Fix Bug #54538 - use of exclusive innodb dictionary lock limits performance.

            
This patch doesn't get rid of the need to acquire the dict_sys->mutex but
reduces the need to keep the mutex locked for the duration of the query
to fsp_get_available_space_in_free_extents() from ha_innobase::info().

rb://390.

storage/innobase/fil/fil0fil.c

@@ -966,6 +966,8 @@ try_again:
 	HASH_SEARCH(name_hash, system->name_hash, ut_fold_string(name), space,
 		    0 == strcmp(name, space->name));
 	if (space != NULL) {
+		ibool	success;
+
 		ut_print_timestamp(stderr);
 		fprintf(stderr,
 			"  InnoDB: Warning: trying to init to the"
@@ -1002,9 +1004,10 @@ try_again:
 
 		namesake_id = space->id;
 
-		mutex_exit(&(system->mutex));
+		success = fil_space_free(namesake_id, FALSE);
+		ut_a(success);
 
-		fil_space_free(namesake_id);
+		mutex_exit(&(system->mutex));
 
 		goto try_again;
 	}
@@ -1127,6 +1130,33 @@ fil_assign_new_space_id(void)
 	return(id);
 }
 
+/***********************************************************************
+Check if the space id exists in the cache, complain to stderr if the
+space id cannot be found. */
+static
+fil_space_t*
+fil_space_search(
+/*=============*/
+			/* out: file space instance*/
+	ulint	id)	/* in: space id */
+{
+	fil_space_t*	space;
+
+	ut_ad(mutex_own(&fil_system->mutex));
+
+	HASH_SEARCH(hash, fil_system->spaces, id, space, space->id == id);
+
+	if (space == NULL) {
+		ut_print_timestamp(stderr);
+		fprintf(stderr,
+			"  InnoDB: Error: trying to remove tablespace %lu"
+			" from the cache but\n"
+			"InnoDB: it is not there.\n", (ulong) id);
+	}
+
+	return(space);
+}
+
 /***********************************************************************
 Frees a space object from the tablespace memory cache. Closes the files in
 the chain but does not delete them. There must not be any pending i/o's or
@@ -1135,27 +1165,21 @@ flushes on the files. */
 ibool
 fil_space_free(
 /*===========*/
-			/* out: TRUE if success */
-	ulint	id)	/* in: space id */
+				/* out: TRUE if success */
+	ulint	id,		/* in: space id */
+	ibool	x_latched)	/* in: TRUE if caller has space->latch
+				in X mode */
 {
 	fil_system_t*	system = fil_system;
 	fil_space_t*	space;
 	fil_space_t*	namespace;
 	fil_node_t*	fil_node;
 
-	mutex_enter(&(system->mutex));
+	ut_ad(mutex_own(&fil_system->mutex));
 
-	HASH_SEARCH(hash, system->spaces, id, space, space->id == id);
-
-	if (!space) {
-		ut_print_timestamp(stderr);
-		fprintf(stderr,
-			"  InnoDB: Error: trying to remove tablespace %lu"
-			" from the cache but\n"
-			"InnoDB: it is not there.\n", (ulong) id);
-
-		mutex_exit(&(system->mutex));
+	space = fil_space_search(id);
 
+	if (space == NULL) {
 		return(FALSE);
 	}
 
@@ -1191,7 +1215,9 @@ fil_space_free(
 
 	ut_a(0 == UT_LIST_GET_LEN(space->chain));
 
-	mutex_exit(&(system->mutex));
+	if (x_latched) {
+		rw_lock_x_unlock(&space->latch);
+	}
 
 	rw_lock_free(&(space->latch));
 
@@ -2048,6 +2074,19 @@ try_again:
 	path = mem_strdup(space->name);
 
 	mutex_exit(&(system->mutex));
+
+	/* Important: We rely on the data dictionary mutex to ensure
+	that a race is not possible here. It should serialize the tablespace
+	drop/free. We acquire an X latch only to avoid a race condition
+	when accessing the tablespace instance via:
+
+	  fsp_get_available_space_in_free_extents().
+
+	There our main motivation is to reduce the contention on the
+	dictionary mutex and not correctness. */
+
+	rw_lock_x_lock(&space->latch);
+
 #ifndef UNIV_HOTBACKUP
 	/* Invalidate in the buffer pool all pages belonging to the
 	tablespace. Since we have set space->is_being_deleted = TRUE, readahead
@@ -2060,7 +2099,11 @@ try_again:
 #endif
 	/* printf("Deleting tablespace %s id %lu\n", space->name, id); */
 
-	success = fil_space_free(id);
+	mutex_enter(&system->mutex);
+
+	success = fil_space_free(id, TRUE);
+
+	mutex_exit(&system->mutex);
 
 	if (success) {
 		success = os_file_delete(path);
@@ -2068,6 +2111,8 @@ try_again:
 		if (!success) {
 			success = os_file_delete_if_exists(path);
 		}
+	} else {
+		rw_lock_x_unlock(&space->latch);
 	}
 
 	if (success) {
@@ -4569,3 +4614,28 @@ fil_page_get_type(
 
 	return(mach_read_from_2(page + FIL_PAGE_TYPE));
 }
+
+/***********************************************************************
+Returns TRUE if a single-table tablespace is being deleted. */
+
+ibool
+fil_tablespace_is_being_deleted(
+/*============================*/
+				/* out: TRUE if space is being deleted */
+	ulint		id)	/* in: space id */
+{
+	fil_space_t*	space;
+	ibool		is_being_deleted;
+
+	mutex_enter(&fil_system->mutex);
+
+	HASH_SEARCH(hash, fil_system->spaces, id, space, space->id == id);
+
+	ut_a(space != NULL);
+
+	is_being_deleted = space->is_being_deleted;
+
+	mutex_exit(&fil_system->mutex);
+
+	return(is_being_deleted);
+}

storage/innobase/fsp/fsp0fsp.c

@@ -2842,12 +2842,61 @@ fsp_get_available_space_in_free_extents(
 
 	ut_ad(!mutex_own(&kernel_mutex));
 
+	/* The convoluted mutex acquire is to overcome latching order
+	issues: The problem is that the fil_mutex is at a lower level
+	than the tablespace latch and the buffer pool mutex. We have to
+	first prevent any operations on the file system by acquiring the
+	dictionary mutex. Then acquire the tablespace latch to obey the
+	latching order and then release the dictionary mutex. That way we
+	ensure that the tablespace instance can't be freed while we are
+	examining its contents (see fil_space_free()).
+
+	However, there is one further complication, we release the fil_mutex
+	when we need to invalidate the the pages in the buffer pool and we
+	reacquire the fil_mutex when deleting and freeing the tablespace
+	instance in fil0fil.c. Here we need to account for that situation
+	too. */
+
+	dict_mutex_enter_for_mysql();
+
+	/* At this stage there is no guarantee that the tablespace even
+	exists in the cache. */
+
+	if (fil_tablespace_deleted_or_being_deleted_in_mem(space, -1)) {
+
+		dict_mutex_exit_for_mysql();
+
+		return(ULLINT_UNDEFINED);
+	}
+
 	mtr_start(&mtr);
 
 	latch = fil_space_get_latch(space);
 
+	/* This should ensure that the tablespace instance can't be freed
+	by another thread. However, the tablespace pages can still be freed
+	from the buffer pool. We need to check for that again. */
+
 	mtr_x_lock(latch, &mtr);
 
+	dict_mutex_exit_for_mysql();
+
+	/* At this point it is possible for the tablespace to be deleted and
+	its pages removed from the buffer pool. We need to check for that
+	situation. However, the tablespace instance can't be deleted because
+	our latching above should ensure that. */
+
+	if (fil_tablespace_is_being_deleted(space)) {
+
+		mtr_commit(&mtr);
+
+		return(ULLINT_UNDEFINED);
+	}
+
+	/* From here on even if the user has dropped the tablespace, the
+	pages _must_ still exist in the buffer pool and the tablespace
+	instance _must be in the file system hash table. */
+
 	space_header = fsp_get_space_header(space, &mtr);
 
 	size = mtr_read_ulint(space_header + FSP_SIZE, MLOG_4BYTES, &mtr);

storage/innobase/handler/ha_innodb.cc

@@ -6485,20 +6485,12 @@ ha_innobase::info(
 		so the "old" value can remain. delete_length is initialized
 		to 0 in the ha_statistics' constructor. */
 		if (!(flag & HA_STATUS_NO_LOCK)) {
+			ullint	avail_space;
 
-			/* lock the data dictionary to avoid races with
-			ibd_file_missing and tablespace_discarded */
-			row_mysql_lock_data_dictionary(prebuilt->trx);
-
-			/* ib_table->space must be an existent tablespace */
-			if (!ib_table->ibd_file_missing
-			    && !ib_table->tablespace_discarded) {
-
-				stats.delete_length =
-					fsp_get_available_space_in_free_extents(
-						ib_table->space) * 1024;
-			} else {
+			avail_space = fsp_get_available_space_in_free_extents(
+					ib_table->space);
 
+			if (avail_space == ULLINT_UNDEFINED) {
 				THD*	thd;
 
 				thd = ha_thd();
@@ -6515,9 +6507,9 @@ ha_innobase::info(
 					ib_table->name);
 
 				stats.delete_length = 0;
+			} else {
+				stats.delete_length = avail_space * 1024;
 			}
-
-			row_mysql_unlock_data_dictionary(prebuilt->trx);
 		}
 
 		stats.check_time = 0;

storage/innobase/include/fil0fil.h

@@ -202,8 +202,10 @@ the chain but does not delete them. */
 ibool
 fil_space_free(
 /*===========*/
-			/* out: TRUE if success */
-	ulint	id);	/* in: space id */
+				/* out: TRUE if success */
+	ulint	id,		/* in: space id */
+	ibool	x_latched);	/* in: TRUE if caller has space->latch
+				in X mode */
 /***********************************************************************
 Returns the size of the space in pages. The tablespace must be cached in the
 memory cache. */
@@ -710,6 +712,14 @@ fil_page_get_type(
 			written to page, the return value not defined */
 	byte*	page);	/* in: file page */
 
+/***********************************************************************
+Returns TRUE if a single-table tablespace is being deleted. */
+
+ibool
+fil_tablespace_is_being_deleted(
+/*============================*/
+				/* out: TRUE if space is being deleted */
+	ulint		id);	/* in: space id */
 
 typedef	struct fil_space_struct	fil_space_t;
 

storage/innobase/include/univ.i

@@ -234,6 +234,12 @@ typedef unsigned long long int	ullint;
 /* Maximum value for a ulint */
 #define ULINT_MAX		((ulint)(-2))
 
+/* THe 'undefined' value for ullint */
+#define ULLINT_UNDEFINED        ((ullint)(-1))
+
+/* Maximum value for a ullint */
+#define ULLINT_MAX		((ullint)(-2))
+
 /* This 'ibool' type is used within Innobase. Remember that different included
 headers may define 'bool' differently. Do not assume that 'bool' is a ulint! */
 #define ibool			ulint