This keeps the behavior the same but allows the heartbeat to be used as a crude timer
without including everything the housekeeper.h header includes.
Moved the protocol level updates on the DCBs last_read value into dcb_read.
This keeps the implementation nice and compact.
mysql_backend.c:gw_change_user: instead of setting flags directly to replybuffer, set server command to backend's protocol object and reply the message directly to backend where flags are set and it is replied back to client.
RWSplit router handles query processing errors that happened in router by sending parse error reply to client. routeQuery fails only when backend has failed.
http://bugs.skysql.com/show_bug.cgi?id=665http://bugs.skysql.com/show_bug.cgi?id=664http://bugs.skysql.com/show_bug.cgi?id=649
dcb.c:dcb_final_free: (665):set dcb->session->client pointer to NULL so that it won't be read anymore and other threads won't try to close it.
dcb_final_free:(664):don't free dcb->data, it is either freed in session_alloc if session creation fails or in session_free only.
session.c:if session creation fails, free dcb->data and remove links between client DCB and session.
mysql_backend.c:(665):gw_backend_close:check that session->client isn't NULL and that client DCB's state is still polling before calling dcb_close for it.
mysql_client.c:gw_mysql_do_authentication:if anything fails, and session_alloc won't be called, free dcb->data.
mysql_common.c:gw_send_authentication_to_backend:if session is already closing then return with error.
readwritesplit.h:Removed invalid macros which assumed that ROUTER_CLIENT_SES->rses_backend_ref always pointed to valid and used backend reference and thus included potential risk of NULL-pointer refernce.
mysql_backend.c and mysql_client.c:avoid executing CHK_PROTOCOL(p) after original DCB has been released the memory.
readwritesplit.c:Replaced RSES_CLEINT_DCB macro with a function which returns client DCB for a given router client session.
dcb.c:dcb_clone: set same close function than in the original DCB so that routing session will become closed when client DCB is closed.
session.h:introduce a new state which indicates that SESSION structure can be freed next, SESSION_STATE_TO_BE_FREED. State is needed to separate the state before and after memory free. Memory is freed in different situation depending on whether the session is parent or child session. Child sessions are freed in their parent.
Also introduce a new member in SESSION struct, ses_is_child, to indicate whether SESSION has a parent session which is responsible for releasing the memory of it.
session.c:session_unlink_dcb:when last session reference is removed set SESSION->state=SESSION_STATE_TO_BE_FREED
session.c:session_free:only free the memory if session is not child session. If it is child, session is left to be freed by the parent (in tee filter's freeSession, for example).
tee.c:newSession:mark branch session to child.
tee.c:freeSession:if child session memory is ready to be freed, free it.
mysql_mon.c:monitorMain:when backend server's status changes so that it is not running anymore or doesn't have any of the states - master, slave, ndb, joined - call for each DCB the callback-function DCB_REASON_NOT_RESPONDING if specified. Earlier callbacks were called if there was any change in the state of the server being monitored.
mysql_mon.c: Added back server state logging to Debug build.
query_classifier.cc: removed some extra debug code, cleaned up a bit function documentation.
mysql_client_server_protocol.h, mysql_backend.c, mysql_common.c: changed some variables to signed ones to enable checking of calculations in the code.
skygw_utils.cc: removed erroneous debug assertion.
void session_enable_log(SESSION* ses, logfile_id_t id)
and
void session_disable_log(SESSION* ses, logfile_id_t id)
Which switch specific log type on/off if the log type in question is not generally enabled.
Each thread carries a thread-specific struct log_info_t which includes members for current session id and bitfield for enabled log types for the current session. That information is checked before actual log write functions are called.
Each file where session-specific logging is used, must include the following exports:
/** Defined in log_manager.cc */
extern int lm_enabled_logfiles_bitmask;
extern size_t log_ses_count[];
extern __thread log_info_t tls_log_info;
if (func.auth ==)gw_change_user->gw_send_change_user_to_backend is called before backend has its scramble, auth packet is set to backend's delauqueue instead of writing it to backend. When backend_write_delayqueue is called COM_CHANGE_USER packets are rewritten with backend's current data.
mysql_backend.c, mysql_client.c free error message GWBUF after calling handleError
readconnroute.c:handleError send error message to client before returning.
readwritesplit.c:handleError don't free error message buffer anymore since the caller of handleError frees it.
gw_change_user returned error code in cases where authentication failed. That is not right behavior in that context. gw_change_user succeeds to route query and should return succees to router's routeQuery. Possible error is handled by creating error message and by feeding it to client DCB's event queue.
Added function poll.c:poll_add_epollin_event_to_dcb which adds given event and related GWBUF to given DCB. The DCB is then added to event queue list so that its even becomes executed as soon as any of the threads start to process events next time.
The problem solved is case where COM_CHANGE_USER failed due to authentication failure. An error should be sent to the client but it was sent once per failed backend authentication instead of just once.
Thread which failed in authentication bypassed router and thus replies weren't filtered out. Solution adds an event and a GWBUF including error message to DCB and the DCB is then added to event queue. Thus the error message appeares like it was an incoming event and becomes processed properly.
An addition was made to mysql_backend.c:gw_read_backend_event so that self-generated event can be processed. Even if socket read doesn't manage to read anything, function proceeds if there is enough data in DCB's readqueue.
Fix ti bug #602, http://bugs.mariadb.com/show_bug.cgi?id=602
Added function sescmd_response_complete which checks if the read data completes session command response packet(s). Function is called only if there's a sessoin command being executed.
