8115 parallel zfs mount

Mounting of filesystems in "filesystem/local" is done using `zfs mount -a`,
which mounts each filesystems serially. The bottleneck for each mount is
the I/O done to load metadata for each filesystem. As such, mounting
filesystems using a parallel algorithm should be a big win, and bring down
the runtime of "filesystem/local"'s start method.

illumos/illumos-gate@591e0e133f9980083db5d64ac33a30bcc3382ff7

Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Pavel Zakharov <pavel.zakharov@delphix.com>
Reviewed by: Brad Lewis <brad.lewis@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Paul Dagnelie <pcd@delphix.com>
Reviewed by: Prashanth Sreenivasa <pks@delphix.com>
Approved by: Matt Ahrens <mahrens@delphix.com>
Author:     Sebastien Roy <seb@delphix.com>

6 files changed, 725 insertions, 90 deletions

diff --git a/lib/libzfs/common/libzfs.h b/lib/libzfs/common/libzfs.h
index 9dc2b02e147f..d296fed59f01 100644
--- a/lib/libzfs/common/libzfs.h
+++ b/lib/libzfs/common/libzfs.h
@@ -576,12 +576,11 @@ typedef struct get_all_cb {
 	zfs_handle_t	**cb_handles;
 	size_t		cb_alloc;
 	size_t		cb_used;
-	boolean_t	cb_verbose;
-	int		(*cb_getone)(zfs_handle_t *, void *);
 } get_all_cb_t;
 
+void zfs_foreach_mountpoint(libzfs_handle_t *, zfs_handle_t **, size_t,
+    zfs_iter_f, void *, boolean_t);
 void libzfs_add_handle(get_all_cb_t *, zfs_handle_t *);
-int libzfs_dataset_cmp(const void *, const void *);
 
 /*
  * Functions to create and destroy datasets.
diff --git a/lib/libzfs/common/libzfs_dataset.c b/lib/libzfs/common/libzfs_dataset.c
index 011c2653a152..556538a4402b 100644
--- a/lib/libzfs/common/libzfs_dataset.c
+++ b/lib/libzfs/common/libzfs_dataset.c
@@ -54,6 +54,7 @@
 #include <idmap.h>
 #include <aclutils.h>
 #include <directory.h>
+#include <time.h>
 
 #include <sys/dnode.h>
 #include <sys/spa.h>
@@ -785,6 +786,8 @@ libzfs_mnttab_cache_compare(const void *arg1, const void *arg2)
 void
 libzfs_mnttab_init(libzfs_handle_t *hdl)
 {
+	(void) mutex_init(&hdl->libzfs_mnttab_cache_lock,
+	    LOCK_NORMAL | LOCK_ERRORCHECK, NULL);
 	assert(avl_numnodes(&hdl->libzfs_mnttab_cache) == 0);
 	avl_create(&hdl->libzfs_mnttab_cache, libzfs_mnttab_cache_compare,
 	    sizeof (mnttab_node_t), offsetof(mnttab_node_t, mtn_node));
@@ -825,6 +828,7 @@ libzfs_mnttab_fini(libzfs_handle_t *hdl)
 		free(mtn);
 	}
 	avl_destroy(&hdl->libzfs_mnttab_cache);
+	(void) mutex_destroy(&hdl->libzfs_mnttab_cache_lock);
 }
 
 void
@@ -839,6 +843,7 @@ libzfs_mnttab_find(libzfs_handle_t *hdl, const char *fsname,
 {
 	mnttab_node_t find;
 	mnttab_node_t *mtn;
+	int ret = ENOENT;
 
 	if (!hdl->libzfs_mnttab_enable) {
 		struct mnttab srch = { 0 };
@@ -854,6 +859,7 @@ libzfs_mnttab_find(libzfs_handle_t *hdl, const char *fsname,
 			return (ENOENT);
 	}
 
+	mutex_enter(&hdl->libzfs_mnttab_cache_lock);
 	if (avl_numnodes(&hdl->libzfs_mnttab_cache) == 0)
 		libzfs_mnttab_update(hdl);
 
@@ -861,9 +867,10 @@ libzfs_mnttab_find(libzfs_handle_t *hdl, const char *fsname,
 	mtn = avl_find(&hdl->libzfs_mnttab_cache, &find, NULL);
 	if (mtn) {
 		*entry = mtn->mtn_mt;
-		return (0);
+		ret = 0;
 	}
-	return (ENOENT);
+	mutex_exit(&hdl->libzfs_mnttab_cache_lock);
+	return (ret);
 }
 
 void
@@ -872,14 +879,16 @@ libzfs_mnttab_add(libzfs_handle_t *hdl, const char *special,
 {
 	mnttab_node_t *mtn;
 
-	if (avl_numnodes(&hdl->libzfs_mnttab_cache) == 0)
-		return;
-	mtn = zfs_alloc(hdl, sizeof (mnttab_node_t));
-	mtn->mtn_mt.mnt_special = zfs_strdup(hdl, special);
-	mtn->mtn_mt.mnt_mountp = zfs_strdup(hdl, mountp);
-	mtn->mtn_mt.mnt_fstype = zfs_strdup(hdl, MNTTYPE_ZFS);
-	mtn->mtn_mt.mnt_mntopts = zfs_strdup(hdl, mntopts);
-	avl_add(&hdl->libzfs_mnttab_cache, mtn);
+	mutex_enter(&hdl->libzfs_mnttab_cache_lock);
+	if (avl_numnodes(&hdl->libzfs_mnttab_cache) != 0) {
+		mtn = zfs_alloc(hdl, sizeof (mnttab_node_t));
+		mtn->mtn_mt.mnt_special = zfs_strdup(hdl, special);
+		mtn->mtn_mt.mnt_mountp = zfs_strdup(hdl, mountp);
+		mtn->mtn_mt.mnt_fstype = zfs_strdup(hdl, MNTTYPE_ZFS);
+		mtn->mtn_mt.mnt_mntopts = zfs_strdup(hdl, mntopts);
+		avl_add(&hdl->libzfs_mnttab_cache, mtn);
+	}
+	mutex_exit(&hdl->libzfs_mnttab_cache_lock);
 }
 
 void
@@ -888,6 +897,7 @@ libzfs_mnttab_remove(libzfs_handle_t *hdl, const char *fsname)
 	mnttab_node_t find;
 	mnttab_node_t *ret;
 
+	mutex_enter(&hdl->libzfs_mnttab_cache_lock);
 	find.mtn_mt.mnt_special = (char *)fsname;
 	if ((ret = avl_find(&hdl->libzfs_mnttab_cache, (void *)&find, NULL))
 	    != NULL) {
@@ -898,6 +908,7 @@ libzfs_mnttab_remove(libzfs_handle_t *hdl, const char *fsname)
 		free(ret->mtn_mt.mnt_mntopts);
 		free(ret);
 	}
+	mutex_exit(&hdl->libzfs_mnttab_cache_lock);
 }
 
 int
diff --git a/lib/libzfs/common/libzfs_impl.h b/lib/libzfs/common/libzfs_impl.h
index 50f48fd7932d..cd9a53d91fa8 100644
--- a/lib/libzfs/common/libzfs_impl.h
+++ b/lib/libzfs/common/libzfs_impl.h
@@ -22,7 +22,7 @@
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  * Copyright (c) 2011 Pawel Jakub Dawidek. All rights reserved.
- * Copyright (c) 2011, 2016 by Delphix. All rights reserved.
+ * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
  */
 
 #ifndef	_LIBZFS_IMPL_H
@@ -33,6 +33,7 @@
 #include <sys/nvpair.h>
 #include <sys/dmu.h>
 #include <sys/zfs_ioctl.h>
+#include <synch.h>
 
 #include <libuutil.h>
 #include <libzfs.h>
@@ -73,6 +74,13 @@ struct libzfs_handle {
 	int libzfs_storeerr; /* stuff error messages into buffer */
 	void *libzfs_sharehdl; /* libshare handle */
 	boolean_t libzfs_mnttab_enable;
+	/*
+	 * We need a lock to handle the case where parallel mount
+	 * threads are populating the mnttab cache simultaneously. The
+	 * lock only protects the integrity of the avl tree, and does
+	 * not protect the contents of the mnttab entries themselves.
+	 */
+	mutex_t libzfs_mnttab_cache_lock;
 	avl_tree_t libzfs_mnttab_cache;
 	int libzfs_pool_iter;
 	topo_hdl_t *libzfs_topo_hdl;
diff --git a/lib/libzfs/common/libzfs_mount.c b/lib/libzfs/common/libzfs_mount.c
index 9fd37825a31c..cf15735f3f9c 100644
--- a/lib/libzfs/common/libzfs_mount.c
+++ b/lib/libzfs/common/libzfs_mount.c
@@ -22,7 +22,7 @@
 /*
  * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
- * Copyright (c) 2014, 2016 by Delphix. All rights reserved.
+ * Copyright (c) 2014, 2017 by Delphix. All rights reserved.
  * Copyright 2016 Igor Kozhukhov <ikozhukhov@gmail.com>
  * Copyright 2017 Joyent, Inc.
  * Copyright 2017 RackTop Systems.
@@ -34,25 +34,25 @@
  * they are used by mount and unmount and when changing a filesystem's
  * mountpoint.
  *
- * 	zfs_is_mounted()
- * 	zfs_mount()
- * 	zfs_unmount()
- * 	zfs_unmountall()
+ *	zfs_is_mounted()
+ *	zfs_mount()
+ *	zfs_unmount()
+ *	zfs_unmountall()
  *
  * This file also contains the functions used to manage sharing filesystems via
  * NFS and iSCSI:
  *
- * 	zfs_is_shared()
- * 	zfs_share()
- * 	zfs_unshare()
+ *	zfs_is_shared()
+ *	zfs_share()
+ *	zfs_unshare()
  *
- * 	zfs_is_shared_nfs()
- * 	zfs_is_shared_smb()
- * 	zfs_share_proto()
- * 	zfs_shareall();
- * 	zfs_unshare_nfs()
- * 	zfs_unshare_smb()
- * 	zfs_unshareall_nfs()
+ *	zfs_is_shared_nfs()
+ *	zfs_is_shared_smb()
+ *	zfs_share_proto()
+ *	zfs_shareall();
+ *	zfs_unshare_nfs()
+ *	zfs_unshare_smb()
+ *	zfs_unshareall_nfs()
  *	zfs_unshareall_smb()
  *	zfs_unshareall()
  *	zfs_unshareall_bypath()
@@ -60,8 +60,8 @@
  * The following functions are available for pool consumers, and will
  * mount/unmount and share/unshare all datasets within pool:
  *
- * 	zpool_enable_datasets()
- * 	zpool_disable_datasets()
+ *	zpool_enable_datasets()
+ *	zpool_disable_datasets()
  */
 
 #include <dirent.h>
@@ -83,11 +83,15 @@
 #include <libzfs.h>
 
 #include "libzfs_impl.h"
+#include "libzfs_taskq.h"
 
 #include <libshare.h>
 #include <sys/systeminfo.h>
 #define	MAXISALEN	257	/* based on sysinfo(2) man page */
 
+static int mount_tq_nthr = 512;	/* taskq threads for multi-threaded mounting */
+
+static void zfs_mount_task(void *);
 static int zfs_share_proto(zfs_handle_t *, zfs_share_proto_t *);
 zfs_share_type_t zfs_is_shared_proto(zfs_handle_t *, char **,
     zfs_share_proto_t);
@@ -1077,25 +1081,32 @@ remove_mountpoint(zfs_handle_t *zhp)
 	}
 }
 
+/*
+ * Add the given zfs handle to the cb_handles array, dynamically reallocating
+ * the array if it is out of space.
+ */
 void
 libzfs_add_handle(get_all_cb_t *cbp, zfs_handle_t *zhp)
 {
 	if (cbp->cb_alloc == cbp->cb_used) {
 		size_t newsz;
-		void *ptr;
+		zfs_handle_t **newhandles;
 
-		newsz = cbp->cb_alloc ? cbp->cb_alloc * 2 : 64;
-		ptr = zfs_realloc(zhp->zfs_hdl,
-		    cbp->cb_handles, cbp->cb_alloc * sizeof (void *),
-		    newsz * sizeof (void *));
-		cbp->cb_handles = ptr;
+		newsz = cbp->cb_alloc != 0 ? cbp->cb_alloc * 2 : 64;
+		newhandles = zfs_realloc(zhp->zfs_hdl,
+		    cbp->cb_handles, cbp->cb_alloc * sizeof (zfs_handle_t *),
+		    newsz * sizeof (zfs_handle_t *));
+		cbp->cb_handles = newhandles;
 		cbp->cb_alloc = newsz;
 	}
 	cbp->cb_handles[cbp->cb_used++] = zhp;
 }
 
+/*
+ * Recursive helper function used during file system enumeration
+ */
 static int
-mount_cb(zfs_handle_t *zhp, void *data)
+zfs_iter_cb(zfs_handle_t *zhp, void *data)
 {
 	get_all_cb_t *cbp = data;
 
@@ -1121,104 +1132,350 @@ mount_cb(zfs_handle_t *zhp, void *data)
 	}
 
 	libzfs_add_handle(cbp, zhp);
-	if (zfs_iter_filesystems(zhp, mount_cb, cbp) != 0) {
+	if (zfs_iter_filesystems(zhp, zfs_iter_cb, cbp) != 0) {
 		zfs_close(zhp);
 		return (-1);
 	}
 	return (0);
 }
 
+/*
+ * Sort comparator that compares two mountpoint paths. We sort these paths so
+ * that subdirectories immediately follow their parents. This means that we
+ * effectively treat the '/' character as the lowest value non-nul char. An
+ * example sorted list using this comparator would look like:
+ *
+ * /foo
+ * /foo/bar
+ * /foo/bar/baz
+ * /foo/baz
+ * /foo.bar
+ *
+ * The mounting code depends on this ordering to deterministically iterate
+ * over filesystems in order to spawn parallel mount tasks.
+ */
 int
-libzfs_dataset_cmp(const void *a, const void *b)
+mountpoint_cmp(const void *arga, const void *argb)
 {
-	zfs_handle_t **za = (zfs_handle_t **)a;
-	zfs_handle_t **zb = (zfs_handle_t **)b;
+	zfs_handle_t *const *zap = arga;
+	zfs_handle_t *za = *zap;
+	zfs_handle_t *const *zbp = argb;
+	zfs_handle_t *zb = *zbp;
 	char mounta[MAXPATHLEN];
 	char mountb[MAXPATHLEN];
+	const char *a = mounta;
+	const char *b = mountb;
 	boolean_t gota, gotb;
 
-	if ((gota = (zfs_get_type(*za) == ZFS_TYPE_FILESYSTEM)) != 0)
-		verify(zfs_prop_get(*za, ZFS_PROP_MOUNTPOINT, mounta,
+	gota = (zfs_get_type(za) == ZFS_TYPE_FILESYSTEM);
+	if (gota) {
+		verify(zfs_prop_get(za, ZFS_PROP_MOUNTPOINT, mounta,
 		    sizeof (mounta), NULL, NULL, 0, B_FALSE) == 0);
-	if ((gotb = (zfs_get_type(*zb) == ZFS_TYPE_FILESYSTEM)) != 0)
-		verify(zfs_prop_get(*zb, ZFS_PROP_MOUNTPOINT, mountb,
+	}
+	gotb = (zfs_get_type(zb) == ZFS_TYPE_FILESYSTEM);
+	if (gotb) {
+		verify(zfs_prop_get(zb, ZFS_PROP_MOUNTPOINT, mountb,
 		    sizeof (mountb), NULL, NULL, 0, B_FALSE) == 0);
+	}
 
-	if (gota && gotb)
-		return (strcmp(mounta, mountb));
+	if (gota && gotb) {
+		while (*a != '\0' && (*a == *b)) {
+			a++;
+			b++;
+		}
+		if (*a == *b)
+			return (0);
+		if (*a == '\0')
+			return (-1);
+		if (*b == '\0')
+			return (1);
+		if (*a == '/')
+			return (-1);
+		if (*b == '/')
+			return (1);
+		return (*a < *b ? -1 : *a > *b);
+	}
 
 	if (gota)
 		return (-1);
 	if (gotb)
 		return (1);
 
-	return (strcmp(zfs_get_name(a), zfs_get_name(b)));
+	/*
+	 * If neither filesystem has a mountpoint, revert to sorting by
+	 * dataset name.
+	 */
+	return (strcmp(zfs_get_name(za), zfs_get_name(zb)));
+}
+
+/*
+ * Return true if path2 is a child of path1.
+ */
+static boolean_t
+libzfs_path_contains(const char *path1, const char *path2)
+{
+	return (strstr(path2, path1) == path2 && path2[strlen(path1)] == '/');
+}
+
+/*
+ * Given a mountpoint specified by idx in the handles array, find the first
+ * non-descendent of that mountpoint and return its index. Descendant paths
+ * start with the parent's path. This function relies on the ordering
+ * enforced by mountpoint_cmp().
+ */
+static int
+non_descendant_idx(zfs_handle_t **handles, size_t num_handles, int idx)
+{
+	char parent[ZFS_MAXPROPLEN];
+	char child[ZFS_MAXPROPLEN];
+	int i;
+
+	verify(zfs_prop_get(handles[idx], ZFS_PROP_MOUNTPOINT, parent,
+	    sizeof (parent), NULL, NULL, 0, B_FALSE) == 0);
+
+	for (i = idx + 1; i < num_handles; i++) {
+		verify(zfs_prop_get(handles[i], ZFS_PROP_MOUNTPOINT, child,
+		    sizeof (child), NULL, NULL, 0, B_FALSE) == 0);
+		if (!libzfs_path_contains(parent, child))
+			break;
+	}
+	return (i);
+}
+
+typedef struct mnt_param {
+	libzfs_handle_t	*mnt_hdl;
+	zfs_taskq_t	*mnt_tq;
+	zfs_handle_t	**mnt_zhps; /* filesystems to mount */
+	size_t		mnt_num_handles;
+	int		mnt_idx;	/* Index of selected entry to mount */
+	zfs_iter_f	mnt_func;
+	void		*mnt_data;
+} mnt_param_t;
+
+/*
+ * Allocate and populate the parameter struct for mount function, and
+ * schedule mounting of the entry selected by idx.
+ */
+static void
+zfs_dispatch_mount(libzfs_handle_t *hdl, zfs_handle_t **handles,
+    size_t num_handles, int idx, zfs_iter_f func, void *data, zfs_taskq_t *tq)
+{
+	mnt_param_t *mnt_param = zfs_alloc(hdl, sizeof (mnt_param_t));
+
+	mnt_param->mnt_hdl = hdl;
+	mnt_param->mnt_tq = tq;
+	mnt_param->mnt_zhps = handles;
+	mnt_param->mnt_num_handles = num_handles;
+	mnt_param->mnt_idx = idx;
+	mnt_param->mnt_func = func;
+	mnt_param->mnt_data = data;
+
+	(void) zfs_taskq_dispatch(tq, zfs_mount_task, (void*)mnt_param,
+	    ZFS_TQ_SLEEP);
+}
+
+/*
+ * This is the structure used to keep state of mounting or sharing operations
+ * during a call to zpool_enable_datasets().
+ */
+typedef struct mount_state {
+	/*
+	 * ms_mntstatus is set to -1 if any mount fails. While multiple threads
+	 * could update this variable concurrently, no synchronization is
+	 * needed as it's only ever set to -1.
+	 */
+	int		ms_mntstatus;
+	int		ms_mntflags;
+	const char	*ms_mntopts;
+} mount_state_t;
+
+static int
+zfs_mount_one(zfs_handle_t *zhp, void *arg)
+{
+	mount_state_t *ms = arg;
+	int ret = 0;
+
+	if (zfs_mount(zhp, ms->ms_mntopts, ms->ms_mntflags) != 0)
+		ret = ms->ms_mntstatus = -1;
+	return (ret);
+}
+
+static int
+zfs_share_one(zfs_handle_t *zhp, void *arg)
+{
+	mount_state_t *ms = arg;
+	int ret = 0;
+
+	if (zfs_share(zhp) != 0)
+		ret = ms->ms_mntstatus = -1;
+	return (ret);
+}
+
+/*
+ * Task queue function to mount one file system. On completion, it finds and
+ * schedules its children to be mounted. This depends on the sorting done in
+ * zfs_foreach_mountpoint(). Note that the degenerate case (chain of entries
+ * each descending from the previous) will have no parallelism since we always
+ * have to wait for the parent to finish mounting before we can schedule
+ * its children.
+ */
+static void
+zfs_mount_task(void *arg)
+{
+	mnt_param_t *mp = arg;
+	int idx = mp->mnt_idx;
+	zfs_handle_t **handles = mp->mnt_zhps;
+	size_t num_handles = mp->mnt_num_handles;
+	char mountpoint[ZFS_MAXPROPLEN];
+
+	verify(zfs_prop_get(handles[idx], ZFS_PROP_MOUNTPOINT, mountpoint,
+	    sizeof (mountpoint), NULL, NULL, 0, B_FALSE) == 0);
+
+	if (mp->mnt_func(handles[idx], mp->mnt_data) != 0)
+		return;
+
+	/*
+	 * We dispatch tasks to mount filesystems with mountpoints underneath
+	 * this one. We do this by dispatching the next filesystem with a
+	 * descendant mountpoint of the one we just mounted, then skip all of
+	 * its descendants, dispatch the next descendant mountpoint, and so on.
+	 * The non_descendant_idx() function skips over filesystems that are
+	 * descendants of the filesystem we just dispatched.
+	 */
+	for (int i = idx + 1; i < num_handles;
+	    i = non_descendant_idx(handles, num_handles, i)) {
+		char child[ZFS_MAXPROPLEN];
+		verify(zfs_prop_get(handles[i], ZFS_PROP_MOUNTPOINT,
+		    child, sizeof (child), NULL, NULL, 0, B_FALSE) == 0);
+
+		if (!libzfs_path_contains(mountpoint, child))
+			break; /* not a descendant, return */
+		zfs_dispatch_mount(mp->mnt_hdl, handles, num_handles, i,
+		    mp->mnt_func, mp->mnt_data, mp->mnt_tq);
+	}
+	free(mp);
+}
+
+/*
+ * Issue the func callback for each ZFS handle contained in the handles
+ * array. This function is used to mount all datasets, and so this function
+ * guarantees that filesystems for parent mountpoints are called before their
+ * children. As such, before issuing any callbacks, we first sort the array
+ * of handles by mountpoint.
+ *
+ * Callbacks are issued in one of two ways:
+ *
+ * 1. Sequentially: If the parallel argument is B_FALSE or the ZFS_SERIAL_MOUNT
+ *    environment variable is set, then we issue callbacks sequentially.
+ *
+ * 2. In parallel: If the parallel argument is B_TRUE and the ZFS_SERIAL_MOUNT
+ *    environment variable is not set, then we use a taskq to dispatch threads
+ *    to mount filesystems is parallel. This function dispatches tasks to mount
+ *    the filesystems at the top-level mountpoints, and these tasks in turn
+ *    are responsible for recursively mounting filesystems in their children
+ *    mountpoints.
+ */
+void
+zfs_foreach_mountpoint(libzfs_handle_t *hdl, zfs_handle_t **handles,
+    size_t num_handles, zfs_iter_f func, void *data, boolean_t parallel)
+{
+	/*
+	 * The ZFS_SERIAL_MOUNT environment variable is an undocumented
+	 * variable that can be used as a convenience to do a/b comparison
+	 * of serial vs. parallel mounting.
+	 */
+	boolean_t serial_mount = !parallel ||
+	    (getenv("ZFS_SERIAL_MOUNT") != NULL);
+
+	/*
+	 * Sort the datasets by mountpoint. See mountpoint_cmp for details
+	 * of how these are sorted.
+	 */
+	qsort(handles, num_handles, sizeof (zfs_handle_t *), mountpoint_cmp);
+
+	if (serial_mount) {
+		for (int i = 0; i < num_handles; i++) {
+			func(handles[i], data);
+		}
+		return;
+	}
+
+	/*
+	 * Issue the callback function for each dataset using a parallel
+	 * algorithm that uses a taskq to manage threads.
+	 */
+	zfs_taskq_t *tq = zfs_taskq_create("mount_taskq", mount_tq_nthr, 0,
+	    mount_tq_nthr, mount_tq_nthr, ZFS_TASKQ_PREPOPULATE);
+
+	/*
+	 * There may be multiple "top level" mountpoints outside of the pool's
+	 * root mountpoint, e.g.: /foo /bar. Dispatch a mount task for each of
+	 * these.
+	 */
+	for (int i = 0; i < num_handles;
+	    i = non_descendant_idx(handles, num_handles, i)) {
+		zfs_dispatch_mount(hdl, handles, num_handles, i, func, data,
+		    tq);
+	}
+
+	zfs_taskq_wait(tq); /* wait for all scheduled mounts to complete */
+	zfs_taskq_destroy(tq);
 }
 
 /*
  * Mount and share all datasets within the given pool.  This assumes that no
- * datasets within the pool are currently mounted.  Because users can create
- * complicated nested hierarchies of mountpoints, we first gather all the
- * datasets and mountpoints within the pool, and sort them by mountpoint.  Once
- * we have the list of all filesystems, we iterate over them in order and mount
- * and/or share each one.
+ * datasets within the pool are currently mounted.
  */
 #pragma weak zpool_mount_datasets = zpool_enable_datasets
 int
 zpool_enable_datasets(zpool_handle_t *zhp, const char *mntopts, int flags)
 {
 	get_all_cb_t cb = { 0 };
-	libzfs_handle_t *hdl = zhp->zpool_hdl;
+	mount_state_t ms = { 0 };
 	zfs_handle_t *zfsp;
-	int i, ret = -1;
-	int *good;
+	sa_init_selective_arg_t sharearg;
+	int ret = 0;
 
-	/*
-	 * Gather all non-snap datasets within the pool.
-	 */
-	if ((zfsp = zfs_open(hdl, zhp->zpool_name, ZFS_TYPE_DATASET)) == NULL)
+	if ((zfsp = zfs_open(zhp->zpool_hdl, zhp->zpool_name,
+	    ZFS_TYPE_DATASET)) == NULL)
 		goto out;
 
-	libzfs_add_handle(&cb, zfsp);
-	if (zfs_iter_filesystems(zfsp, mount_cb, &cb) != 0)
-		goto out;
-	/*
-	 * Sort the datasets by mountpoint.
-	 */
-	qsort(cb.cb_handles, cb.cb_used, sizeof (void *),
-	    libzfs_dataset_cmp);
 
 	/*
-	 * And mount all the datasets, keeping track of which ones
-	 * succeeded or failed.
+	 * Gather all non-snapshot datasets within the pool. Start by adding
+	 * the root filesystem for this pool to the list, and then iterate
+	 * over all child filesystems.
 	 */
-	if ((good = zfs_alloc(zhp->zpool_hdl,
-	    cb.cb_used * sizeof (int))) == NULL)
+	libzfs_add_handle(&cb, zfsp);
+	if (zfs_iter_filesystems(zfsp, zfs_iter_cb, &cb) != 0)
 		goto out;
 
-	ret = 0;
-	for (i = 0; i < cb.cb_used; i++) {
-		if (zfs_mount(cb.cb_handles[i], mntopts, flags) != 0)
-			ret = -1;
-		else
-			good[i] = 1;
-	}
+	ms.ms_mntopts = mntopts;
+	ms.ms_mntflags = flags;
+	zfs_foreach_mountpoint(zhp->zpool_hdl, cb.cb_handles, cb.cb_used,
+	    zfs_mount_one, &ms, B_TRUE);
+	if (ms.ms_mntstatus != 0)
+		ret = ms.ms_mntstatus;
 
 	/*
-	 * Then share all the ones that need to be shared. This needs
-	 * to be a separate pass in order to avoid excessive reloading
-	 * of the configuration. Good should never be NULL since
-	 * zfs_alloc is supposed to exit if memory isn't available.
+	 * Share all filesystems that need to be shared. This needs to be
+	 * a separate pass because libshare is not mt-safe, and so we need
+	 * to share serially.
 	 */
-	for (i = 0; i < cb.cb_used; i++) {
-		if (good[i] && zfs_share(cb.cb_handles[i]) != 0)
-			ret = -1;
-	}
+	sharearg.zhandle_arr = cb.cb_handles;
+	sharearg.zhandle_len = cb.cb_used;
+	if ((ret = zfs_init_libshare_arg(zhp->zpool_hdl,
+	    SA_INIT_SHARE_API_SELECTIVE, &sharearg)) != 0)
+		goto out;
 
-	free(good);
+	ms.ms_mntstatus = 0;
+	zfs_foreach_mountpoint(zhp->zpool_hdl, cb.cb_handles, cb.cb_used,
+	    zfs_share_one, &ms, B_FALSE);
+	if (ms.ms_mntstatus != 0)
+		ret = ms.ms_mntstatus;
 
 out:
-	for (i = 0; i < cb.cb_used; i++)
+	for (int i = 0; i < cb.cb_used; i++)
 		zfs_close(cb.cb_handles[i]);
 	free(cb.cb_handles);
 
diff --git a/lib/libzfs/common/libzfs_taskq.c b/lib/libzfs/common/libzfs_taskq.c
new file mode 100644
index 000000000000..28bf64971084
--- /dev/null
+++ b/lib/libzfs/common/libzfs_taskq.c
@@ -0,0 +1,297 @@
+/*
+ * CDDL HEADER START
+ *
+ * The contents of this file are subject to the terms of the
+ * Common Development and Distribution License (the "License").
+ * You may not use this file except in compliance with the License.
+ *
+ * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
+ * or http://www.opensolaris.org/os/licensing.
+ * See the License for the specific language governing permissions
+ * and limitations under the License.
+ *
+ * When distributing Covered Code, include this CDDL HEADER in each
+ * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
+ * If applicable, add the following below this CDDL HEADER, with the
+ * fields enclosed by brackets "[]" replaced with your own identifying
+ * information: Portions Copyright [yyyy] [name of copyright owner]
+ *
+ * CDDL HEADER END
+ */
+/*
+ * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
+ * Use is subject to license terms.
+ */
+/*
+ * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
+ * Copyright 2012 Garrett D'Amore <garrett@damore.org>.  All rights reserved.
+ * Copyright (c) 2014, 2018 by Delphix. All rights reserved.
+ */
+
+#include <thread.h>
+#include <synch.h>
+#include <unistd.h>
+#include <string.h>
+#include <errno.h>
+#include <sys/debug.h>
+#include <sys/sysmacros.h>
+
+#include "libzfs_taskq.h"
+
+#define	ZFS_TASKQ_ACTIVE	0x00010000
+#define	ZFS_TASKQ_NAMELEN	31
+
+typedef struct zfs_taskq_ent {
+	struct zfs_taskq_ent	*ztqent_next;
+	struct zfs_taskq_ent	*ztqent_prev;
+	ztask_func_t		*ztqent_func;
+	void			*ztqent_arg;
+	uintptr_t		ztqent_flags;
+} zfs_taskq_ent_t;
+
+struct zfs_taskq {
+	char		ztq_name[ZFS_TASKQ_NAMELEN + 1];
+	mutex_t		ztq_lock;
+	rwlock_t	ztq_threadlock;
+	cond_t		ztq_dispatch_cv;
+	cond_t		ztq_wait_cv;
+	thread_t	*ztq_threadlist;
+	int		ztq_flags;
+	int		ztq_active;
+	int		ztq_nthreads;
+	int		ztq_nalloc;
+	int		ztq_minalloc;
+	int		ztq_maxalloc;
+	cond_t		ztq_maxalloc_cv;
+	int		ztq_maxalloc_wait;
+	zfs_taskq_ent_t	*ztq_freelist;
+	zfs_taskq_ent_t	ztq_task;
+};
+
+static zfs_taskq_ent_t *
+ztask_alloc(zfs_taskq_t *ztq, int ztqflags)
+{
+	zfs_taskq_ent_t *t;
+	timestruc_t ts;
+	int err;
+
+again:	if ((t = ztq->ztq_freelist) != NULL &&
+	    ztq->ztq_nalloc >= ztq->ztq_minalloc) {
+		ztq->ztq_freelist = t->ztqent_next;
+	} else {
+		if (ztq->ztq_nalloc >= ztq->ztq_maxalloc) {
+			if (!(ztqflags & UMEM_NOFAIL))
+				return (NULL);
+
+			/*
+			 * We don't want to exceed ztq_maxalloc, but we can't
+			 * wait for other tasks to complete (and thus free up
+			 * task structures) without risking deadlock with
+			 * the caller.  So, we just delay for one second
+			 * to throttle the allocation rate. If we have tasks
+			 * complete before one second timeout expires then
+			 * zfs_taskq_ent_free will signal us and we will
+			 * immediately retry the allocation.
+			 */
+			ztq->ztq_maxalloc_wait++;
+
+			ts.tv_sec = 1;
+			ts.tv_nsec = 0;
+			err = cond_reltimedwait(&ztq->ztq_maxalloc_cv,
+			    &ztq->ztq_lock, &ts);
+
+			ztq->ztq_maxalloc_wait--;
+			if (err == 0)
+				goto again;		/* signaled */
+		}
+		mutex_exit(&ztq->ztq_lock);
+
+		t = umem_alloc(sizeof (zfs_taskq_ent_t), ztqflags);
+
+		mutex_enter(&ztq->ztq_lock);
+		if (t != NULL)
+			ztq->ztq_nalloc++;
+	}
+	return (t);
+}
+
+static void
+ztask_free(zfs_taskq_t *ztq, zfs_taskq_ent_t *t)
+{
+	if (ztq->ztq_nalloc <= ztq->ztq_minalloc) {
+		t->ztqent_next = ztq->ztq_freelist;
+		ztq->ztq_freelist = t;
+	} else {
+		ztq->ztq_nalloc--;
+		mutex_exit(&ztq->ztq_lock);
+		umem_free(t, sizeof (zfs_taskq_ent_t));
+		mutex_enter(&ztq->ztq_lock);
+	}
+
+	if (ztq->ztq_maxalloc_wait)
+		VERIFY0(cond_signal(&ztq->ztq_maxalloc_cv));
+}
+
+zfs_taskqid_t
+zfs_taskq_dispatch(zfs_taskq_t *ztq, ztask_func_t func, void *arg,
+    uint_t ztqflags)
+{
+	zfs_taskq_ent_t *t;
+
+	mutex_enter(&ztq->ztq_lock);
+	ASSERT(ztq->ztq_flags & ZFS_TASKQ_ACTIVE);
+	if ((t = ztask_alloc(ztq, ztqflags)) == NULL) {
+		mutex_exit(&ztq->ztq_lock);
+		return (0);
+	}
+	if (ztqflags & ZFS_TQ_FRONT) {
+		t->ztqent_next = ztq->ztq_task.ztqent_next;
+		t->ztqent_prev = &ztq->ztq_task;
+	} else {
+		t->ztqent_next = &ztq->ztq_task;
+		t->ztqent_prev = ztq->ztq_task.ztqent_prev;
+	}
+	t->ztqent_next->ztqent_prev = t;
+	t->ztqent_prev->ztqent_next = t;
+	t->ztqent_func = func;
+	t->ztqent_arg = arg;
+	t->ztqent_flags = 0;
+	VERIFY0(cond_signal(&ztq->ztq_dispatch_cv));
+	mutex_exit(&ztq->ztq_lock);
+	return (1);
+}
+
+void
+zfs_taskq_wait(zfs_taskq_t *ztq)
+{
+	mutex_enter(&ztq->ztq_lock);
+	while (ztq->ztq_task.ztqent_next != &ztq->ztq_task ||
+	    ztq->ztq_active != 0) {
+		int ret = cond_wait(&ztq->ztq_wait_cv, &ztq->ztq_lock);
+		VERIFY(ret == 0 || ret == EINTR);
+	}
+	mutex_exit(&ztq->ztq_lock);
+}
+
+static void *
+zfs_taskq_thread(void *arg)
+{
+	zfs_taskq_t *ztq = arg;
+	zfs_taskq_ent_t *t;
+	boolean_t prealloc;
+
+	mutex_enter(&ztq->ztq_lock);
+	while (ztq->ztq_flags & ZFS_TASKQ_ACTIVE) {
+		if ((t = ztq->ztq_task.ztqent_next) == &ztq->ztq_task) {
+			int ret;
+			if (--ztq->ztq_active == 0)
+				VERIFY0(cond_broadcast(&ztq->ztq_wait_cv));
+			ret = cond_wait(&ztq->ztq_dispatch_cv, &ztq->ztq_lock);
+			VERIFY(ret == 0 || ret == EINTR);
+			ztq->ztq_active++;
+			continue;
+		}
+		t->ztqent_prev->ztqent_next = t->ztqent_next;
+		t->ztqent_next->ztqent_prev = t->ztqent_prev;
+		t->ztqent_next = NULL;
+		t->ztqent_prev = NULL;
+		prealloc = t->ztqent_flags & ZFS_TQENT_FLAG_PREALLOC;
+		mutex_exit(&ztq->ztq_lock);
+
+		VERIFY0(rw_rdlock(&ztq->ztq_threadlock));
+		t->ztqent_func(t->ztqent_arg);
+		VERIFY0(rw_unlock(&ztq->ztq_threadlock));
+
+		mutex_enter(&ztq->ztq_lock);
+		if (!prealloc)
+			ztask_free(ztq, t);
+	}
+	ztq->ztq_nthreads--;
+	VERIFY0(cond_broadcast(&ztq->ztq_wait_cv));
+	mutex_exit(&ztq->ztq_lock);
+	return (NULL);
+}
+
+/*ARGSUSED*/
+zfs_taskq_t *
+zfs_taskq_create(const char *name, int nthreads, pri_t pri, int minalloc,
+    int maxalloc, uint_t flags)
+{
+	zfs_taskq_t *ztq = umem_zalloc(sizeof (zfs_taskq_t), UMEM_NOFAIL);
+	int t;
+
+	ASSERT3S(nthreads, >=, 1);
+
+	VERIFY0(rwlock_init(&ztq->ztq_threadlock, USYNC_THREAD, NULL));
+	VERIFY0(cond_init(&ztq->ztq_dispatch_cv, USYNC_THREAD, NULL));
+	VERIFY0(cond_init(&ztq->ztq_wait_cv, USYNC_THREAD, NULL));
+	VERIFY0(cond_init(&ztq->ztq_maxalloc_cv, USYNC_THREAD, NULL));
+	VERIFY0(mutex_init(
+	    &ztq->ztq_lock, LOCK_NORMAL | LOCK_ERRORCHECK, NULL));
+
+	(void) strncpy(ztq->ztq_name, name, ZFS_TASKQ_NAMELEN + 1);
+
+	ztq->ztq_flags = flags | ZFS_TASKQ_ACTIVE;
+	ztq->ztq_active = nthreads;
+	ztq->ztq_nthreads = nthreads;
+	ztq->ztq_minalloc = minalloc;
+	ztq->ztq_maxalloc = maxalloc;
+	ztq->ztq_task.ztqent_next = &ztq->ztq_task;
+	ztq->ztq_task.ztqent_prev = &ztq->ztq_task;
+	ztq->ztq_threadlist =
+	    umem_alloc(nthreads * sizeof (thread_t), UMEM_NOFAIL);
+
+	if (flags & ZFS_TASKQ_PREPOPULATE) {
+		mutex_enter(&ztq->ztq_lock);
+		while (minalloc-- > 0)
+			ztask_free(ztq, ztask_alloc(ztq, UMEM_NOFAIL));
+		mutex_exit(&ztq->ztq_lock);
+	}
+
+	for (t = 0; t < nthreads; t++) {
+		(void) thr_create(0, 0, zfs_taskq_thread,
+		    ztq, THR_BOUND, &ztq->ztq_threadlist[t]);
+	}
+
+	return (ztq);
+}
+
+void
+zfs_taskq_destroy(zfs_taskq_t *ztq)
+{
+	int t;
+	int nthreads = ztq->ztq_nthreads;
+
+	zfs_taskq_wait(ztq);
+
+	mutex_enter(&ztq->ztq_lock);
+
+	ztq->ztq_flags &= ~ZFS_TASKQ_ACTIVE;
+	VERIFY0(cond_broadcast(&ztq->ztq_dispatch_cv));
+
+	while (ztq->ztq_nthreads != 0) {
+		int ret = cond_wait(&ztq->ztq_wait_cv, &ztq->ztq_lock);
+		VERIFY(ret == 0 || ret == EINTR);
+	}
+
+	ztq->ztq_minalloc = 0;
+	while (ztq->ztq_nalloc != 0) {
+		ASSERT(ztq->ztq_freelist != NULL);
+		ztask_free(ztq, ztask_alloc(ztq, UMEM_NOFAIL));
+	}
+
+	mutex_exit(&ztq->ztq_lock);
+
+	for (t = 0; t < nthreads; t++)
+		(void) thr_join(ztq->ztq_threadlist[t], NULL, NULL);
+
+	umem_free(ztq->ztq_threadlist, nthreads * sizeof (thread_t));
+
+	VERIFY0(rwlock_destroy(&ztq->ztq_threadlock));
+	VERIFY0(cond_destroy(&ztq->ztq_dispatch_cv));
+	VERIFY0(cond_destroy(&ztq->ztq_wait_cv));
+	VERIFY0(cond_destroy(&ztq->ztq_maxalloc_cv));
+	VERIFY0(mutex_destroy(&ztq->ztq_lock));
+
+	umem_free(ztq, sizeof (zfs_taskq_t));
+}
diff --git a/lib/libzfs/common/libzfs_taskq.h b/lib/libzfs/common/libzfs_taskq.h
new file mode 100644
index 000000000000..7ac045738c36
--- /dev/null
+++ b/lib/libzfs/common/libzfs_taskq.h
@@ -0,0 +1,63 @@
+/*
+ * CDDL HEADER START
+ *
+ * The contents of this file are subject to the terms of the
+ * Common Development and Distribution License (the "License").
+ * You may not use this file except in compliance with the License.
+ *
+ * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
+ * or http://www.opensolaris.org/os/licensing.
+ * See the License for the specific language governing permissions
+ * and limitations under the License.
+ *
+ * When distributing Covered Code, include this CDDL HEADER in each
+ * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
+ * If applicable, add the following below this CDDL HEADER, with the
+ * fields enclosed by brackets "[]" replaced with your own identifying
+ * information: Portions Copyright [yyyy] [name of copyright owner]
+ *
+ * CDDL HEADER END
+ */
+/*
+ * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
+ * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
+ * Copyright (c) 2012, Joyent, Inc. All rights reserved.
+ */
+
+#ifndef	_ZFS_TASKQ_H
+#define	_ZFS_TASKQ_H
+
+#include <stdint.h>
+#include <umem.h>
+
+#ifdef	__cplusplus
+extern "C" {
+#endif
+
+typedef struct zfs_taskq zfs_taskq_t;
+typedef uintptr_t zfs_taskqid_t;
+typedef void (ztask_func_t)(void *);
+
+#define	ZFS_TQENT_FLAG_PREALLOC		0x1	/* taskq_dispatch_ent used */
+
+#define	ZFS_TASKQ_PREPOPULATE		0x0001
+
+#define	ZFS_TQ_SLEEP	UMEM_NOFAIL	/* Can block for memory */
+#define	ZFS_TQ_NOSLEEP	UMEM_DEFAULT	/* cannot block for memory; may fail */
+#define	ZFS_TQ_FRONT	0x08		/* Queue in front */
+
+extern zfs_taskq_t	*zfs_taskq_create(const char *, int, pri_t, int,
+	int, uint_t);
+extern void		zfs_taskq_destroy(zfs_taskq_t *);
+
+extern zfs_taskqid_t	zfs_taskq_dispatch(zfs_taskq_t *, ztask_func_t,
+	void *, uint_t);
+
+extern void		zfs_taskq_wait(zfs_taskq_t *);
+
+#ifdef	__cplusplus
+}
+#endif
+
+#endif	/* _ZFS_TASKQ_H */