MDEV-14756 - Remove trx_sys_t::rw_trx_list

Reduce divergence between trx_sys_t::rw_trx_hash and trx_sys_t::rw_trx_list
by not adding recovered COMMITTED transactions to trx_sys_t::rw_trx_list.

Such transactions are discarded immediately without creating trx object.

This also required to split rollback and cleanup phases of recovery. To
reflect these updates the following renames happened:
trx_rollback_or_clean_all_recovered() -> trx_rollback_all_recovered()
trx_rollback_or_clean_is_active -> trx_rollback_is_active
trx_rollback_or_clean_recovered() -> trx_rollback_recovered()
trx_cleanup_at_db_startup() -> trx_cleanup_recovered()

Also removed a hack from lock_trx_release_locks(). Instead let recovery
rollback thread to skip committed XA transactions.

storage/innobase/btr/btr0cur.cc

@@ -7782,7 +7782,7 @@ btr_free_externally_stored_field(
 					 MLOG_4BYTES, &mtr);
 			/* Zero out the BLOB length.  If the server
 			crashes during the execution of this function,
-			trx_rollback_or_clean_all_recovered() could
+			trx_rollback_all_recovered() could
 			dereference the half-deleted BLOB, fetching a
 			wrong prefix for the BLOB. */
 			mlog_write_ulint(field_ref + BTR_EXTERN_LEN + 4,

storage/innobase/include/trx0roll.h

@@ -33,7 +33,7 @@ Created 3/26/1996 Heikki Tuuri
 #include "mtr0mtr.h"
 #include "trx0sys.h"
 
-extern bool		trx_rollback_or_clean_is_active;
+extern bool		trx_rollback_is_active;
 extern const trx_t*	trx_roll_crash_recv_trx;
 
 /*******************************************************************//**
@@ -71,12 +71,11 @@ trx_roll_must_shutdown();
 Rollback or clean up any incomplete transactions which were
 encountered in crash recovery.  If the transaction already was
 committed, then we clean up a possible insert undo log. If the
-transaction was not yet committed, then we roll it back. */
+transaction was not yet committed, then we roll it back.
+@param all true=roll back all recovered active transactions;
+false=roll back any incomplete dictionary transaction */
 void
-trx_rollback_or_clean_recovered(
-/*============================*/
-	ibool	all);	/*!< in: FALSE=roll back dictionary transactions;
-			TRUE=roll back all non-PREPARED transactions */
+trx_rollback_recovered(bool all);
 /*******************************************************************//**
 Rollback or clean up any incomplete transactions which were
 encountered in crash recovery.  If the transaction already was
@@ -86,11 +85,7 @@ Note: this is done in a background thread.
 @return a dummy parameter */
 extern "C"
 os_thread_ret_t
-DECLARE_THREAD(trx_rollback_or_clean_all_recovered)(
-/*================================================*/
-	void*	arg MY_ATTRIBUTE((unused)));
-			/*!< in: a dummy parameter required by
-			os_thread_create */
+DECLARE_THREAD(trx_rollback_all_recovered)(void*);
 /*********************************************************************//**
 Creates a rollback command node struct.
 @return own: rollback node struct */

storage/innobase/include/trx0trx.h

@@ -246,14 +246,6 @@ trx_commit_low(
 	trx_t*	trx,	/*!< in/out: transaction */
 	mtr_t*	mtr);	/*!< in/out: mini-transaction (will be committed),
 			or NULL if trx made no modifications */
-/****************************************************************//**
-Cleans up a transaction at database startup. The cleanup is needed if
-the transaction already got to the middle of a commit when the database
-crashed, and we cannot roll it back. */
-void
-trx_cleanup_at_db_startup(
-/*======================*/
-	trx_t*	trx);	/*!< in: transaction */
 /**********************************************************************//**
 Does the transaction commit for MySQL.
 @return DB_SUCCESS or error number */

storage/innobase/lock/lock0lock.cc

@@ -7630,8 +7630,6 @@ lock_trx_release_locks(
 		lock_mutex_enter();
 	}
 
-	trx_mutex_enter(trx);
-
 	/* The following assignment makes the transaction committed in memory
 	and makes its changes to data visible to other transactions.
 	NOTE that there is a small discrepancy from the strict formal
@@ -7647,7 +7645,9 @@ lock_trx_release_locks(
 	committed. */
 
 	/*--------------------------------------*/
+	trx_mutex_enter(trx);
 	trx->state = TRX_STATE_COMMITTED_IN_MEMORY;
+	trx_mutex_exit(trx);
 	/*--------------------------------------*/
 
 	if (trx->is_referenced()) {
@@ -7658,41 +7658,18 @@ lock_trx_release_locks(
 
 		while (trx->is_referenced()) {
 
-			trx_mutex_exit(trx);
-
 			DEBUG_SYNC_C("waiting_trx_is_not_referenced");
 
 			/** Doing an implicit to explicit conversion
 			should not be expensive. */
 			ut_delay(ut_rnd_interval(0, srv_spin_wait_delay));
-
-			trx_mutex_enter(trx);
 		}
 
-		trx_mutex_exit(trx);
-
 		lock_mutex_enter();
-
-		trx_mutex_enter(trx);
 	}
 
 	ut_ad(!trx->is_referenced());
 
-	/* If the background thread trx_rollback_or_clean_recovered()
-	is still active then there is a chance that the rollback
-	thread may see this trx as COMMITTED_IN_MEMORY and goes ahead
-	to clean it up calling trx_cleanup_at_db_startup(). This can
-	happen in the case we are committing a trx here that is left
-	in PREPARED state during the crash. Note that commit of the
-	rollback of a PREPARED trx happens in the recovery thread
-	while the rollback of other transactions happen in the
-	background thread. To avoid this race we unconditionally unset
-	the is_recovered flag. */
-
-	trx->is_recovered = false;
-
-	trx_mutex_exit(trx);
-
 	if (release_lock) {
 
 		lock_release(trx);

storage/innobase/log/log0log.cc

@@ -1949,7 +1949,7 @@ logs_empty_and_mark_files_at_shutdown(void)
 		goto wait_suspend_loop;
 	} else if (btr_defragment_thread_active) {
 		thread_name = "btr_defragment_thread";
-	} else if (srv_fast_shutdown != 2 && trx_rollback_or_clean_is_active) {
+	} else if (srv_fast_shutdown != 2 && trx_rollback_is_active) {
 		thread_name = "rollback of recovered transactions";
 	} else {
 		thread_name = NULL;

storage/innobase/log/log0recv.cc

@@ -3426,8 +3426,8 @@ recv_recovery_rollback_active(void)
 		/* Rollback the uncommitted transactions which have no user
 		session */
 
-		trx_rollback_or_clean_is_active = true;
-		os_thread_create(trx_rollback_or_clean_all_recovered, 0, 0);
+		trx_rollback_is_active = true;
+		os_thread_create(trx_rollback_all_recovered, 0, 0);
 	}
 }
 

storage/innobase/row/row0ext.cc

@@ -70,7 +70,7 @@ row_ext_cache_fill(
 		} else {
 			/* Fetch at most ext->max_len of the column.
 			The column should be non-empty.  However,
-			trx_rollback_or_clean_all_recovered() may try to
+			trx_rollback_all_recovered() may try to
 			access a half-deleted BLOB if the server previously
 			crashed during the execution of
 			btr_free_externally_stored_field(). */

storage/innobase/srv/srv0start.cc

@@ -1649,7 +1649,7 @@ innobase_start_or_create_for_mysql()
 			    + 1 /* dict_stats_thread */
 			    + 1 /* fts_optimize_thread */
 			    + 1 /* recv_writer_thread */
-			    + 1 /* trx_rollback_or_clean_all_recovered */
+			    + 1 /* trx_rollback_all_recovered */
 			    + 128 /* added as margin, for use of
 				  InnoDB Memcached etc. */
 			    + max_connections
@@ -2471,7 +2471,7 @@ innobase_start_or_create_for_mysql()
 		The data dictionary latch should guarantee that there is at
 		most one data dictionary transaction active at a time. */
 		if (srv_force_recovery < SRV_FORCE_NO_TRX_UNDO) {
-			trx_rollback_or_clean_recovered(FALSE);
+			trx_rollback_recovered(false);
 		}
 
 		/* Fix-up truncate of tables in the system tablespace

storage/innobase/trx/trx0purge.cc

@@ -300,7 +300,7 @@ trx_purge_add_undo_to_history(const trx_t* trx, trx_undo_t*& undo, mtr_t* mtr)
 	continue to execute user transactions. */
 	ut_ad(srv_undo_sources
 	      || ((srv_startup_is_before_trx_rollback_phase
-		   || trx_rollback_or_clean_is_active)
+		   || trx_rollback_is_active)
 		  && purge_sys->state == PURGE_STATE_INIT)
 	      || (srv_force_recovery >= SRV_FORCE_NO_BACKGROUND
 		  && purge_sys->state == PURGE_STATE_DISABLED)

storage/innobase/trx/trx0roll.cc

@@ -54,8 +54,8 @@ Created 3/26/1996 Heikki Tuuri
 rollback */
 static const ulint TRX_ROLL_TRUNC_THRESHOLD = 1;
 
-/** true if trx_rollback_or_clean_all_recovered() thread is active */
-bool			trx_rollback_or_clean_is_active;
+/** true if trx_rollback_all_recovered() thread is active */
+bool			trx_rollback_is_active;
 
 /** In crash recovery, the current trx to be rolled back; NULL otherwise */
 const trx_t*		trx_roll_crash_recv_trx;
@@ -734,7 +734,7 @@ ibool
 trx_rollback_resurrected(
 /*=====================*/
 	trx_t*	trx,	/*!< in: transaction to rollback or clean */
-	ibool*	all)	/*!< in/out: FALSE=roll back dictionary transactions;
+	bool*	all)	/*!< in/out: FALSE=roll back dictionary transactions;
 			TRUE=roll back all non-PREPARED transactions */
 {
 	ut_ad(trx_sys_mutex_own());
@@ -752,14 +752,6 @@ trx_rollback_resurrected(
 	}
 
 	switch (trx->state) {
-	case TRX_STATE_COMMITTED_IN_MEMORY:
-		trx_mutex_exit(trx);
-		trx_sys_mutex_exit();
-		ib::info() << "Cleaning up trx with id " << ib::hex(trx->id);
-
-		trx_cleanup_at_db_startup(trx);
-		trx_free_resurrected(trx);
-		return(TRUE);
 	case TRX_STATE_ACTIVE:
 		if (!srv_is_being_started
 		    && !srv_undo_sources && srv_fast_shutdown) {
@@ -767,7 +759,7 @@ trx_rollback_resurrected(
 			trx->state = TRX_STATE_PREPARED;
 			trx_sys->n_prepared_trx++;
 			trx_sys->n_prepared_recovered_trx++;
-			*all = FALSE;
+			*all = false;
 			goto func_exit;
 		}
 		trx_mutex_exit(trx);
@@ -788,6 +780,8 @@ trx_rollback_resurrected(
 			return(TRUE);
 		}
 		return(FALSE);
+	case TRX_STATE_COMMITTED_IN_MEMORY:
+		ut_ad(trx->xid);
 	case TRX_STATE_PREPARED:
 		goto func_exit;
 	case TRX_STATE_NOT_STARTED:
@@ -848,12 +842,11 @@ trx_roll_must_shutdown()
 Rollback or clean up any incomplete transactions which were
 encountered in crash recovery.  If the transaction already was
 committed, then we clean up a possible insert undo log. If the
-transaction was not yet committed, then we roll it back. */
+transaction was not yet committed, then we roll it back.
+@param all true=roll back all recovered active transactions;
+false=roll back any incomplete dictionary transaction */
 void
-trx_rollback_or_clean_recovered(
-/*============================*/
-	ibool	all)	/*!< in: FALSE=roll back dictionary transactions;
-			TRUE=roll back all non-PREPARED transactions */
+trx_rollback_recovered(bool all)
 {
 	trx_t*	trx;
 
@@ -917,11 +910,7 @@ Note: this is done in a background thread.
 @return a dummy parameter */
 extern "C"
 os_thread_ret_t
-DECLARE_THREAD(trx_rollback_or_clean_all_recovered)(
-/*================================================*/
-	void*	arg MY_ATTRIBUTE((unused)))
-			/*!< in: a dummy parameter required by
-			os_thread_create */
+DECLARE_THREAD(trx_rollback_all_recovered)(void*)
 {
 	my_thread_init();
 	ut_ad(!srv_read_only_mode);
@@ -930,9 +919,9 @@ DECLARE_THREAD(trx_rollback_or_clean_all_recovered)(
 	pfs_register_thread(trx_rollback_clean_thread_key);
 #endif /* UNIV_PFS_THREAD */
 
-	trx_rollback_or_clean_recovered(TRUE);
+	trx_rollback_recovered(true);
 
-	trx_rollback_or_clean_is_active = false;
+	trx_rollback_is_active = false;
 
 	my_thread_end();
 	/* We count the number of threads in os_thread_exit(). A created

storage/innobase/trx/trx0undo.cc

@@ -1643,7 +1643,6 @@ trx_undo_free_prepared(
 			/* fall through */
 		case TRX_UNDO_ACTIVE:
 			/* lock_trx_release_locks() assigns
-			trx->is_recovered=false and
 			trx->state = TRX_STATE_COMMITTED_IN_MEMORY,
 			also for transactions that we faked
 			to TRX_STATE_PREPARED in trx_rollback_resurrected(). */
@@ -1673,7 +1672,6 @@ trx_undo_free_prepared(
 			/* fall through */
 		case TRX_UNDO_ACTIVE:
 			/* lock_trx_release_locks() assigns
-			trx->is_recovered=false and
 			trx->state = TRX_STATE_COMMITTED_IN_MEMORY,
 			also for transactions that we faked
 			to TRX_STATE_PREPARED in trx_rollback_resurrected(). */