both blocks with a one-line comment pointing at the new plugin owner. Plugin lifecycle: - genai_init() now calls genai_register_admin_tables() right after genai_register_admin_commands(). Both run during chassis Phase C (init), before admin module bootstrap so the chassis dispatcher merges plugin tables into the admin DB schema before any DML. Test extension (test/tap/tests/unit/genai_plugin_load_unit-t.cpp): - Updated the "skeleton registers no tables (yet)" assertions to expect 6 admin / 3 config / 5 stats — matching plugin_tables.cpp. - 25/25 tests passing. What this changes for operators (no-op): - Same SQL: SHOW TABLES still shows mcp_*, runtime_mcp_*, stats_mcp_*. - Same DDL: column lists, primary keys, defaults — all preserved. - Same admindb / configdb / statsdb attachment.	2 months ago
Rene Cannao	f90307ad75	feat(plugins/genai): Step 4.F — register LOAD/SAVE MCP QUERY RULES TO/FROM RUNTIME ... Adds the last LOAD/SAVE pair the pre-4.C admin SQL surface had: the MCP query rules cache held by Query_Tool_Handler's Discovery_Schema catalog. New helpers in plugins/genai/src/plugin_main.cpp: mcp_load_query_rules_to_runtime(ctx) SELECT * FROM main.mcp_query_rules WHERE active=1 ORDER BY rule_id → catalog->load_mcp_query_rules(rs). Mirrors the pre-4.C ProxySQL_Admin::load_mcp_query_rules_to_runtime. mcp_save_query_rules_from_runtime(ctx, runtime) Pulls catalog->get_mcp_query_rules() and REPLACEs the rows in main.mcp_query_rules (or runtime_mcp_query_rules when runtime=true). Mirrors the pre-4.C ProxySQL_Admin::save_mcp_query_rules_from_runtime. Both helpers fail cleanly (return false, log to stderr) when the MCP listener isn't running — there's no in-memory Query_Tool_Handler cache to read or write. Pre-4.C behaviour was to attempt a load_mcp_server() self-recovery in that case; we no longer have that shortcut from outside the listener lifecycle, so callers get a clear error instead of an opaque retry. New verbs registered in plugin_commands.cpp: LOAD MCP QUERY RULES TO RUNTIME aliases: FROM MEMORY, FROM MEM, TO RUN SAVE MCP QUERY RULES TO MEMORY aliases: TO MEM, FROM RUNTIME, FROM RUN Required new core-side includes in plugin_main.cpp for the Discovery_Schema and Query_Tool_Handler types these helpers traffic in. The .so already linked against them via the chassis --whole-archive trick on libproxysql.a; this is just a header dependency. Test extension (test/tap/tests/unit/genai_plugin_load_unit-t.cpp): - 6 new assertions for the QUERY RULES canonical + 5 aliases. - 25/25 tests passing. Carve-out scoreboard, runtime reconfig surface restored: LOAD MCP VARIABLES TO RUNTIME / FROM MEMORY / FROM MEM / TO RUN LOAD MCP PROFILES TO RUNTIME / FROM MEMORY / FROM MEM / TO RUN LOAD MCP QUERY RULES TO RUNTIME / FROM MEMORY / FROM MEM / TO RUN SAVE MCP VARIABLES TO MEMORY / TO MEM / FROM RUNTIME / FROM RUN SAVE MCP QUERY RULES TO MEMORY / TO MEM / FROM RUNTIME / FROM RUN Still pending: SET mcp-X=Y (chassis ABI gap, separate enhancement).	2 months ago
Rene Cannao	7a4f08416e	feat(plugins/genai): Step 4.F — register SAVE MCP VARIABLES TO MEMORY ... Adds the SAVE direction (runtime → main.global_variables) to the plugin's admin SQL surface, mirroring the LOAD direction wired in the previous commit. New helper plugins/genai/src/plugin_main.cpp::mcp_save_variables_to_admindb Walks ctx.mcp->get_variables_list(), reads each runtime value via ctx.mcp->get_variable, and REPLACEs the corresponding "mcp-<name>" rows in main.global_variables. Owns its own MCP_Threads_Handler wrlock so callers don't have to coordinate. Mirrors the pre-4.C flush_mcp_variables___runtime_to_database with runtime=false (no runtime_global_variables view write). New verb registered in plugins/genai/src/plugin_commands.cpp: SAVE MCP VARIABLES TO MEMORY (canonical) aliases: SAVE MCP VARIABLES TO MEM SAVE MCP VARIABLES FROM RUNTIME SAVE MCP VARIABLES FROM RUN Same alias convention as the LOAD verbs and as plugins/mysqlx. The on-disk write ("SAVE MCP VARIABLES TO DISK") stays in core's admin_handler_command_load_or_save — it's a pure SQL DML that copies main.* to disk.* and doesn't need plugin runtime state. Test extension (test/tap/tests/unit/genai_plugin_load_unit-t.cpp): - 4 new TAP assertions for the SAVE canonical + 3 aliases. - 19/19 tests passing. Carve-out scoreboard, runtime reconfig surface restored: LOAD MCP VARIABLES TO RUNTIME / FROM MEMORY / FROM MEM / TO RUN LOAD MCP PROFILES TO RUNTIME / FROM MEMORY / FROM MEM / TO RUN SAVE MCP VARIABLES TO MEMORY / TO MEM / FROM RUNTIME / FROM RUN Still pending: - SET mcp-X=Y (chassis ABI lacks per-variable namespace registration; would need a new ABI surface). - LOAD/SAVE MCP QUERY RULES TO/FROM RUNTIME (Query_Tool_Handler in plugin owns the runtime cache; helpers come in next commit).	2 months ago
Rene Cannao	ae13958995	feat(plugins/genai): Step 4.F (continued) — register LOAD MCP {VARIABLES,PROFILES} TO RUNTIME ... Adds plugins/genai/src/plugin_commands.cpp registering two admin SQL verbs through the chassis command registry, restoring runtime reconfiguration that 4.C disabled. Verbs registered (each with FROM MEMORY / FROM MEM / TO RUN aliases): LOAD MCP VARIABLES TO RUNTIME -> mcp_load_variables_from_admindb() -> mcp_start_listener_if_enabled() Re-pushes mcp-* values from main.global_variables into the running MCP_Threads_Handler, then re-evaluates the listener state (start if newly enabled). Replaces the FIXME-stubbed ProxySQL_Admin::flush_mcp_variables___database_to_runtime path. LOAD MCP PROFILES TO RUNTIME -> mcp_load_target_auth_map_from_admindb() Refreshes runtime_mcp_auth_profiles / runtime_mcp_target_profiles from main.* and rebuilds the in-memory target_auth_map. Replaces the FIXME-stubbed init_mcp_variables auth-map block. The chassis dispatcher in lib/Admin_Handler.cpp:3987 runs BEFORE admin_handler_command_load_or_save, so any admin SQL matching a registered plugin verb routes to the plugin and the (now no-op) hardcoded core handlers never fire. No core changes needed. What's still stubbed (deferred to follow-up commits): - SAVE MCP VARIABLES TO MEMORY / FROM RUNTIME (needs runtime -> admindb push in plugin) - SAVE MCP PROFILES TO DISK (writes runtime view back to disk.*) - LOAD/SAVE MCP QUERY RULES (Query_Tool_Handler in plugin owns them, but the runtime->admindb push isn't wired through plugin yet) - SET mcp-X=Y (chassis ABI has no register_variable_namespace; would need either a new ABI surface or per-variable register_command — defer to a chassis enhancement) Refactor: moved mcp_load_variables_from_admindb, mcp_load_target_auth_map_from_admindb, mcp_start_listener_if_enabled out of the anonymous namespace in plugin_main.cpp so plugin_commands.cpp can call them. Forward declarations added to genai_plugin.h. Test extension (test/tap/tests/unit/genai_plugin_load_unit-t.cpp): - 7 new TAP assertions verifying canonical SQL and aliases resolve through the plugin manager's resolve_alias_to_canonical, which is the same path admin SQL dispatch uses. - 15/15 tests passing.	2 months ago