MDEV-14756 - Remove trx_sys_t::rw_trx_list

Use atomic operations when accessing trx_sys_t::max_trx_id. We can't yet
move trx_sys_t::get_new_trx_id() out of mutex because it must be updated
atomically along with trx_sys_t::rw_trx_ids.

storage/innobase/handler/ha_innodb.cc

@@ -3645,7 +3645,7 @@ static ulonglong innodb_prepare_commit_versioned(THD* thd, ulonglong *trx_id)
 				DBUG_ASSERT(trx->rsegs.m_redo.rseg);
 
 				mutex_enter(&trx_sys->mutex);
-				trx_id_t commit_id = trx_sys_get_new_trx_id();
+				trx_id_t commit_id = trx_sys->get_new_trx_id();
 				mutex_exit(&trx_sys->mutex);
 
 				return commit_id;
@@ -19907,7 +19907,7 @@ wsrep_fake_trx_id(
 	THD		*thd)	/*!< in: user thread handle */
 {
 	mutex_enter(&trx_sys->mutex);
-	trx_id_t trx_id = trx_sys_get_new_trx_id();
+	trx_id_t trx_id = trx_sys->get_new_trx_id();
 	mutex_exit(&trx_sys->mutex);
 	WSREP_DEBUG("innodb fake trx id: " TRX_ID_FMT " thd: %s",
 		    trx_id, wsrep_thd_query(thd));

storage/innobase/include/lock0lock.h

@@ -609,7 +609,7 @@ lock_report_trx_id_insanity(
 	const rec_t*	rec,		/*!< in: user record */
 	dict_index_t*	index,		/*!< in: index */
 	const ulint*	offsets,	/*!< in: rec_get_offsets(rec, index) */
-	trx_id_t	max_trx_id);	/*!< in: trx_sys_get_max_trx_id() */
+	trx_id_t	max_trx_id);	/*!< in: trx_sys->get_max_trx_id() */
 /*********************************************************************//**
 Prints info of locks for all transactions.
 @return FALSE if not able to obtain lock mutex and exits without
@@ -827,7 +827,6 @@ Set the lock system timeout event. */
 void
 lock_set_timeout_event();
 /*====================*/
-#ifdef UNIV_DEBUG
 /*********************************************************************//**
 Checks that a transaction id is sensible, i.e., not in the future.
 @return true if ok */
@@ -837,8 +836,8 @@ lock_check_trx_id_sanity(
 	trx_id_t	trx_id,		/*!< in: trx id */
 	const rec_t*	rec,		/*!< in: user record */
 	dict_index_t*	index,		/*!< in: index */
-	const ulint*	offsets)	/*!< in: rec_get_offsets(rec, index) */
-	MY_ATTRIBUTE((warn_unused_result));
+	const ulint*	offsets);	/*!< in: rec_get_offsets(rec, index) */
+#ifdef UNIV_DEBUG
 /*******************************************************************//**
 Check if the transaction holds any locks on the sys tables
 or its records.

storage/innobase/include/trx0sys.h

@@ -132,21 +132,6 @@ trx_sysf_rseg_set_page_no(
 	ulint		page_no,	/*!< in: page number, FIL_NULL if
 					the slot is reset to unused */
 	mtr_t*		mtr);		/*!< in: mtr */
-/*****************************************************************//**
-Allocates a new transaction id.
-@return new, allocated trx id */
-UNIV_INLINE
-trx_id_t
-trx_sys_get_new_trx_id();
-/*===================*/
-/*****************************************************************//**
-Determines the maximum transaction id.
-@return maximum currently allocated trx id; will be stale after the
-next call to trx_sys_get_new_trx_id() */
-UNIV_INLINE
-trx_id_t
-trx_sys_get_max_trx_id(void);
-/*========================*/
 
 #ifdef UNIV_DEBUG
 /* Flag to control TRX_RSEG_N_SLOTS behavior debugging. */
@@ -419,6 +404,11 @@ FIL_PAGE_ARCH_LOG_NO_OR_SPACE_ID. */
 
 /** Size of the doublewrite block in pages */
 #define TRX_SYS_DOUBLEWRITE_BLOCK_SIZE	FSP_EXTENT_SIZE
+
+/** When a trx id which is zero modulo this number (which must be a power of
+two) is assigned, the field TRX_SYS_TRX_ID_STORE on the transaction system
+page is updated */
+#define TRX_SYS_TRX_ID_WRITE_MARGIN	((trx_id_t) 256)
 /* @} */
 
 trx_t* current_trx();
@@ -847,20 +837,24 @@ class rw_trx_hash_t
 
 /** The transaction system central memory data structure. */
 struct trx_sys_t {
+private:
+  /**
+    The smallest number not yet assigned as a transaction id or transaction
+    number. Accessed and updated with atomic operations.
+  */
+
+  char pad0[CACHE_LINE_SIZE];
+  trx_id_t m_max_trx_id;
+  char pad1[CACHE_LINE_SIZE];
+
 
+public:
 	TrxSysMutex	mutex;		/*!< mutex protecting most fields in
 					this structure except when noted
 					otherwise */
 
 	MVCC*		mvcc;		/*!< Multi version concurrency control
 					manager */
-	volatile trx_id_t
-			max_trx_id;	/*!< The smallest number not yet
-					assigned as a transaction id or
-					transaction number. This is declared
-					volatile because it can be accessed
-					without holding any mutex during
-					AC-NL-RO view creation. */
 	trx_ut_list_t	serialisation_list;
 					/*!< Ordered on trx_t::no of all the
 					currenrtly active RW transactions */
@@ -870,7 +864,7 @@ struct trx_sys_t {
 #endif /* UNIV_DEBUG */
 
 	/** Avoid false sharing */
-	const char	pad2[CACHE_LINE_SIZE];
+	char	pad2[CACHE_LINE_SIZE];
 	trx_ut_list_t	mysql_trx_list;	/*!< List of transactions created
 					for MySQL. All user transactions are
 					on mysql_trx_list. The rw_trx_hash
@@ -891,11 +885,11 @@ struct trx_sys_t {
 					consistent snapshot. */
 
 	/** Avoid false sharing */
-	const char	pad3[CACHE_LINE_SIZE];
+	char	pad3[CACHE_LINE_SIZE];
 	/** Temporary rollback segments */
 	trx_rseg_t*	temp_rsegs[TRX_SYS_N_RSEGS];
 	/** Avoid false sharing */
-	const char	pad4[CACHE_LINE_SIZE];
+	char	pad4[CACHE_LINE_SIZE];
 
 	trx_rseg_t*	rseg_array[TRX_SYS_N_RSEGS];
 					/*!< Pointer array to rollback
@@ -910,7 +904,7 @@ struct trx_sys_t {
 					transactions), protected by
 					rseg->mutex */
 
-  const char rw_trx_hash_pre_pad[CACHE_LINE_SIZE];
+  char rw_trx_hash_pre_pad[CACHE_LINE_SIZE];
 
 
   /**
@@ -919,7 +913,7 @@ struct trx_sys_t {
   */
 
   rw_trx_hash_t rw_trx_hash;
-  const char rw_trx_hash_post_pad[CACHE_LINE_SIZE];
+  char rw_trx_hash_post_pad[CACHE_LINE_SIZE];
 
 	ulint		n_prepared_trx;	/*!< Number of transactions currently
 					in the XA PREPARED state */
@@ -940,18 +934,64 @@ struct trx_sys_t {
     must look at the trx->state to find out if the minimum trx id transaction
     itself is active, or already committed.)
 
-    @return the minimum trx id, or trx_sys->max_trx_id if the trx list is empty
+    @return the minimum trx id, or m_max_trx_id if the trx list is empty
   */
 
   trx_id_t get_min_trx_id()
   {
-    trx_id_t id= trx_sys_get_max_trx_id();
+    trx_id_t id= get_max_trx_id();
     rw_trx_hash.iterate(reinterpret_cast<my_hash_walk_action>
                         (get_min_trx_id_callback), &id);
     return id;
   }
 
 
+  /**
+    Determines the maximum transaction id.
+
+    @return maximum currently allocated trx id; will be stale after the
+            next call to trx_sys->get_new_trx_id()
+  */
+
+  trx_id_t get_max_trx_id(void)
+  {
+    return static_cast<trx_id_t>
+           (my_atomic_load64_explicit(reinterpret_cast<int64*>(&m_max_trx_id),
+                                      MY_MEMORY_ORDER_RELAXED));
+  }
+
+
+  /**
+    Allocates a new transaction id.
+
+    VERY important: after the database is started, m_max_trx_id value is
+    divisible by TRX_SYS_TRX_ID_WRITE_MARGIN, and the following if
+    will evaluate to TRUE when this function is first time called,
+    and the value for trx id will be written to disk-based header!
+    Thus trx id values will not overlap when the database is
+    repeatedly started!
+
+    @return new, allocated trx id
+  */
+
+  trx_id_t get_new_trx_id()
+  {
+    ut_ad(mutex_own(&trx_sys->mutex));
+    trx_id_t id= static_cast<trx_id_t>(my_atomic_add64_explicit(
+      reinterpret_cast<int64*>(&m_max_trx_id), 1, MY_MEMORY_ORDER_RELAXED));
+
+    if (UNIV_UNLIKELY(!(id % TRX_SYS_TRX_ID_WRITE_MARGIN)))
+      flush_max_trx_id();
+    return(id);
+  }
+
+
+  void init_max_trx_id(trx_id_t value)
+  {
+    m_max_trx_id= value;
+  }
+
+
 private:
   static my_bool get_min_trx_id_callback(rw_trx_hash_element_t *element,
                                          trx_id_t *id)
@@ -966,12 +1006,14 @@ struct trx_sys_t {
     }
     return 0;
   }
-};
 
-/** When a trx id which is zero modulo this number (which must be a power of
-two) is assigned, the field TRX_SYS_TRX_ID_STORE on the transaction system
-page is updated */
-#define TRX_SYS_TRX_ID_WRITE_MARGIN	((trx_id_t) 256)
+
+  /**
+    Writes the value of m_max_trx_id to the file based trx system header.
+  */
+
+  void flush_max_trx_id();
+};
 
 /** Test if trx_sys->mutex is owned. */
 #define trx_sys_mutex_own() (trx_sys->mutex.is_owned())

storage/innobase/include/trx0sys.ic

@@ -45,12 +45,6 @@ typedef byte	trx_sysf_rseg_t;
 /* Size of a rollback segment specification slot */
 #define TRX_SYS_RSEG_SLOT_SIZE	8
 
-/*****************************************************************//**
-Writes the value of max_trx_id to the file based trx system header. */
-void
-trx_sys_flush_max_trx_id(void);
-/*==========================*/
-
 /** Checks if a page address is the trx sys header page.
 @param[in]	page_id	page id
 @return true if trx sys header page */
@@ -191,58 +185,3 @@ trx_write_trx_id(
 	ut_ad(id > 0);
 	mach_write_to_6(ptr, id);
 }
-
-/*****************************************************************//**
-Allocates a new transaction id.
-@return new, allocated trx id */
-UNIV_INLINE
-trx_id_t
-trx_sys_get_new_trx_id()
-/*====================*/
-{
-	/* wsrep_fake_trx_id  violates this assert */
-	ut_ad(trx_sys_mutex_own());
-
-	/* VERY important: after the database is started, max_trx_id value is
-	divisible by TRX_SYS_TRX_ID_WRITE_MARGIN, and the following if
-	will evaluate to TRUE when this function is first time called,
-	and the value for trx id will be written to disk-based header!
-	Thus trx id values will not overlap when the database is
-	repeatedly started! */
-
-	if (!(trx_sys->max_trx_id % TRX_SYS_TRX_ID_WRITE_MARGIN)) {
-
-		trx_sys_flush_max_trx_id();
-	}
-
-	return(trx_sys->max_trx_id++);
-}
-
-/*****************************************************************//**
-Determines the maximum transaction id.
-@return maximum currently allocated trx id; will be stale after the
-next call to trx_sys_get_new_trx_id() */
-UNIV_INLINE
-trx_id_t
-trx_sys_get_max_trx_id(void)
-/*========================*/
-{
-	ut_ad(!trx_sys_mutex_own());
-
-#if UNIV_WORD_SIZE < DATA_TRX_ID_LEN
-	/* Avoid torn reads. */
-
-	trx_sys_mutex_enter();
-
-	trx_id_t	max_trx_id = trx_sys->max_trx_id;
-
-	trx_sys_mutex_exit();
-
-	return(max_trx_id);
-#else
-	/* Perform a dirty read. Callers should be prepared for stale
-	values, and we know that the value fits in a machine word, so
-	that it will be read and written atomically. */
-	return(trx_sys->max_trx_id);
-#endif /* UNIV_WORD_SIZE < DATA_TRX_ID_LEN */
-}

storage/innobase/lock/lock0lock.cc

@@ -354,7 +354,7 @@ lock_report_trx_id_insanity(
 	const rec_t*	rec,		/*!< in: user record */
 	dict_index_t*	index,		/*!< in: index */
 	const ulint*	offsets,	/*!< in: rec_get_offsets(rec, index) */
-	trx_id_t	max_trx_id)	/*!< in: trx_sys_get_max_trx_id() */
+	trx_id_t	max_trx_id)	/*!< in: trx_sys->get_max_trx_id() */
 {
 	ut_ad(rec_offs_validate(rec, index, offsets));
 	ut_ad(!rec_is_default_row(rec, index));
@@ -371,11 +371,6 @@ lock_report_trx_id_insanity(
 /*********************************************************************//**
 Checks that a transaction id is sensible, i.e., not in the future.
 @return true if ok */
-#ifdef UNIV_DEBUG
-
-#else
-static MY_ATTRIBUTE((warn_unused_result))
-#endif
 bool
 lock_check_trx_id_sanity(
 /*=====================*/
@@ -387,15 +382,14 @@ lock_check_trx_id_sanity(
 	ut_ad(rec_offs_validate(rec, index, offsets));
 	ut_ad(!rec_is_default_row(rec, index));
 
-	trx_id_t	max_trx_id = trx_sys_get_max_trx_id();
-	bool		is_ok = trx_id < max_trx_id;
+	trx_id_t	max_trx_id = trx_sys->get_max_trx_id();
 
-	if (!is_ok) {
+	if (trx_id >= max_trx_id) {
 		lock_report_trx_id_insanity(
 			trx_id, rec, index, offsets, max_trx_id);
+                return false;
 	}
-
-	return(is_ok);
+	return(true);
 }
 
 /*********************************************************************//**
@@ -5215,7 +5209,7 @@ lock_release(
 {
 	lock_t*		lock;
 	ulint		count = 0;
-	trx_id_t	max_trx_id = trx_sys_get_max_trx_id();
+	trx_id_t	max_trx_id = trx_sys->get_max_trx_id();
 
 	ut_ad(lock_mutex_own());
 	ut_ad(!trx_mutex_own(trx));
@@ -5639,7 +5633,7 @@ lock_print_info_summary(
 	      "------------\n", file);
 
 	fprintf(file, "Trx id counter " TRX_ID_FMT "\n",
-		trx_sys_get_max_trx_id());
+		trx_sys->get_max_trx_id());
 
 	fprintf(file,
 		"Purge done for trx's n:o < " TRX_ID_FMT

storage/innobase/page/page0page.cc

@@ -2432,7 +2432,7 @@ page_validate(
 	    && page_is_leaf(page)
 	    && !page_is_empty(page)) {
 		trx_id_t	max_trx_id	= page_get_max_trx_id(page);
-		trx_id_t	sys_max_trx_id	= trx_sys_get_max_trx_id();
+		trx_id_t	sys_max_trx_id	= trx_sys->get_max_trx_id();
 
 		if (max_trx_id == 0 || max_trx_id > sys_max_trx_id) {
 			ib::error() << "PAGE_MAX_TRX_ID out of bounds: "

storage/innobase/read/read0read.cc

@@ -459,7 +459,7 @@ ReadView::prepare(trx_id_t id)
 
 	m_creator_trx_id = id;
 
-	m_low_limit_no = m_low_limit_id = trx_sys->max_trx_id;
+	m_low_limit_no = m_low_limit_id = trx_sys->get_max_trx_id();
 
 	if (!trx_sys->rw_trx_ids.empty()) {
 		copy_trx_ids(trx_sys->rw_trx_ids);
@@ -580,7 +580,7 @@ MVCC::view_open(ReadView*& view, trx_t* trx)
 
 			view->m_closed = false;
 
-			if (view->m_low_limit_id == trx_sys_get_max_trx_id()) {
+			if (view->m_low_limit_id == trx_sys->get_max_trx_id()) {
 				return;
 			} else {
 				view->m_closed = true;

storage/innobase/row/row0vers.cc

@@ -126,14 +126,7 @@ row_vers_impl_x_locked_low(
 	if (trx == 0) {
 		/* The transaction that modified or inserted clust_rec is no
 		longer active, or it is corrupt: no implicit lock on rec */
-		trx_sys_mutex_enter();
-		bool corrupt = trx_id >= trx_sys->max_trx_id;
-		trx_sys_mutex_exit();
-		if (corrupt) {
-			lock_report_trx_id_insanity(
-				trx_id, clust_rec, clust_index, clust_offsets,
-				trx_sys_get_max_trx_id());
-		}
+		lock_check_trx_id_sanity(trx_id, clust_rec, clust_index, clust_offsets);
 		mem_heap_free(heap);
 		DBUG_RETURN(0);
 	}

storage/innobase/trx/trx0sys.cc

@@ -58,7 +58,7 @@ ReadView::check_trx_id_sanity(
 	trx_id_t		id,
 	const table_name_t&	name)
 {
-	if (id >= trx_sys->max_trx_id) {
+	if (id >= trx_sys->get_max_trx_id()) {
 
 		ib::warn() << "A transaction id"
 			   << " in a record of table "
@@ -89,33 +89,25 @@ ReadView::check_trx_id_sanity(
 uint	trx_rseg_n_slots_debug = 0;
 #endif
 
-/*****************************************************************//**
-Writes the value of max_trx_id to the file based trx system header. */
-void
-trx_sys_flush_max_trx_id(void)
-/*==========================*/
-{
-	mtr_t		mtr;
-	trx_sysf_t*	sys_header;
-
-	/* wsrep_fake_trx_id  violates this assert
-	Copied from trx_sys_get_new_trx_id
-	*/
-	ut_ad(trx_sys_mutex_own());
-
-	if (!srv_read_only_mode) {
-		mtr_start(&mtr);
 
-		sys_header = trx_sysf_get(&mtr);
-
-		mlog_write_ull(
-			sys_header + TRX_SYS_TRX_ID_STORE,
-			trx_sys->max_trx_id, &mtr);
+/**
+  Writes the value of m_max_trx_id to the file based trx system header.
+*/
 
-		mtr_commit(&mtr);
+void trx_sys_t::flush_max_trx_id()
+{
+  ut_ad(trx_sys->mutex.is_owned());
+  if (!srv_read_only_mode)
+  {
+    mtr_t mtr;
+    mtr.start();
+    mlog_write_ull(trx_sysf_get(&mtr) + TRX_SYS_TRX_ID_STORE,
+                   trx_sys->get_max_trx_id(), &mtr);
+    mtr.commit();
   }
 }
 
+
 /*****************************************************************//**
 Updates the offset information about the end of the MySQL binlog entry
 which corresponds to the transaction just being committed. In a MySQL
@@ -432,7 +424,7 @@ trx_sys_init_at_db_start()
 
 	/* VERY important: after the database is started, max_trx_id value is
 	divisible by TRX_SYS_TRX_ID_WRITE_MARGIN, and the 'if' in
-	trx_sys_get_new_trx_id will evaluate to TRUE when the function
+	trx_sys->get_new_trx_id will evaluate to TRUE when the function
 	is first time called, and the value for trx id will be written
 	to the disk-based header! Thus trx id values will not overlap when
 	the database is repeatedly started! */
@@ -442,13 +434,13 @@ trx_sys_init_at_db_start()
 
 	sys_header = trx_sysf_get(&mtr);
 
-	trx_sys->max_trx_id = 2 * TRX_SYS_TRX_ID_WRITE_MARGIN
+	trx_sys->init_max_trx_id(2 * TRX_SYS_TRX_ID_WRITE_MARGIN
 		+ ut_uint64_align_up(mach_read_from_8(sys_header
 						   + TRX_SYS_TRX_ID_STORE),
-				     TRX_SYS_TRX_ID_WRITE_MARGIN);
+				     TRX_SYS_TRX_ID_WRITE_MARGIN));
 
 	mtr.commit();
-	ut_d(trx_sys->rw_max_trx_id = trx_sys->max_trx_id);
+	ut_d(trx_sys->rw_max_trx_id = trx_sys->get_max_trx_id());
 
 	trx_dummy_sess = sess_open();
 

storage/innobase/trx/trx0trx.cc

@@ -969,7 +969,7 @@ trx_lists_init_at_db_start()
 			" cleaned up in total " << rows_to_undo
 			<< " row operations to undo";
 
-		ib::info() << "Trx id counter is " << trx_sys->max_trx_id;
+		ib::info() << "Trx id counter is " << trx_sys->get_max_trx_id();
 	}
 
 	std::sort(trx_sys->rw_trx_ids.begin(), trx_sys->rw_trx_ids.end());
@@ -1089,7 +1089,7 @@ trx_t::assign_temp_rseg()
 
 	if (id == 0) {
 		mutex_enter(&trx_sys->mutex);
-		id = trx_sys_get_new_trx_id();
+		id = trx_sys->get_new_trx_id();
 		trx_sys->rw_trx_ids.push_back(id);
 		mutex_exit(&trx_sys->mutex);
 		trx_sys->rw_trx_hash.insert(this);
@@ -1180,7 +1180,7 @@ trx_start_low(
 
 		trx_sys_mutex_enter();
 
-		trx->id = trx_sys_get_new_trx_id();
+		trx->id = trx_sys->get_new_trx_id();
 
 		trx_sys->rw_trx_ids.push_back(trx->id);
 
@@ -1212,7 +1212,7 @@ trx_start_low(
 
 				ut_ad(!srv_read_only_mode);
 
-				trx->id = trx_sys_get_new_trx_id();
+				trx->id = trx_sys->get_new_trx_id();
 
 				trx_sys->rw_trx_ids.push_back(trx->id);
 
@@ -1249,7 +1249,7 @@ trx_serialise(trx_t* trx, trx_rseg_t* rseg)
 
 	trx_sys_mutex_enter();
 
-	trx->no = trx_sys_get_new_trx_id();
+	trx->no = trx_sys->get_new_trx_id();
 
 	/* Track the minimum serialisation number. */
 	UT_LIST_ADD_LAST(trx_sys->serialisation_list, trx);
@@ -2768,7 +2768,7 @@ trx_set_rw_mode(
 	ut_ad(trx->rsegs.m_redo.rseg != 0);
 
 	mutex_enter(&trx_sys->mutex);
-	trx->id = trx_sys_get_new_trx_id();
+	trx->id = trx_sys->get_new_trx_id();
 
 	trx_sys->rw_trx_ids.push_back(trx->id);