Add support for parallel pool exports

Changed spa_export_common() such that it no longer holds the
spa_namespace_lock for the entire duration and instead sets
spa_export_thread to indicate an import is in progress on the
spa.  This allows for an export to a diffent pool to proceed
in parallel while an export is still processing potentially
long operations like spa_unload_log_sm_flush_all().

Calls like spa_lookup() and spa_vdev_enter() that rely on
the spa_namespace_lock to serialize them against a concurrent
export, now wait for any in-progress export thread to complete
before proceeding.

The 'zpool import -a' sub-command also provides multi-threaded
support, using a thread pool to submit the exports in parallel.

Sponsored-By: Klara Inc.
Sponsored-by: Wasabi Technology, Inc.

Signed-off-by: Don Brady <[email protected]>

cmd/zpool/zpool_main.c

@@ -2030,10 +2030,19 @@ zpool_do_destroy(int argc, char **argv)
 }
 
 typedef struct export_cbdata {
+	tpool_t *tpool;
+	pthread_mutex_t mnttab_lock;
 	boolean_t force;
 	boolean_t hardforce;
+	int retval;
 } export_cbdata_t;
 
+
+typedef struct {
+	char *aea_poolname;
+	export_cbdata_t	*aea_cbdata;
+} async_export_args_t;
+
 /*
  * Export one pool
  */
@@ -2042,11 +2051,20 @@ zpool_export_one(zpool_handle_t *zhp, void *data)
 {
 	export_cbdata_t *cb = data;
 
-	if (zpool_disable_datasets(zhp, cb->force) != 0)
-		return (1);
+	/*
+	 * zpool_disable_datasets() is not thread-safe for mnttab access.
+	 * So we serialize access here for 'zpool export -a' parallel case.
+	 */
+	if (cb->tpool != NULL)
+		pthread_mutex_lock(&cb->mnttab_lock);
 
-	/* The history must be logged as part of the export */
-	log_history = B_FALSE;
+	int retval = zpool_disable_datasets(zhp, cb->force);
+
+	if (cb->tpool != NULL)
+		pthread_mutex_unlock(&cb->mnttab_lock);
+
+	if (retval)
+		return (1);
 
 	if (cb->hardforce) {
 		if (zpool_export_force(zhp, history_str) != 0)
@@ -2058,6 +2076,48 @@ zpool_export_one(zpool_handle_t *zhp, void *data)
 	return (0);
 }
 
+/*
+ * Asynchronous export request
+ */
+static void
+zpool_export_task(void *arg)
+{
+	async_export_args_t *aea = arg;
+
+	zpool_handle_t *zhp = zpool_open(g_zfs, aea->aea_poolname);
+	if (zhp != NULL) {
+		int ret = zpool_export_one(zhp, aea->aea_cbdata);
+		if (ret != 0)
+			aea->aea_cbdata->retval = ret;
+		zpool_close(zhp);
+	} else {
+		aea->aea_cbdata->retval = 1;
+	}
+
+	free(aea->aea_poolname);
+	free(aea);
+}
+
+/*
+ * Process an export request in parallel
+ */
+static int
+zpool_export_one_async(zpool_handle_t *zhp, void *data)
+{
+	tpool_t *tpool = ((export_cbdata_t *)data)->tpool;
+	async_export_args_t *aea = safe_malloc(sizeof (async_export_args_t));
+
+	/* save pool name since zhp will go out of scope */
+	aea->aea_poolname = strdup(zpool_get_name(zhp));
+	aea->aea_cbdata = data;
+
+	/* ship off actual export to another thread */
+	if (tpool_dispatch(tpool, zpool_export_task, (void *)aea) != 0)
+		return (errno);	/* unlikely */
+	else
+		return (0);
+}
+
 /*
  * zpool export [-f] <pool> ...
  *
@@ -2098,17 +2158,33 @@ zpool_do_export(int argc, char **argv)
 
 	cb.force = force;
 	cb.hardforce = hardforce;
+	cb.tpool = NULL;
+	cb.retval = 0;
 	argc -= optind;
 	argv += optind;
 
+	/* The history will be logged as part of the export itself */
+	log_history = B_FALSE;
+
 	if (do_all) {
 		if (argc != 0) {
 			(void) fprintf(stderr, gettext("too many arguments\n"));
 			usage(B_FALSE);
 		}
 
-		return (for_each_pool(argc, argv, B_TRUE, NULL,
-		    ZFS_TYPE_POOL, B_FALSE, zpool_export_one, &cb));
+		cb.tpool = tpool_create(1, 5 * sysconf(_SC_NPROCESSORS_ONLN),
+		    0, NULL);
+		pthread_mutex_init(&cb.mnttab_lock, NULL);
+
+		/* Asynchronously call zpool_export_one using thread pool */
+		ret = for_each_pool(argc, argv, B_TRUE, NULL, ZFS_TYPE_POOL,
+		    B_FALSE, zpool_export_one_async, &cb);
+
+		tpool_wait(cb.tpool);
+		tpool_destroy(cb.tpool);
+		(void) pthread_mutex_destroy(&cb.mnttab_lock);
+
+		return (ret | cb.retval);
 	}
 
 	/* check arguments */

include/sys/spa_impl.h

@@ -237,6 +237,7 @@ struct spa {
 	dsl_pool_t	*spa_dsl_pool;
 	boolean_t	spa_is_initializing;	/* true while opening pool */
 	boolean_t	spa_is_exporting;	/* true while exporting pool */
+	kthread_t	*spa_export_thread;	/* valid during pool export */
 	kthread_t	*spa_load_thread;	/* loading, no namespace lock */
 	metaslab_class_t *spa_normal_class;	/* normal data class */
 	metaslab_class_t *spa_log_class;	/* intent log data class */

module/zfs/arc.c

@@ -8143,11 +8143,11 @@ l2arc_dev_get_next(void)
 
 		ASSERT3P(next, !=, NULL);
 	} while (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild ||
-	    next->l2ad_trim_all);
+	    next->l2ad_trim_all || next->l2ad_spa->spa_is_exporting);
 
 	/* if we were unable to find any usable vdevs, return NULL */
 	if (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild ||
-	    next->l2ad_trim_all)
+	    next->l2ad_trim_all || next->l2ad_spa->spa_is_exporting)
 		next = NULL;
 
 	l2arc_dev_last = next;

module/zfs/spa.c

@@ -34,6 +34,7 @@
  * Copyright (c) 2017, Intel Corporation.
  * Copyright (c) 2021, Colm Buckley <[email protected]>
  * Copyright (c) 2023 Hewlett Packard Enterprise Development LP.
+ * Copyright (c) 2024, Klara Inc.
  */
 
 /*
@@ -1991,7 +1992,8 @@ spa_destroy_aux_threads(spa_t *spa)
 static void
 spa_unload(spa_t *spa)
 {
-	ASSERT(MUTEX_HELD(&spa_namespace_lock));
+	ASSERT(MUTEX_HELD(&spa_namespace_lock) ||
+	    spa->spa_export_thread == curthread);
 	ASSERT(spa_state(spa) != POOL_STATE_UNINITIALIZED);
 
 	spa_import_progress_remove(spa_guid(spa));
@@ -6955,7 +6957,7 @@ static int
 spa_export_common(const char *pool, int new_state, nvlist_t **oldconfig,
     boolean_t force, boolean_t hardforce)
 {
-	int error;
+	int error = 0;
 	spa_t *spa;
 	hrtime_t export_start = gethrtime();
 
@@ -6979,8 +6981,8 @@ spa_export_common(const char *pool, int new_state, nvlist_t **oldconfig,
 	spa->spa_is_exporting = B_TRUE;
 
 	/*
-	 * Put a hold on the pool, drop the namespace lock, stop async tasks,
-	 * reacquire the namespace lock, and see if we can export.
+	 * Put a hold on the pool, drop the namespace lock, stop async tasks
+	 * and see if we can export.
 	 */
 	spa_open_ref(spa, FTAG);
 	mutex_exit(&spa_namespace_lock);
@@ -6990,10 +6992,18 @@ spa_export_common(const char *pool, int new_state, nvlist_t **oldconfig,
 		taskq_wait(spa->spa_zvol_taskq);
 	}
 	mutex_enter(&spa_namespace_lock);
+	spa->spa_export_thread = curthread;
 	spa_close(spa, FTAG);
+	mutex_exit(&spa_namespace_lock);
+
+	/*
+	 * At this point we no longer hold the spa_namespace_lock and
+	 * the spa_export_thread indicates that an export is in progress.
+	 */
 
 	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
 		goto export_spa;
+
 	/*
 	 * The pool will be in core if it's openable, in which case we can
 	 * modify its state.  Objsets may be open only because they're dirty,
@@ -7089,6 +7099,10 @@ spa_export_common(const char *pool, int new_state, nvlist_t **oldconfig,
 	if (oldconfig && spa->spa_config)
 		*oldconfig = fnvlist_dup(spa->spa_config);
 
+	if (new_state == POOL_STATE_EXPORTED)
+		zio_handle_export_delay(spa, gethrtime() - export_start);
+
+	mutex_enter(&spa_namespace_lock);
 	if (new_state != POOL_STATE_UNINITIALIZED) {
 		if (!hardforce)
 			spa_write_cachefile(spa, B_TRUE, B_TRUE, B_FALSE);
@@ -7100,17 +7114,25 @@ spa_export_common(const char *pool, int new_state, nvlist_t **oldconfig,
 		 * we make sure to reset the exporting flag.
 		 */
 		spa->spa_is_exporting = B_FALSE;
+		spa->spa_export_thread = NULL;
 	}
 
-	if (new_state == POOL_STATE_EXPORTED)
-		zio_handle_export_delay(spa, gethrtime() - export_start);
-
+	/*
+	 * Wake up any waiters on spa_namespace_lock
+	 * They need to re-attempt a spa_lookup()
+	 */
+	cv_broadcast(&spa_namespace_cv);
 	mutex_exit(&spa_namespace_lock);
 	return (0);
 
 fail:
+	mutex_enter(&spa_namespace_lock);
 	spa->spa_is_exporting = B_FALSE;
+	spa->spa_export_thread = NULL;
 	spa_async_resume(spa);
+
+	/* Wake up any waiters on spa_namespace_lock */
+	cv_broadcast(&spa_namespace_cv);
 	mutex_exit(&spa_namespace_lock);
 	return (error);
 }

module/zfs/spa_misc.c

@@ -27,7 +27,7 @@
  * Copyright (c) 2017 Datto Inc.
  * Copyright (c) 2017, Intel Corporation.
  * Copyright (c) 2019, loli10K <[email protected]>. All rights reserved.
- * Copyright (c) 2023, Klara Inc.
+ * Copyright (c) 2023, 2024, Klara Inc.
  */
 
 #include <sys/zfs_context.h>
@@ -82,8 +82,8 @@
  *		- Check if spa_refcount is zero
  *		- Rename a spa_t
  *		- add/remove/attach/detach devices
- *		- Held for the duration of create/destroy/export
- *		- Held at the start and end of import
+ *		- Held for the duration of create/destroy
+ *		- Held at the start and end of import and export
  *
  *	It does not need to handle recursion.  A create or destroy may
  *	reference objects (files or zvols) in other pools, but by
@@ -635,8 +635,14 @@ spa_lookup(const char *name)
 	if (spa == NULL)
 		return (NULL);
 
-	if (spa->spa_load_thread != NULL &&
-	    spa->spa_load_thread != curthread) {
+	/*
+	 * Avoid racing with import/export, which don't hold the namespace
+	 * lock for their entire duration.
+	 */
+	if ((spa->spa_load_thread != NULL &&
+	    spa->spa_load_thread != curthread) ||
+	    (spa->spa_export_thread != NULL &&
+	    spa->spa_export_thread != curthread)) {
 		cv_wait(&spa_namespace_cv, &spa_namespace_lock);
 		goto retry;
 	}
@@ -937,14 +943,15 @@ spa_open_ref(spa_t *spa, const void *tag)
 
 /*
  * Remove a reference to the given spa_t.  Must have at least one reference, or
- * have the namespace lock held.
+ * have the namespace lock held or be part of a pool import/export.
  */
 void
 spa_close(spa_t *spa, const void *tag)
 {
 	ASSERT(zfs_refcount_count(&spa->spa_refcount) > spa->spa_minref ||
 	    MUTEX_HELD(&spa_namespace_lock) ||
-	    spa->spa_load_thread == curthread);
+	    spa->spa_load_thread == curthread ||
+	    spa->spa_export_thread == curthread);
 	(void) zfs_refcount_remove(&spa->spa_refcount, tag);
 }
 
@@ -964,13 +971,15 @@ spa_async_close(spa_t *spa, const void *tag)
 
 /*
  * Check to see if the spa refcount is zero.  Must be called with
- * spa_namespace_lock held.  We really compare against spa_minref, which is the
- * number of references acquired when opening a pool
+ * spa_namespace_lock held or be the spa export thread.  We really
+ * compare against spa_minref, which is the  number of references
+ * acquired when opening a pool
  */
 boolean_t
 spa_refcount_zero(spa_t *spa)
 {
-	ASSERT(MUTEX_HELD(&spa_namespace_lock));
+	ASSERT(MUTEX_HELD(&spa_namespace_lock) ||
+	    spa->spa_export_thread == curthread);
 
 	return (zfs_refcount_count(&spa->spa_refcount) == spa->spa_minref);
 }
@@ -1218,6 +1227,21 @@ spa_vdev_enter(spa_t *spa)
 	mutex_enter(&spa->spa_vdev_top_lock);
 	mutex_enter(&spa_namespace_lock);
 
+	/*
+	 * We have a reference on the spa and a spa export could be
+	 * starting but no longer holding the spa_namespace_lock. So
+	 * check if there is an export and if so wait. It will fail
+	 * fast (EBUSY) since we are still holding a spa reference.
+	 *
+	 * Note that we can be woken by a different spa transitioning
+	 * through an import/export, so we must wait for our condition
+	 * to change before proceeding.
+	 */
+	while (spa->spa_export_thread != NULL &&
+	    spa->spa_export_thread != curthread) {
+		cv_wait(&spa_namespace_cv, &spa_namespace_lock);
+	}
+
 	vdev_autotrim_stop_all(spa);
 
 	return (spa_vdev_config_enter(spa));
@@ -1235,6 +1259,12 @@ spa_vdev_detach_enter(spa_t *spa, uint64_t guid)
 	mutex_enter(&spa->spa_vdev_top_lock);
 	mutex_enter(&spa_namespace_lock);
 
+	/* See comment in spa_vdev_enter() */
+	while (spa->spa_export_thread != NULL &&
+	    spa->spa_export_thread != curthread) {
+		cv_wait(&spa_namespace_cv, &spa_namespace_lock);
+	}
+
 	vdev_autotrim_stop_all(spa);
 
 	if (guid != 0) {

module/zfs/vdev_initialize.c

@@ -682,7 +682,8 @@ vdev_initialize_stop_wait(spa_t *spa, list_t *vd_list)
 	(void) spa;
 	vdev_t *vd;
 
-	ASSERT(MUTEX_HELD(&spa_namespace_lock));
+	ASSERT(MUTEX_HELD(&spa_namespace_lock) ||
+	    spa->spa_export_thread == curthread);
 
 	while ((vd = list_remove_head(vd_list)) != NULL) {
 		mutex_enter(&vd->vdev_initialize_lock);
@@ -724,7 +725,8 @@ vdev_initialize_stop(vdev_t *vd, vdev_initializing_state_t tgt_state,
 	if (vd_list == NULL) {
 		vdev_initialize_stop_wait_impl(vd);
 	} else {
-		ASSERT(MUTEX_HELD(&spa_namespace_lock));
+		ASSERT(MUTEX_HELD(&spa_namespace_lock) ||
+		    vd->vdev_spa->spa_export_thread == curthread);
 		list_insert_tail(vd_list, vd);
 	}
 }
@@ -756,7 +758,8 @@ vdev_initialize_stop_all(vdev_t *vd, vdev_initializing_state_t tgt_state)
 	spa_t *spa = vd->vdev_spa;
 	list_t vd_list;
 
-	ASSERT(MUTEX_HELD(&spa_namespace_lock));
+	ASSERT(MUTEX_HELD(&spa_namespace_lock) ||
+	    spa->spa_export_thread == curthread);
 
 	list_create(&vd_list, sizeof (vdev_t),
 	    offsetof(vdev_t, vdev_initialize_node));

module/zfs/vdev_rebuild.c

@@ -1087,7 +1087,8 @@ vdev_rebuild_stop_wait(vdev_t *vd)
 {
 	spa_t *spa = vd->vdev_spa;
 
-	ASSERT(MUTEX_HELD(&spa_namespace_lock));
+	ASSERT(MUTEX_HELD(&spa_namespace_lock) ||
+	    spa->spa_export_thread == curthread);
 
 	if (vd == spa->spa_root_vdev) {
 		for (uint64_t i = 0; i < vd->vdev_children; i++)