Parallel pool import

This commit allow spa_load() to drop the spa_namespace_lock so
that imports can happen concurrently. Prior to dropping the
spa_namespace_lock, the import logic will set the spa_load_thread
value to track the thread which is doing the import.

Consumers of spa_lookup() retain the same behavior by blocking
when either a thread is holding the spa_namespace_lock or the
spa_load_thread value is set. This will ensure that critical
concurrent operations cannot take place while a pool is being
imported.

The zpool command is also enhanced to provide multi-threaded support
when invoking zpool import -a.

Lastly, zinject provides a mechanism to insert artificial delays
when importing a pool and new zfs tests are added to verify parallel
import functionality.

Contributions-by: Don Brady <don.brady@klarasystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Wilson <gwilson@delphix.com>
Closes #16093

module/zfs/spa.c

@@ -3273,8 +3273,6 @@ spa_spawn_aux_threads(spa_t *spa)
 {
 	ASSERT(spa_writeable(spa));
 
-	ASSERT(MUTEX_HELD(&spa_namespace_lock));
-
 	spa_start_raidz_expansion_thread(spa);
 	spa_start_indirect_condensing_thread(spa);
 	spa_start_livelist_destroy_thread(spa);
@@ -4981,7 +4979,8 @@ spa_ld_read_checkpoint_txg(spa_t *spa)
 	int error = 0;
 
 	ASSERT0(spa->spa_checkpoint_txg);
-	ASSERT(MUTEX_HELD(&spa_namespace_lock));
+	ASSERT(MUTEX_HELD(&spa_namespace_lock) ||
+	    spa->spa_load_thread == curthread);
 
 	error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 	    DMU_POOL_ZPOOL_CHECKPOINT, sizeof (uint64_t),
@@ -5228,6 +5227,7 @@ spa_load_impl(spa_t *spa, spa_import_type_t type, const char **ereport)
 	boolean_t checkpoint_rewind =
 	    (spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT);
 	boolean_t update_config_cache = B_FALSE;
+	hrtime_t load_start = gethrtime();
 
 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
 	ASSERT(spa->spa_config_source != SPA_CONFIG_SRC_NONE);
@@ -5272,13 +5272,19 @@ spa_load_impl(spa_t *spa, spa_import_type_t type, const char **ereport)
 			return (error);
 	}
 
+	/*
+	 * Drop the namespace lock for the rest of the function.
+	 */
+	spa->spa_load_thread = curthread;
+	mutex_exit(&spa_namespace_lock);
+
 	/*
 	 * Retrieve the checkpoint txg if the pool has a checkpoint.
 	 */
 	spa_import_progress_set_notes(spa, "Loading checkpoint txg");
 	error = spa_ld_read_checkpoint_txg(spa);
 	if (error != 0)
-		return (error);
+		goto fail;
 
 	/*
 	 * Retrieve the mapping of indirect vdevs. Those vdevs were removed
@@ -5291,7 +5297,7 @@ spa_load_impl(spa_t *spa, spa_import_type_t type, const char **ereport)
 	spa_import_progress_set_notes(spa, "Loading indirect vdev metadata");
 	error = spa_ld_open_indirect_vdev_metadata(spa);
 	if (error != 0)
-		return (error);
+		goto fail;
 
 	/*
 	 * Retrieve the full list of active features from the MOS and check if
@@ -5300,7 +5306,7 @@ spa_load_impl(spa_t *spa, spa_import_type_t type, const char **ereport)
 	spa_import_progress_set_notes(spa, "Checking feature flags");
 	error = spa_ld_check_features(spa, &missing_feat_write);
 	if (error != 0)
-		return (error);
+		goto fail;
 
 	/*
 	 * Load several special directories from the MOS needed by the dsl_pool
@@ -5309,7 +5315,7 @@ spa_load_impl(spa_t *spa, spa_import_type_t type, const char **ereport)
 	spa_import_progress_set_notes(spa, "Loading special MOS directories");
 	error = spa_ld_load_special_directories(spa);
 	if (error != 0)
-		return (error);
+		goto fail;
 
 	/*
 	 * Retrieve pool properties from the MOS.
@@ -5317,7 +5323,7 @@ spa_load_impl(spa_t *spa, spa_import_type_t type, const char **ereport)
 	spa_import_progress_set_notes(spa, "Loading properties");
 	error = spa_ld_get_props(spa);
 	if (error != 0)
-		return (error);
+		goto fail;
 
 	/*
 	 * Retrieve the list of auxiliary devices - cache devices and spares -
@@ -5326,7 +5332,7 @@ spa_load_impl(spa_t *spa, spa_import_type_t type, const char **ereport)
 	spa_import_progress_set_notes(spa, "Loading AUX vdevs");
 	error = spa_ld_open_aux_vdevs(spa, type);
 	if (error != 0)
-		return (error);
+		goto fail;
 
 	/*
 	 * Load the metadata for all vdevs. Also check if unopenable devices
@@ -5335,17 +5341,17 @@ spa_load_impl(spa_t *spa, spa_import_type_t type, const char **ereport)
 	spa_import_progress_set_notes(spa, "Loading vdev metadata");
 	error = spa_ld_load_vdev_metadata(spa);
 	if (error != 0)
-		return (error);
+		goto fail;
 
 	spa_import_progress_set_notes(spa, "Loading dedup tables");
 	error = spa_ld_load_dedup_tables(spa);
 	if (error != 0)
-		return (error);
+		goto fail;
 
 	spa_import_progress_set_notes(spa, "Loading BRT");
 	error = spa_ld_load_brt(spa);
 	if (error != 0)
-		return (error);
+		goto fail;
 
 	/*
 	 * Verify the logs now to make sure we don't have any unexpected errors
@@ -5354,7 +5360,7 @@ spa_load_impl(spa_t *spa, spa_import_type_t type, const char **ereport)
 	spa_import_progress_set_notes(spa, "Verifying Log Devices");
 	error = spa_ld_verify_logs(spa, type, ereport);
 	if (error != 0)
-		return (error);
+		goto fail;
 
 	if (missing_feat_write) {
 		ASSERT(spa->spa_load_state == SPA_LOAD_TRYIMPORT);
@@ -5364,8 +5370,9 @@ spa_load_impl(spa_t *spa, spa_import_type_t type, const char **ereport)
 		 * read-only mode but not read-write mode. We now have enough
 		 * information and can return to userland.
 		 */
-		return (spa_vdev_err(spa->spa_root_vdev, VDEV_AUX_UNSUP_FEAT,
-		    ENOTSUP));
+		error = spa_vdev_err(spa->spa_root_vdev, VDEV_AUX_UNSUP_FEAT,
+		    ENOTSUP);
+		goto fail;
 	}
 
 	/*
@@ -5376,7 +5383,7 @@ spa_load_impl(spa_t *spa, spa_import_type_t type, const char **ereport)
 	spa_import_progress_set_notes(spa, "Verifying pool data");
 	error = spa_ld_verify_pool_data(spa);
 	if (error != 0)
-		return (error);
+		goto fail;
 
 	/*
 	 * Calculate the deflated space for the pool. This must be done before
@@ -5501,13 +5508,19 @@ spa_load_impl(spa_t *spa, spa_import_type_t type, const char **ereport)
 		spa_config_exit(spa, SCL_CONFIG, FTAG);
 		spa_import_progress_set_notes(spa, "Finished importing");
 	}
+	zio_handle_import_delay(spa, gethrtime() - load_start);
 
 	spa_import_progress_remove(spa_guid(spa));
 	spa_async_request(spa, SPA_ASYNC_L2CACHE_REBUILD);
 
 	spa_load_note(spa, "LOADED");
+fail:
+	mutex_enter(&spa_namespace_lock);
+	spa->spa_load_thread = NULL;
+	cv_broadcast(&spa_namespace_cv);
+
+	return (error);
 
-	return (0);
 }
 
 static int
@@ -6757,9 +6770,14 @@ spa_tryimport(nvlist_t *tryconfig)
 	/*
 	 * Create and initialize the spa structure.
 	 */
+	char *name = kmem_alloc(MAXPATHLEN, KM_SLEEP);
+	(void) snprintf(name, MAXPATHLEN, "%s-%llx-%s",
+	    TRYIMPORT_NAME, (u_longlong_t)curthread, poolname);
+
 	mutex_enter(&spa_namespace_lock);
-	spa = spa_add(TRYIMPORT_NAME, tryconfig, NULL);
+	spa = spa_add(name, tryconfig, NULL);
 	spa_activate(spa, SPA_MODE_READ);
+	kmem_free(name, MAXPATHLEN);
 
 	/*
 	 * Rewind pool if a max txg was provided.
@@ -6874,6 +6892,7 @@ spa_export_common(const char *pool, int new_state, nvlist_t **oldconfig,
 {
 	int error;
 	spa_t *spa;
+	hrtime_t export_start = gethrtime();
 
 	if (oldconfig)
 		*oldconfig = NULL;
@@ -7018,6 +7037,9 @@ export_spa:
 		spa->spa_is_exporting = B_FALSE;
 	}
 
+	if (new_state == POOL_STATE_EXPORTED)
+		zio_handle_export_delay(spa, gethrtime() - export_start);
+
 	mutex_exit(&spa_namespace_lock);
 	return (0);
 

module/zfs/spa_misc.c

@@ -20,7 +20,7 @@
  */
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
- * Copyright (c) 2011, 2019 by Delphix. All rights reserved.
+ * Copyright (c) 2011, 2024 by Delphix. All rights reserved.
  * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  * Copyright 2013 Saso Kiselkov. All rights reserved.
@@ -82,7 +82,8 @@
  *		- Check if spa_refcount is zero
  *		- Rename a spa_t
  *		- add/remove/attach/detach devices
- *		- Held for the duration of create/destroy/import/export
+ *		- Held for the duration of create/destroy/export
+ *		- Held at the start and end of import
  *
  *	It does not need to handle recursion.  A create or destroy may
  *	reference objects (files or zvols) in other pools, but by
@@ -235,9 +236,9 @@
  * locking is, always, based on spa_namespace_lock and spa_config_lock[].
  */
 
-static avl_tree_t spa_namespace_avl;
+avl_tree_t spa_namespace_avl;
 kmutex_t spa_namespace_lock;
-static kcondvar_t spa_namespace_cv;
+kcondvar_t spa_namespace_cv;
 static const int spa_max_replication_override = SPA_DVAS_PER_BP;
 
 static kmutex_t spa_spare_lock;
@@ -619,6 +620,7 @@ spa_lookup(const char *name)
 
 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
 
+retry:
 	(void) strlcpy(search.spa_name, name, sizeof (search.spa_name));
 
 	/*
@@ -630,6 +632,14 @@ spa_lookup(const char *name)
 		*cp = '\0';
 
 	spa = avl_find(&spa_namespace_avl, &search, &where);
+	if (spa == NULL)
+		return (NULL);
+
+	if (spa->spa_load_thread != NULL &&
+	    spa->spa_load_thread != curthread) {
+		cv_wait(&spa_namespace_cv, &spa_namespace_lock);
+		goto retry;
+	}
 
 	return (spa);
 }
@@ -728,6 +738,7 @@ spa_add(const char *name, nvlist_t *config, const char *altroot)
 	spa_config_lock_init(spa);
 	spa_stats_init(spa);
 
+	ASSERT(MUTEX_HELD(&spa_namespace_lock));
 	avl_add(&spa_namespace_avl, spa);
 
 	/*
@@ -826,7 +837,6 @@ spa_remove(spa_t *spa)
 	nvlist_free(spa->spa_config_splitting);
 
 	avl_remove(&spa_namespace_avl, spa);
-	cv_broadcast(&spa_namespace_cv);
 
 	if (spa->spa_root)
 		spa_strfree(spa->spa_root);
@@ -920,7 +930,8 @@ void
 spa_open_ref(spa_t *spa, const void *tag)
 {
 	ASSERT(zfs_refcount_count(&spa->spa_refcount) >= spa->spa_minref ||
-	    MUTEX_HELD(&spa_namespace_lock));
+	    MUTEX_HELD(&spa_namespace_lock) ||
+	    spa->spa_load_thread == curthread);
 	(void) zfs_refcount_add(&spa->spa_refcount, tag);
 }
 
@@ -932,7 +943,8 @@ void
 spa_close(spa_t *spa, const void *tag)
 {
 	ASSERT(zfs_refcount_count(&spa->spa_refcount) > spa->spa_minref ||
-	    MUTEX_HELD(&spa_namespace_lock));
+	    MUTEX_HELD(&spa_namespace_lock) ||
+	    spa->spa_load_thread == curthread);
 	(void) zfs_refcount_remove(&spa->spa_refcount, tag);
 }
 

module/zfs/vdev_initialize.c

@@ -20,7 +20,7 @@
  */
 
 /*
- * Copyright (c) 2016, 2019 by Delphix. All rights reserved.
+ * Copyright (c) 2016, 2024 by Delphix. All rights reserved.
  */
 
 #include <sys/spa.h>
@@ -775,7 +775,8 @@ vdev_initialize_stop_all(vdev_t *vd, vdev_initializing_state_t tgt_state)
 void
 vdev_initialize_restart(vdev_t *vd)
 {
-	ASSERT(MUTEX_HELD(&spa_namespace_lock));
+	ASSERT(MUTEX_HELD(&spa_namespace_lock) ||
+	    vd->vdev_spa->spa_load_thread == curthread);
 	ASSERT(!spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER));
 
 	if (vd->vdev_leaf_zap != 0) {

module/zfs/vdev_rebuild.c

@@ -23,6 +23,7 @@
  * Copyright (c) 2018, Intel Corporation.
  * Copyright (c) 2020 by Lawrence Livermore National Security, LLC.
  * Copyright (c) 2022 Hewlett Packard Enterprise Development LP.
+ * Copyright (c) 2024 by Delphix. All rights reserved.
  */
 
 #include <sys/vdev_impl.h>
@@ -1071,7 +1072,8 @@ vdev_rebuild_restart_impl(vdev_t *vd)
 void
 vdev_rebuild_restart(spa_t *spa)
 {
-	ASSERT(MUTEX_HELD(&spa_namespace_lock));
+	ASSERT(MUTEX_HELD(&spa_namespace_lock) ||
+	    spa->spa_load_thread == curthread);
 
 	vdev_rebuild_restart_impl(spa->spa_root_vdev);
 }

module/zfs/vdev_trim.c

@@ -20,7 +20,7 @@
  */
 
 /*
- * Copyright (c) 2016 by Delphix. All rights reserved.
+ * Copyright (c) 2016, 2024 by Delphix. All rights reserved.
  * Copyright (c) 2019 by Lawrence Livermore National Security, LLC.
  * Copyright (c) 2021 Hewlett Packard Enterprise Development LP
  * Copyright 2023 RackTop Systems, Inc.
@@ -1148,7 +1148,8 @@ vdev_trim_stop_all(vdev_t *vd, vdev_trim_state_t tgt_state)
 void
 vdev_trim_restart(vdev_t *vd)
 {
-	ASSERT(MUTEX_HELD(&spa_namespace_lock));
+	ASSERT(MUTEX_HELD(&spa_namespace_lock) ||
+	    vd->vdev_spa->spa_load_thread == curthread);
 	ASSERT(!spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER));
 
 	if (vd->vdev_leaf_zap != 0) {
@@ -1568,8 +1569,8 @@ vdev_autotrim_stop_all(spa_t *spa)
 void
 vdev_autotrim_restart(spa_t *spa)
 {
-	ASSERT(MUTEX_HELD(&spa_namespace_lock));
-
+	ASSERT(MUTEX_HELD(&spa_namespace_lock) ||
+	    spa->spa_load_thread == curthread);
 	if (spa->spa_autotrim)
 		vdev_autotrim(spa);
 }

module/zfs/zio_inject.c

@@ -22,6 +22,7 @@
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  * Copyright (c) 2012, 2015 by Delphix. All rights reserved.
  * Copyright (c) 2017, Intel Corporation.
+ * Copyright (c) 2024, Klara Inc.
  */
 
 /*
@@ -59,6 +60,7 @@ uint32_t zio_injection_enabled = 0;
 typedef struct inject_handler {
 	int			zi_id;
 	spa_t			*zi_spa;
+	char			*zi_spa_name; /* ZINJECT_DELAY_IMPORT only */
 	zinject_record_t	zi_record;
 	uint64_t		*zi_lanes;
 	int			zi_next_lane;
@@ -703,6 +705,63 @@ zio_handle_io_delay(zio_t *zio)
 	return (min_target);
 }
 
+static void
+zio_handle_pool_delay(spa_t *spa, hrtime_t elapsed, zinject_type_t command)
+{
+	inject_handler_t *handler;
+	hrtime_t delay = 0;
+	int id = 0;
+
+	rw_enter(&inject_lock, RW_READER);
+
+	for (handler = list_head(&inject_handlers);
+	    handler != NULL && handler->zi_record.zi_cmd == command;
+	    handler = list_next(&inject_handlers, handler)) {
+		ASSERT3P(handler->zi_spa_name, !=, NULL);
+		if (strcmp(spa_name(spa), handler->zi_spa_name) == 0) {
+			uint64_t pause =
+			    SEC2NSEC(handler->zi_record.zi_duration);
+			if (pause > elapsed) {
+				delay = pause - elapsed;
+			}
+			id = handler->zi_id;
+			break;
+		}
+	}
+
+	rw_exit(&inject_lock);
+
+	if (delay) {
+		if (command == ZINJECT_DELAY_IMPORT) {
+			spa_import_progress_set_notes(spa, "injecting %llu "
+			    "sec delay", (u_longlong_t)NSEC2SEC(delay));
+		}
+		zfs_sleep_until(gethrtime() + delay);
+	}
+	if (id) {
+		/* all done with this one-shot handler */
+		zio_clear_fault(id);
+	}
+}
+
+/*
+ * For testing, inject a delay during an import
+ */
+void
+zio_handle_import_delay(spa_t *spa, hrtime_t elapsed)
+{
+	zio_handle_pool_delay(spa, elapsed, ZINJECT_DELAY_IMPORT);
+}
+
+/*
+ * For testing, inject a delay during an export
+ */
+void
+zio_handle_export_delay(spa_t *spa, hrtime_t elapsed)
+{
+	zio_handle_pool_delay(spa, elapsed, ZINJECT_DELAY_EXPORT);
+}
+
 static int
 zio_calculate_range(const char *pool, zinject_record_t *record)
 {
@@ -760,6 +819,28 @@ zio_calculate_range(const char *pool, zinject_record_t *record)
 	return (0);
 }
 
+static boolean_t
+zio_pool_handler_exists(const char *name, zinject_type_t command)
+{
+	boolean_t exists = B_FALSE;
+
+	rw_enter(&inject_lock, RW_READER);
+	for (inject_handler_t *handler = list_head(&inject_handlers);
+	    handler != NULL; handler = list_next(&inject_handlers, handler)) {
+		if (command != handler->zi_record.zi_cmd)
+			continue;
+
+		const char *pool = (handler->zi_spa_name != NULL) ?
+		    handler->zi_spa_name : spa_name(handler->zi_spa);
+		if (strcmp(name, pool) == 0) {
+			exists = B_TRUE;
+			break;
+		}
+	}
+	rw_exit(&inject_lock);
+
+	return (exists);
+}
 /*
  * Create a new handler for the given record.  We add it to the list, adding
  * a reference to the spa_t in the process.  We increment zio_injection_enabled,
@@ -810,16 +891,42 @@ zio_inject_fault(char *name, int flags, int *id, zinject_record_t *record)
 
 	if (!(flags & ZINJECT_NULL)) {
 		/*
-		 * spa_inject_ref() will add an injection reference, which will
-		 * prevent the pool from being removed from the namespace while
-		 * still allowing it to be unloaded.
+		 * Pool delays for import or export don't take an
+		 * injection reference on the spa. Instead they
+		 * rely on matching by name.
 		 */
-		if ((spa = spa_inject_addref(name)) == NULL)
-			return (SET_ERROR(ENOENT));
+		if (record->zi_cmd == ZINJECT_DELAY_IMPORT ||
+		    record->zi_cmd == ZINJECT_DELAY_EXPORT) {
+			if (record->zi_duration <= 0)
+				return (SET_ERROR(EINVAL));
+			/*
+			 * Only one import | export delay handler per pool.
+			 */
+			if (zio_pool_handler_exists(name, record->zi_cmd))
+				return (SET_ERROR(EEXIST));
+
+			mutex_enter(&spa_namespace_lock);
+			boolean_t has_spa = spa_lookup(name) != NULL;
+			mutex_exit(&spa_namespace_lock);
+
+			if (record->zi_cmd == ZINJECT_DELAY_IMPORT && has_spa)
+				return (SET_ERROR(EEXIST));
+			if (record->zi_cmd == ZINJECT_DELAY_EXPORT && !has_spa)
+				return (SET_ERROR(ENOENT));
+			spa = NULL;
+		} else {
+			/*
+			 * spa_inject_ref() will add an injection reference,
+			 * which will prevent the pool from being removed
+			 * from the namespace while still allowing it to be
+			 * unloaded.
+			 */
+			if ((spa = spa_inject_addref(name)) == NULL)
+				return (SET_ERROR(ENOENT));
+		}
 
 		handler = kmem_alloc(sizeof (inject_handler_t), KM_SLEEP);
-
-		handler->zi_spa = spa;
+		handler->zi_spa = spa;	/* note: can be NULL */
 		handler->zi_record = *record;
 
 		if (handler->zi_record.zi_cmd == ZINJECT_DELAY_IO) {
@@ -832,6 +939,11 @@ zio_inject_fault(char *name, int flags, int *id, zinject_record_t *record)
 			handler->zi_next_lane = 0;
 		}
 
+		if (handler->zi_spa == NULL)
+			handler->zi_spa_name = spa_strdup(name);
+		else
+			handler->zi_spa_name = NULL;
+
 		rw_enter(&inject_lock, RW_WRITER);
 
 		/*
@@ -891,7 +1003,11 @@ zio_inject_list_next(int *id, char *name, size_t buflen,
 	if (handler) {
 		*record = handler->zi_record;
 		*id = handler->zi_id;
-		(void) strlcpy(name, spa_name(handler->zi_spa), buflen);
+		ASSERT(handler->zi_spa || handler->zi_spa_name);
+		if (handler->zi_spa != NULL)
+			(void) strlcpy(name, spa_name(handler->zi_spa), buflen);
+		else
+			(void) strlcpy(name, handler->zi_spa_name, buflen);
 		ret = 0;
 	} else {
 		ret = SET_ERROR(ENOENT);
@@ -941,7 +1057,11 @@ zio_clear_fault(int id)
 		ASSERT3P(handler->zi_lanes, ==, NULL);
 	}
 
-	spa_inject_delref(handler->zi_spa);
+	if (handler->zi_spa_name != NULL)
+		spa_strfree(handler->zi_spa_name);
+
+	if (handler->zi_spa != NULL)
+		spa_inject_delref(handler->zi_spa);
 	kmem_free(handler, sizeof (inject_handler_t));
 	atomic_dec_32(&zio_injection_enabled);