diff --git a/sql/log_event.cc b/sql/log_event.cc index bfcac10e138..2a54d80178c 100644 --- a/sql/log_event.cc +++ b/sql/log_event.cc @@ -1244,7 +1244,7 @@ Log_event* Log_event::read_log_event(const char* buf, uint event_len, break; #ifdef HAVE_REPLICATION case SLAVE_EVENT: /* can never happen (unused event) */ - ev = new Slave_log_event(buf, event_len); + ev = new Slave_log_event(buf, event_len, description_event); break; #endif /* HAVE_REPLICATION */ case CREATE_FILE_EVENT: @@ -1332,8 +1332,10 @@ Log_event* Log_event::read_log_event(const char* buf, uint event_len, (because constructor is "void") ; so instead we leave the pointer we wanted to allocate (e.g. 'query') to 0 and we test it in is_valid(). Same for Format_description_log_event, member 'post_header_len'. + + SLAVE_EVENT is never used, so it should not be read ever. */ - if (!ev || !ev->is_valid()) + if (!ev || !ev->is_valid() || (event_type == SLAVE_EVENT)) { DBUG_PRINT("error",("Found invalid event in binary log")); @@ -6117,8 +6119,12 @@ void Slave_log_event::init_from_mem_pool(int data_size) /** This code is not used, so has not been updated to be format-tolerant. */ -Slave_log_event::Slave_log_event(const char* buf, uint event_len) - :Log_event(buf,0) /*unused event*/ ,mem_pool(0),master_host(0) +/* We are using description_event so that slave does not crash on Log_event + constructor */ +Slave_log_event::Slave_log_event(const char* buf, + uint event_len, + const Format_description_log_event* description_event) + :Log_event(buf,description_event),mem_pool(0),master_host(0) { if (event_len < LOG_EVENT_HEADER_LEN) return; diff --git a/sql/log_event.h b/sql/log_event.h index 5d7250d8ebd..f58e1064d32 100644 --- a/sql/log_event.h +++ b/sql/log_event.h @@ -1846,7 +1846,9 @@ public: void print(FILE* file, PRINT_EVENT_INFO* print_event_info); #endif - Slave_log_event(const char* buf, uint event_len); + Slave_log_event(const char* buf, + uint event_len, + const Format_description_log_event *description_event); ~Slave_log_event(); int get_data_size(); bool is_valid() const { return master_host != 0; } diff --git a/sql/rpl_rli.h b/sql/rpl_rli.h index bf8381933c0..09b58d4f3a0 100644 --- a/sql/rpl_rli.h +++ b/sql/rpl_rli.h @@ -97,6 +97,16 @@ public: */ MYSQL_BIN_LOG relay_log; LOG_INFO linfo; + + /* + cur_log + Pointer that either points at relay_log.get_log_file() or + &rli->cache_buf, depending on whether the log is hot or there was + the need to open a cold relay_log. + + cache_buf + IO_CACHE used when opening cold relay logs. + */ IO_CACHE cache_buf,*cur_log; /* diff --git a/sql/slave.cc b/sql/slave.cc index 3e77a5e7516..677199bfc0e 100644 --- a/sql/slave.cc +++ b/sql/slave.cc @@ -4726,12 +4726,66 @@ static Log_event* next_event(Relay_log_info* rli) DBUG_ASSERT(rli->cur_log_fd == -1); /* - Read pointer has to be at the start since we are the only - reader. - We must keep the LOCK_log to read the 4 first bytes, as this is a hot - log (same as when we call read_log_event() above: for a hot log we - take the mutex). + When the SQL thread is [stopped and] (re)started the + following may happen: + + 1. Log was hot at stop time and remains hot at restart + + SQL thread reads again from hot_log (SQL thread was + reading from the active log when it was stopped and the + very same log is still active on SQL thread restart). + + In this case, my_b_seek is performed on cur_log, while + cur_log points to relay_log.get_log_file(); + + 2. Log was hot at stop time but got cold before restart + + The log was hot when SQL thread stopped, but it is not + anymore when the SQL thread restarts. + + In this case, the SQL thread reopens the log, using + cache_buf, ie, cur_log points to &cache_buf, and thence + its coordinates are reset. + + 3. Log was already cold at stop time + + The log was not hot when the SQL thread stopped, and, of + course, it will not be hot when it restarts. + + In this case, the SQL thread opens the cold log again, + using cache_buf, ie, cur_log points to &cache_buf, and + thence its coordinates are reset. + + 4. Log was hot at stop time, DBA changes to previous cold + log and restarts SQL thread + + The log was hot when the SQL thread was stopped, but the + user changed the coordinates of the SQL thread to + restart from a previous cold log. + + In this case, at start time, cur_log points to a cold + log, opened using &cache_buf as cache, and coordinates + are reset. However, as it moves on to the next logs, it + will eventually reach the hot log. If the hot log is the + same at the time the SQL thread was stopped, then + coordinates were not reset - the cur_log will point to + relay_log.get_log_file(), and not a freshly opened + IO_CACHE through cache_buf. For this reason we need to + deploy a my_b_seek before calling check_binlog_magic at + this point of the code (see: BUG#55263 for more + details). + + NOTES: + - We must keep the LOCK_log to read the 4 first bytes, as + this is a hot log (same as when we call read_log_event() + above: for a hot log we take the mutex). + + - Because of scenario #4 above, we need to have a + my_b_seek here. Otherwise, we might hit the assertion + inside check_binlog_magic. */ + + my_b_seek(cur_log, (my_off_t) 0); if (check_binlog_magic(cur_log,&errmsg)) { if (!hot_log)