X-Git-Url: https://git.camperquake.de/gitweb.cgi?a=blobdiff_plain;f=module%2Fzfs%2Fmetaslab.c;h=cd1b6ce730f4843bbe5d41f9f4b24f39f4aa48f5;hb=refs%2Fheads%2Frertzinger%2Ffeature-zpool-get--p;hp=77556ac5d71cd718ef918343a08851c4cdf85495;hpb=9babb37438b58e77bad04e820d5702e15b79e6a6;p=zfs.git

diff --git a/module/zfs/metaslab.c b/module/zfs/metaslab.c
index 77556ac..cd1b6ce 100644
--- a/module/zfs/metaslab.c
+++ b/module/zfs/metaslab.c
@@ -19,12 +19,11 @@
  * CDDL HEADER END
  */
 /*
- * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
- * Use is subject to license terms.
+ * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2012 by Delphix. All rights reserved.
  */
 
 #include <sys/zfs_context.h>
-#include <sys/spa_impl.h>
 #include <sys/dmu.h>
 #include <sys/dmu_tx.h>
 #include <sys/space_map.h>
@@ -32,12 +31,46 @@
 #include <sys/vdev_impl.h>
 #include <sys/zio.h>
 
+#define WITH_DF_BLOCK_ALLOCATOR
+
+/*
+ * Allow allocations to switch to gang blocks quickly. We do this to
+ * avoid having to load lots of space_maps in a given txg. There are,
+ * however, some cases where we want to avoid "fast" ganging and instead
+ * we want to do an exhaustive search of all metaslabs on this device.
+ * Currently we don't allow any gang, zil, or dump device related allocations
+ * to "fast" gang.
+ */
+#define	CAN_FASTGANG(flags) \
+	(!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
+	METASLAB_GANG_AVOID)))
+
 uint64_t metaslab_aliquot = 512ULL << 10;
 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1;	/* force gang blocks */
 
 /*
+ * The in-core space map representation is more compact than its on-disk form.
+ * The zfs_condense_pct determines how much more compact the in-core
+ * space_map representation must be before we compact it on-disk.
+ * Values should be greater than or equal to 100.
+ */
+int zfs_condense_pct = 200;
+
+/*
+ * This value defines the number of allowed allocation failures per vdev.
+ * If a device reaches this threshold in a given txg then we consider skipping
+ * allocations on that device.
+ */
+int zfs_mg_alloc_failures;
+
+/*
+ * Metaslab debugging: when set, keeps all space maps in core to verify frees.
+ */
+int metaslab_debug = 0;
+
+/*
  * Minimum size which forces the dynamic allocator to change
- * it's allocation strategy. Once the space map cannot satisfy
+ * it's allocation strategy.  Once the space map cannot satisfy
  * an allocation of this size then it switches to using more
  * aggressive strategy (i.e search by size rather than offset).
  */
@@ -49,7 +82,23 @@ uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE;
  * Once the space_map's free space drops below this level we dynamically
  * switch to using best-fit allocations.
  */
-int metaslab_df_free_pct = 30;
+int metaslab_df_free_pct = 4;
+
+/*
+ * A metaslab is considered "free" if it contains a contiguous
+ * segment which is greater than metaslab_min_alloc_size.
+ */
+uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
+
+/*
+ * Max number of space_maps to prefetch.
+ */
+int metaslab_prefetch_limit = SPA_DVAS_PER_BP;
+
+/*
+ * Percentage bonus multiplier for metaslabs that are in the bonus area.
+ */
+int metaslab_smo_bonus_pct = 150;
 
 /*
  * ==========================================================================
@@ -57,14 +106,16 @@ int metaslab_df_free_pct = 30;
  * ==========================================================================
  */
 metaslab_class_t *
-metaslab_class_create(space_map_ops_t *ops)
+metaslab_class_create(spa_t *spa, space_map_ops_t *ops)
 {
 	metaslab_class_t *mc;
 
-	mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
+	mc = kmem_zalloc(sizeof (metaslab_class_t), KM_PUSHPAGE);
 
+	mc->mc_spa = spa;
 	mc->mc_rotor = NULL;
 	mc->mc_ops = ops;
+	mutex_init(&mc->mc_fastwrite_lock, NULL, MUTEX_DEFAULT, NULL);
 
 	return (mc);
 }
@@ -72,58 +123,74 @@ metaslab_class_create(space_map_ops_t *ops)
 void
 metaslab_class_destroy(metaslab_class_t *mc)
 {
-	metaslab_group_t *mg;
-
-	while ((mg = mc->mc_rotor) != NULL) {
-		metaslab_class_remove(mc, mg);
-		metaslab_group_destroy(mg);
-	}
+	ASSERT(mc->mc_rotor == NULL);
+	ASSERT(mc->mc_alloc == 0);
+	ASSERT(mc->mc_deferred == 0);
+	ASSERT(mc->mc_space == 0);
+	ASSERT(mc->mc_dspace == 0);
 
+	mutex_destroy(&mc->mc_fastwrite_lock);
 	kmem_free(mc, sizeof (metaslab_class_t));
 }
 
-void
-metaslab_class_add(metaslab_class_t *mc, metaslab_group_t *mg)
+int
+metaslab_class_validate(metaslab_class_t *mc)
 {
-	metaslab_group_t *mgprev, *mgnext;
+	metaslab_group_t *mg;
+	vdev_t *vd;
 
-	ASSERT(mg->mg_class == NULL);
+	/*
+	 * Must hold one of the spa_config locks.
+	 */
+	ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
+	    spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
 
-	if ((mgprev = mc->mc_rotor) == NULL) {
-		mg->mg_prev = mg;
-		mg->mg_next = mg;
-	} else {
-		mgnext = mgprev->mg_next;
-		mg->mg_prev = mgprev;
-		mg->mg_next = mgnext;
-		mgprev->mg_next = mg;
-		mgnext->mg_prev = mg;
-	}
-	mc->mc_rotor = mg;
-	mg->mg_class = mc;
+	if ((mg = mc->mc_rotor) == NULL)
+		return (0);
+
+	do {
+		vd = mg->mg_vd;
+		ASSERT(vd->vdev_mg != NULL);
+		ASSERT3P(vd->vdev_top, ==, vd);
+		ASSERT3P(mg->mg_class, ==, mc);
+		ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
+	} while ((mg = mg->mg_next) != mc->mc_rotor);
+
+	return (0);
 }
 
 void
-metaslab_class_remove(metaslab_class_t *mc, metaslab_group_t *mg)
+metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
+    int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
 {
-	metaslab_group_t *mgprev, *mgnext;
+	atomic_add_64(&mc->mc_alloc, alloc_delta);
+	atomic_add_64(&mc->mc_deferred, defer_delta);
+	atomic_add_64(&mc->mc_space, space_delta);
+	atomic_add_64(&mc->mc_dspace, dspace_delta);
+}
 
-	ASSERT(mg->mg_class == mc);
+uint64_t
+metaslab_class_get_alloc(metaslab_class_t *mc)
+{
+	return (mc->mc_alloc);
+}
 
-	mgprev = mg->mg_prev;
-	mgnext = mg->mg_next;
+uint64_t
+metaslab_class_get_deferred(metaslab_class_t *mc)
+{
+	return (mc->mc_deferred);
+}
 
-	if (mg == mgnext) {
-		mc->mc_rotor = NULL;
-	} else {
-		mc->mc_rotor = mgnext;
-		mgprev->mg_next = mgnext;
-		mgnext->mg_prev = mgprev;
-	}
+uint64_t
+metaslab_class_get_space(metaslab_class_t *mc)
+{
+	return (mc->mc_space);
+}
 
-	mg->mg_prev = NULL;
-	mg->mg_next = NULL;
-	mg->mg_class = NULL;
+uint64_t
+metaslab_class_get_dspace(metaslab_class_t *mc)
+{
+	return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
 }
 
 /*
@@ -145,9 +212,9 @@ metaslab_compare(const void *x1, const void *x2)
 	/*
 	 * If the weights are identical, use the offset to force uniqueness.
 	 */
-	if (m1->ms_map.sm_start < m2->ms_map.sm_start)
+	if (m1->ms_map->sm_start < m2->ms_map->sm_start)
 		return (-1);
-	if (m1->ms_map.sm_start > m2->ms_map.sm_start)
+	if (m1->ms_map->sm_start > m2->ms_map->sm_start)
 		return (1);
 
 	ASSERT3P(m1, ==, m2);
@@ -160,13 +227,13 @@ metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
 {
 	metaslab_group_t *mg;
 
-	mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
+	mg = kmem_zalloc(sizeof (metaslab_group_t), KM_PUSHPAGE);
 	mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
 	avl_create(&mg->mg_metaslab_tree, metaslab_compare,
 	    sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
-	mg->mg_aliquot = metaslab_aliquot * MAX(1, vd->vdev_children);
 	mg->mg_vd = vd;
-	metaslab_class_add(mc, mg);
+	mg->mg_class = mc;
+	mg->mg_activation_count = 0;
 
 	return (mg);
 }
@@ -174,11 +241,82 @@ metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
 void
 metaslab_group_destroy(metaslab_group_t *mg)
 {
+	ASSERT(mg->mg_prev == NULL);
+	ASSERT(mg->mg_next == NULL);
+	/*
+	 * We may have gone below zero with the activation count
+	 * either because we never activated in the first place or
+	 * because we're done, and possibly removing the vdev.
+	 */
+	ASSERT(mg->mg_activation_count <= 0);
+
 	avl_destroy(&mg->mg_metaslab_tree);
 	mutex_destroy(&mg->mg_lock);
 	kmem_free(mg, sizeof (metaslab_group_t));
 }
 
+void
+metaslab_group_activate(metaslab_group_t *mg)
+{
+	metaslab_class_t *mc = mg->mg_class;
+	metaslab_group_t *mgprev, *mgnext;
+
+	ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
+
+	ASSERT(mc->mc_rotor != mg);
+	ASSERT(mg->mg_prev == NULL);
+	ASSERT(mg->mg_next == NULL);
+	ASSERT(mg->mg_activation_count <= 0);
+
+	if (++mg->mg_activation_count <= 0)
+		return;
+
+	mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
+
+	if ((mgprev = mc->mc_rotor) == NULL) {
+		mg->mg_prev = mg;
+		mg->mg_next = mg;
+	} else {
+		mgnext = mgprev->mg_next;
+		mg->mg_prev = mgprev;
+		mg->mg_next = mgnext;
+		mgprev->mg_next = mg;
+		mgnext->mg_prev = mg;
+	}
+	mc->mc_rotor = mg;
+}
+
+void
+metaslab_group_passivate(metaslab_group_t *mg)
+{
+	metaslab_class_t *mc = mg->mg_class;
+	metaslab_group_t *mgprev, *mgnext;
+
+	ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
+
+	if (--mg->mg_activation_count != 0) {
+		ASSERT(mc->mc_rotor != mg);
+		ASSERT(mg->mg_prev == NULL);
+		ASSERT(mg->mg_next == NULL);
+		ASSERT(mg->mg_activation_count < 0);
+		return;
+	}
+
+	mgprev = mg->mg_prev;
+	mgnext = mg->mg_next;
+
+	if (mg == mgnext) {
+		mc->mc_rotor = NULL;
+	} else {
+		mc->mc_rotor = mgnext;
+		mgprev->mg_next = mgnext;
+		mgnext->mg_prev = mgprev;
+	}
+
+	mg->mg_prev = NULL;
+	mg->mg_next = NULL;
+}
+
 static void
 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
 {
@@ -219,6 +357,35 @@ metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
 }
 
 /*
+ * ==========================================================================
+ * Common allocator routines
+ * ==========================================================================
+ */
+static int
+metaslab_segsize_compare(const void *x1, const void *x2)
+{
+	const space_seg_t *s1 = x1;
+	const space_seg_t *s2 = x2;
+	uint64_t ss_size1 = s1->ss_end - s1->ss_start;
+	uint64_t ss_size2 = s2->ss_end - s2->ss_start;
+
+	if (ss_size1 < ss_size2)
+		return (-1);
+	if (ss_size1 > ss_size2)
+		return (1);
+
+	if (s1->ss_start < s2->ss_start)
+		return (-1);
+	if (s1->ss_start > s2->ss_start)
+		return (1);
+
+	return (0);
+}
+
+#if defined(WITH_FF_BLOCK_ALLOCATOR) || \
+    defined(WITH_DF_BLOCK_ALLOCATOR) || \
+    defined(WITH_CDF_BLOCK_ALLOCATOR)
+/*
  * This is a helper function that can be used by the allocator to find
  * a suitable block to allocate. This will search the specified AVL
  * tree looking for a block that matches the specified criteria.
@@ -257,141 +424,249 @@ metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
 	*cursor = 0;
 	return (metaslab_block_picker(t, cursor, size, align));
 }
+#endif /* WITH_FF/DF/CDF_BLOCK_ALLOCATOR */
 
-/*
- * ==========================================================================
- * The first-fit block allocator
- * ==========================================================================
- */
 static void
-metaslab_ff_load(space_map_t *sm)
+metaslab_pp_load(space_map_t *sm)
 {
+	space_seg_t *ss;
+
 	ASSERT(sm->sm_ppd == NULL);
-	sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_SLEEP);
-	sm->sm_pp_root = NULL;
+	sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_PUSHPAGE);
+
+	sm->sm_pp_root = kmem_alloc(sizeof (avl_tree_t), KM_PUSHPAGE);
+	avl_create(sm->sm_pp_root, metaslab_segsize_compare,
+	    sizeof (space_seg_t), offsetof(struct space_seg, ss_pp_node));
+
+	for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
+		avl_add(sm->sm_pp_root, ss);
 }
 
 static void
-metaslab_ff_unload(space_map_t *sm)
+metaslab_pp_unload(space_map_t *sm)
 {
+	void *cookie = NULL;
+
 	kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
 	sm->sm_ppd = NULL;
-}
 
-static uint64_t
-metaslab_ff_alloc(space_map_t *sm, uint64_t size)
-{
-	avl_tree_t *t = &sm->sm_root;
-	uint64_t align = size & -size;
-	uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
+	while (avl_destroy_nodes(sm->sm_pp_root, &cookie) != NULL) {
+		/* tear down the tree */
+	}
 
-	return (metaslab_block_picker(t, cursor, size, align));
+	avl_destroy(sm->sm_pp_root);
+	kmem_free(sm->sm_pp_root, sizeof (avl_tree_t));
+	sm->sm_pp_root = NULL;
 }
 
 /* ARGSUSED */
 static void
-metaslab_ff_claim(space_map_t *sm, uint64_t start, uint64_t size)
+metaslab_pp_claim(space_map_t *sm, uint64_t start, uint64_t size)
 {
 	/* No need to update cursor */
 }
 
 /* ARGSUSED */
 static void
-metaslab_ff_free(space_map_t *sm, uint64_t start, uint64_t size)
+metaslab_pp_free(space_map_t *sm, uint64_t start, uint64_t size)
 {
 	/* No need to update cursor */
 }
 
+/*
+ * Return the maximum contiguous segment within the metaslab.
+ */
+uint64_t
+metaslab_pp_maxsize(space_map_t *sm)
+{
+	avl_tree_t *t = sm->sm_pp_root;
+	space_seg_t *ss;
+
+	if (t == NULL || (ss = avl_last(t)) == NULL)
+		return (0ULL);
+
+	return (ss->ss_end - ss->ss_start);
+}
+
+#if defined(WITH_FF_BLOCK_ALLOCATOR)
+/*
+ * ==========================================================================
+ * The first-fit block allocator
+ * ==========================================================================
+ */
+static uint64_t
+metaslab_ff_alloc(space_map_t *sm, uint64_t size)
+{
+	avl_tree_t *t = &sm->sm_root;
+	uint64_t align = size & -size;
+	uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
+
+	return (metaslab_block_picker(t, cursor, size, align));
+}
+
+/* ARGSUSED */
+boolean_t
+metaslab_ff_fragmented(space_map_t *sm)
+{
+	return (B_TRUE);
+}
+
 static space_map_ops_t metaslab_ff_ops = {
-	metaslab_ff_load,
-	metaslab_ff_unload,
+	metaslab_pp_load,
+	metaslab_pp_unload,
 	metaslab_ff_alloc,
-	metaslab_ff_claim,
-	metaslab_ff_free,
-	NULL	/* maxsize */
+	metaslab_pp_claim,
+	metaslab_pp_free,
+	metaslab_pp_maxsize,
+	metaslab_ff_fragmented
 };
 
+space_map_ops_t *zfs_metaslab_ops = &metaslab_ff_ops;
+#endif /* WITH_FF_BLOCK_ALLOCATOR */
+
+#if defined(WITH_DF_BLOCK_ALLOCATOR)
 /*
+ * ==========================================================================
  * Dynamic block allocator -
  * Uses the first fit allocation scheme until space get low and then
  * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
  * and metaslab_df_free_pct to determine when to switch the allocation scheme.
+ * ==========================================================================
  */
-
-uint64_t
-metaslab_df_maxsize(space_map_t *sm)
+static uint64_t
+metaslab_df_alloc(space_map_t *sm, uint64_t size)
 {
-	avl_tree_t *t = sm->sm_pp_root;
-	space_seg_t *ss;
+	avl_tree_t *t = &sm->sm_root;
+	uint64_t align = size & -size;
+	uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
+	uint64_t max_size = metaslab_pp_maxsize(sm);
+	int free_pct = sm->sm_space * 100 / sm->sm_size;
 
-	if (t == NULL || (ss = avl_last(t)) == NULL)
-		return (0ULL);
+	ASSERT(MUTEX_HELD(sm->sm_lock));
+	ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
 
-	return (ss->ss_end - ss->ss_start);
+	if (max_size < size)
+		return (-1ULL);
+
+	/*
+	 * If we're running low on space switch to using the size
+	 * sorted AVL tree (best-fit).
+	 */
+	if (max_size < metaslab_df_alloc_threshold ||
+	    free_pct < metaslab_df_free_pct) {
+		t = sm->sm_pp_root;
+		*cursor = 0;
+	}
+
+	return (metaslab_block_picker(t, cursor, size, 1ULL));
 }
 
-static int
-metaslab_df_seg_compare(const void *x1, const void *x2)
+static boolean_t
+metaslab_df_fragmented(space_map_t *sm)
 {
-	const space_seg_t *s1 = x1;
-	const space_seg_t *s2 = x2;
-	uint64_t ss_size1 = s1->ss_end - s1->ss_start;
-	uint64_t ss_size2 = s2->ss_end - s2->ss_start;
-
-	if (ss_size1 < ss_size2)
-		return (-1);
-	if (ss_size1 > ss_size2)
-		return (1);
+	uint64_t max_size = metaslab_pp_maxsize(sm);
+	int free_pct = sm->sm_space * 100 / sm->sm_size;
 
-	if (s1->ss_start < s2->ss_start)
-		return (-1);
-	if (s1->ss_start > s2->ss_start)
-		return (1);
+	if (max_size >= metaslab_df_alloc_threshold &&
+	    free_pct >= metaslab_df_free_pct)
+		return (B_FALSE);
 
-	return (0);
+	return (B_TRUE);
 }
 
-static void
-metaslab_df_load(space_map_t *sm)
+static space_map_ops_t metaslab_df_ops = {
+	metaslab_pp_load,
+	metaslab_pp_unload,
+	metaslab_df_alloc,
+	metaslab_pp_claim,
+	metaslab_pp_free,
+	metaslab_pp_maxsize,
+	metaslab_df_fragmented
+};
+
+space_map_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
+#endif /* WITH_DF_BLOCK_ALLOCATOR */
+
+/*
+ * ==========================================================================
+ * Other experimental allocators
+ * ==========================================================================
+ */
+#if defined(WITH_CDF_BLOCK_ALLOCATOR)
+static uint64_t
+metaslab_cdf_alloc(space_map_t *sm, uint64_t size)
 {
-	space_seg_t *ss;
+	avl_tree_t *t = &sm->sm_root;
+	uint64_t *cursor = (uint64_t *)sm->sm_ppd;
+	uint64_t *extent_end = (uint64_t *)sm->sm_ppd + 1;
+	uint64_t max_size = metaslab_pp_maxsize(sm);
+	uint64_t rsize = size;
+	uint64_t offset = 0;
 
-	ASSERT(sm->sm_ppd == NULL);
-	sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_SLEEP);
+	ASSERT(MUTEX_HELD(sm->sm_lock));
+	ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
 
-	sm->sm_pp_root = kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
-	avl_create(sm->sm_pp_root, metaslab_df_seg_compare,
-	    sizeof (space_seg_t), offsetof(struct space_seg, ss_pp_node));
+	if (max_size < size)
+		return (-1ULL);
 
-	for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
-		avl_add(sm->sm_pp_root, ss);
+	ASSERT3U(*extent_end, >=, *cursor);
+
+	/*
+	 * If we're running low on space switch to using the size
+	 * sorted AVL tree (best-fit).
+	 */
+	if ((*cursor + size) > *extent_end) {
+
+		t = sm->sm_pp_root;
+		*cursor = *extent_end = 0;
+
+		if (max_size > 2 * SPA_MAXBLOCKSIZE)
+			rsize = MIN(metaslab_min_alloc_size, max_size);
+		offset = metaslab_block_picker(t, extent_end, rsize, 1ULL);
+		if (offset != -1)
+			*cursor = offset + size;
+	} else {
+		offset = metaslab_block_picker(t, cursor, rsize, 1ULL);
+	}
+	ASSERT3U(*cursor, <=, *extent_end);
+	return (offset);
 }
 
-static void
-metaslab_df_unload(space_map_t *sm)
+static boolean_t
+metaslab_cdf_fragmented(space_map_t *sm)
 {
-	void *cookie = NULL;
+	uint64_t max_size = metaslab_pp_maxsize(sm);
 
-	kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
-	sm->sm_ppd = NULL;
+	if (max_size > (metaslab_min_alloc_size * 10))
+		return (B_FALSE);
+	return (B_TRUE);
+}
 
-	while (avl_destroy_nodes(sm->sm_pp_root, &cookie) != NULL) {
-		/* tear down the tree */
-	}
+static space_map_ops_t metaslab_cdf_ops = {
+	metaslab_pp_load,
+	metaslab_pp_unload,
+	metaslab_cdf_alloc,
+	metaslab_pp_claim,
+	metaslab_pp_free,
+	metaslab_pp_maxsize,
+	metaslab_cdf_fragmented
+};
 
-	avl_destroy(sm->sm_pp_root);
-	kmem_free(sm->sm_pp_root, sizeof (avl_tree_t));
-	sm->sm_pp_root = NULL;
-}
+space_map_ops_t *zfs_metaslab_ops = &metaslab_cdf_ops;
+#endif /* WITH_CDF_BLOCK_ALLOCATOR */
+
+#if defined(WITH_NDF_BLOCK_ALLOCATOR)
+uint64_t metaslab_ndf_clump_shift = 4;
 
 static uint64_t
-metaslab_df_alloc(space_map_t *sm, uint64_t size)
+metaslab_ndf_alloc(space_map_t *sm, uint64_t size)
 {
 	avl_tree_t *t = &sm->sm_root;
-	uint64_t align = size & -size;
-	uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
-	uint64_t max_size = metaslab_df_maxsize(sm);
-	int free_pct = sm->sm_space * 100 / sm->sm_size;
+	avl_index_t where;
+	space_seg_t *ss, ssearch;
+	uint64_t hbit = highbit(size);
+	uint64_t *cursor = (uint64_t *)sm->sm_ppd + hbit - 1;
+	uint64_t max_size = metaslab_pp_maxsize(sm);
 
 	ASSERT(MUTEX_HELD(sm->sm_lock));
 	ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
@@ -399,43 +674,54 @@ metaslab_df_alloc(space_map_t *sm, uint64_t size)
 	if (max_size < size)
 		return (-1ULL);
 
-	/*
-	 * If we're running low on space switch to using the size
-	 * sorted AVL tree (best-fit).
-	 */
-	if (max_size < metaslab_df_alloc_threshold ||
-	    free_pct < metaslab_df_free_pct) {
+	ssearch.ss_start = *cursor;
+	ssearch.ss_end = *cursor + size;
+
+	ss = avl_find(t, &ssearch, &where);
+	if (ss == NULL || (ss->ss_start + size > ss->ss_end)) {
 		t = sm->sm_pp_root;
-		*cursor = 0;
+
+		ssearch.ss_start = 0;
+		ssearch.ss_end = MIN(max_size,
+		    1ULL << (hbit + metaslab_ndf_clump_shift));
+		ss = avl_find(t, &ssearch, &where);
+		if (ss == NULL)
+			ss = avl_nearest(t, where, AVL_AFTER);
+		ASSERT(ss != NULL);
 	}
 
-	return (metaslab_block_picker(t, cursor, size, 1ULL));
+	if (ss != NULL) {
+		if (ss->ss_start + size <= ss->ss_end) {
+			*cursor = ss->ss_start + size;
+			return (ss->ss_start);
+		}
+	}
+	return (-1ULL);
 }
 
-/* ARGSUSED */
-static void
-metaslab_df_claim(space_map_t *sm, uint64_t start, uint64_t size)
+static boolean_t
+metaslab_ndf_fragmented(space_map_t *sm)
 {
-	/* No need to update cursor */
-}
+	uint64_t max_size = metaslab_pp_maxsize(sm);
 
-/* ARGSUSED */
-static void
-metaslab_df_free(space_map_t *sm, uint64_t start, uint64_t size)
-{
-	/* No need to update cursor */
+	if (max_size > (metaslab_min_alloc_size << metaslab_ndf_clump_shift))
+		return (B_FALSE);
+	return (B_TRUE);
 }
 
-static space_map_ops_t metaslab_df_ops = {
-	metaslab_df_load,
-	metaslab_df_unload,
-	metaslab_df_alloc,
-	metaslab_df_claim,
-	metaslab_df_free,
-	metaslab_df_maxsize
+
+static space_map_ops_t metaslab_ndf_ops = {
+	metaslab_pp_load,
+	metaslab_pp_unload,
+	metaslab_ndf_alloc,
+	metaslab_pp_claim,
+	metaslab_pp_free,
+	metaslab_pp_maxsize,
+	metaslab_ndf_fragmented
 };
 
-space_map_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
+space_map_ops_t *zfs_metaslab_ops = &metaslab_ndf_ops;
+#endif /* WITH_NDF_BLOCK_ALLOCATOR */
 
 /*
  * ==========================================================================
@@ -449,7 +735,7 @@ metaslab_init(metaslab_group_t *mg, space_map_obj_t *smo,
 	vdev_t *vd = mg->mg_vd;
 	metaslab_t *msp;
 
-	msp = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
+	msp = kmem_zalloc(sizeof (metaslab_t), KM_PUSHPAGE);
 	mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
 
 	msp->ms_smo_syncing = *smo;
@@ -461,11 +747,19 @@ metaslab_init(metaslab_group_t *mg, space_map_obj_t *smo,
 	 * addition of new space; and for debugging, it ensures that we'd
 	 * data fault on any attempt to use this metaslab before it's ready.
 	 */
-	space_map_create(&msp->ms_map, start, size,
+	msp->ms_map = kmem_zalloc(sizeof (space_map_t), KM_PUSHPAGE);
+	space_map_create(msp->ms_map, start, size,
 	    vd->vdev_ashift, &msp->ms_lock);
 
 	metaslab_group_add(mg, msp);
 
+	if (metaslab_debug && smo->smo_object != 0) {
+		mutex_enter(&msp->ms_lock);
+		VERIFY(space_map_load(msp->ms_map, mg->mg_class->mc_ops,
+		    SM_FREE, smo, spa_meta_objset(vd->vdev_spa)) == 0);
+		mutex_exit(&msp->ms_lock);
+	}
+
 	/*
 	 * If we're opening an existing pool (txg == 0) or creating
 	 * a new one (txg == TXG_INITIAL), all space is available now.
@@ -476,16 +770,8 @@ metaslab_init(metaslab_group_t *mg, space_map_obj_t *smo,
 		metaslab_sync_done(msp, 0);
 
 	if (txg != 0) {
-		/*
-		 * The vdev is dirty, but the metaslab isn't -- it just needs
-		 * to have metaslab_sync_done() invoked from vdev_sync_done().
-		 * [We could just dirty the metaslab, but that would cause us
-		 * to allocate a space map object for it, which is wasteful
-		 * and would mess up the locality logic in metaslab_weight().]
-		 */
-		ASSERT(TXG_CLEAN(txg) == spa_last_synced_txg(vd->vdev_spa));
 		vdev_dirty(vd, 0, NULL, txg);
-		vdev_dirty(vd, VDD_METASLAB, msp, TXG_CLEAN(txg));
+		vdev_dirty(vd, VDD_METASLAB, msp, txg);
 	}
 
 	return (msp);
@@ -497,21 +783,31 @@ metaslab_fini(metaslab_t *msp)
 	metaslab_group_t *mg = msp->ms_group;
 	int t;
 
-	vdev_space_update(mg->mg_vd, -msp->ms_map.sm_size,
-	    -msp->ms_smo.smo_alloc, B_TRUE);
+	vdev_space_update(mg->mg_vd,
+	    -msp->ms_smo.smo_alloc, 0, -msp->ms_map->sm_size);
 
 	metaslab_group_remove(mg, msp);
 
 	mutex_enter(&msp->ms_lock);
 
-	space_map_unload(&msp->ms_map);
-	space_map_destroy(&msp->ms_map);
+	space_map_unload(msp->ms_map);
+	space_map_destroy(msp->ms_map);
+	kmem_free(msp->ms_map, sizeof (*msp->ms_map));
 
 	for (t = 0; t < TXG_SIZE; t++) {
-		space_map_destroy(&msp->ms_allocmap[t]);
-		space_map_destroy(&msp->ms_freemap[t]);
+		space_map_destroy(msp->ms_allocmap[t]);
+		space_map_destroy(msp->ms_freemap[t]);
+		kmem_free(msp->ms_allocmap[t], sizeof (*msp->ms_allocmap[t]));
+		kmem_free(msp->ms_freemap[t], sizeof (*msp->ms_freemap[t]));
+	}
+
+	for (t = 0; t < TXG_DEFER_SIZE; t++) {
+		space_map_destroy(msp->ms_defermap[t]);
+		kmem_free(msp->ms_defermap[t], sizeof (*msp->ms_defermap[t]));
 	}
 
+	ASSERT0(msp->ms_deferspace);
+
 	mutex_exit(&msp->ms_lock);
 	mutex_destroy(&msp->ms_lock);
 
@@ -522,13 +818,12 @@ metaslab_fini(metaslab_t *msp)
 #define	METASLAB_WEIGHT_SECONDARY	(1ULL << 62)
 #define	METASLAB_ACTIVE_MASK		\
 	(METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
-#define	METASLAB_SMO_BONUS_MULTIPLIER	2
 
 static uint64_t
 metaslab_weight(metaslab_t *msp)
 {
 	metaslab_group_t *mg = msp->ms_group;
-	space_map_t *sm = &msp->ms_map;
+	space_map_t *sm = msp->ms_map;
 	space_map_obj_t *smo = &msp->ms_smo;
 	vdev_t *vd = mg->mg_vd;
 	uint64_t weight, space;
@@ -555,44 +850,91 @@ metaslab_weight(metaslab_t *msp)
 	ASSERT(weight >= space && weight <= 2 * space);
 
 	/*
-	 * For locality, assign higher weight to metaslabs we've used before.
+	 * For locality, assign higher weight to metaslabs which have
+	 * a lower offset than what we've already activated.
 	 */
-	if (smo->smo_object != 0)
-		weight *= METASLAB_SMO_BONUS_MULTIPLIER;
+	if (sm->sm_start <= mg->mg_bonus_area)
+		weight *= (metaslab_smo_bonus_pct / 100);
 	ASSERT(weight >= space &&
-	    weight <= 2 * METASLAB_SMO_BONUS_MULTIPLIER * space);
+	    weight <= 2 * (metaslab_smo_bonus_pct / 100) * space);
+
+	if (sm->sm_loaded && !sm->sm_ops->smop_fragmented(sm)) {
+		/*
+		 * If this metaslab is one we're actively using, adjust its
+		 * weight to make it preferable to any inactive metaslab so
+		 * we'll polish it off.
+		 */
+		weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
+	}
+	return (weight);
+}
+
+static void
+metaslab_prefetch(metaslab_group_t *mg)
+{
+	spa_t *spa = mg->mg_vd->vdev_spa;
+	metaslab_t *msp;
+	avl_tree_t *t = &mg->mg_metaslab_tree;
+	int m;
+
+	mutex_enter(&mg->mg_lock);
 
 	/*
-	 * If this metaslab is one we're actively using, adjust its weight to
-	 * make it preferable to any inactive metaslab so we'll polish it off.
+	 * Prefetch the next potential metaslabs
 	 */
-	weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
+	for (msp = avl_first(t), m = 0; msp; msp = AVL_NEXT(t, msp), m++) {
+		space_map_t *sm = msp->ms_map;
+		space_map_obj_t *smo = &msp->ms_smo;
 
-	return (weight);
+		/* If we have reached our prefetch limit then we're done */
+		if (m >= metaslab_prefetch_limit)
+			break;
+
+		if (!sm->sm_loaded && smo->smo_object != 0) {
+			mutex_exit(&mg->mg_lock);
+			dmu_prefetch(spa_meta_objset(spa), smo->smo_object,
+			    0ULL, smo->smo_objsize);
+			mutex_enter(&mg->mg_lock);
+		}
+	}
+	mutex_exit(&mg->mg_lock);
 }
 
 static int
-metaslab_activate(metaslab_t *msp, uint64_t activation_weight, uint64_t size)
+metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
 {
-	space_map_t *sm = &msp->ms_map;
+	metaslab_group_t *mg = msp->ms_group;
+	space_map_t *sm = msp->ms_map;
 	space_map_ops_t *sm_ops = msp->ms_group->mg_class->mc_ops;
+	int t;
 
 	ASSERT(MUTEX_HELD(&msp->ms_lock));
 
 	if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
-		int error = space_map_load(sm, sm_ops, SM_FREE, &msp->ms_smo,
-		    msp->ms_group->mg_vd->vdev_spa->spa_meta_objset);
-		if (error) {
-			metaslab_group_sort(msp->ms_group, msp, 0);
-			return (error);
+		space_map_load_wait(sm);
+		if (!sm->sm_loaded) {
+			space_map_obj_t *smo = &msp->ms_smo;
+
+			int error = space_map_load(sm, sm_ops, SM_FREE, smo,
+			    spa_meta_objset(msp->ms_group->mg_vd->vdev_spa));
+			if (error)  {
+				metaslab_group_sort(msp->ms_group, msp, 0);
+				return (error);
+			}
+			for (t = 0; t < TXG_DEFER_SIZE; t++)
+				space_map_walk(msp->ms_defermap[t],
+				    space_map_claim, sm);
+
 		}
 
 		/*
-		 * If we were able to load the map then make sure
-		 * that this map is still able to satisfy our request.
+		 * Track the bonus area as we activate new metaslabs.
 		 */
-		if (msp->ms_weight < size)
-			return (ENOSPC);
+		if (sm->sm_start > mg->mg_bonus_area) {
+			mutex_enter(&mg->mg_lock);
+			mg->mg_bonus_area = sm->sm_start;
+			mutex_exit(&mg->mg_lock);
+		}
 
 		metaslab_group_sort(msp->ms_group, msp,
 		    msp->ms_weight | activation_weight);
@@ -611,12 +953,159 @@ metaslab_passivate(metaslab_t *msp, uint64_t size)
 	 * this metaslab again.  In that case, it had better be empty,
 	 * or we would be leaving space on the table.
 	 */
-	ASSERT(size >= SPA_MINBLOCKSIZE || msp->ms_map.sm_space == 0);
+	ASSERT(size >= SPA_MINBLOCKSIZE || msp->ms_map->sm_space == 0);
 	metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
 	ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
 }
 
 /*
+ * Determine if the in-core space map representation can be condensed on-disk.
+ * We would like to use the following criteria to make our decision:
+ *
+ * 1. The size of the space map object should not dramatically increase as a
+ * result of writing out our in-core free map.
+ *
+ * 2. The minimal on-disk space map representation is zfs_condense_pct/100
+ * times the size than the in-core representation (i.e. zfs_condense_pct = 110
+ * and in-core = 1MB, minimal = 1.1.MB).
+ *
+ * Checking the first condition is tricky since we don't want to walk
+ * the entire AVL tree calculating the estimated on-disk size. Instead we
+ * use the size-ordered AVL tree in the space map and calculate the
+ * size required for the largest segment in our in-core free map. If the
+ * size required to represent that segment on disk is larger than the space
+ * map object then we avoid condensing this map.
+ *
+ * To determine the second criterion we use a best-case estimate and assume
+ * each segment can be represented on-disk as a single 64-bit entry. We refer
+ * to this best-case estimate as the space map's minimal form.
+ */
+static boolean_t
+metaslab_should_condense(metaslab_t *msp)
+{
+	space_map_t *sm = msp->ms_map;
+	space_map_obj_t *smo = &msp->ms_smo_syncing;
+	space_seg_t *ss;
+	uint64_t size, entries, segsz;
+
+	ASSERT(MUTEX_HELD(&msp->ms_lock));
+	ASSERT(sm->sm_loaded);
+
+	/*
+	 * Use the sm_pp_root AVL tree, which is ordered by size, to obtain
+	 * the largest segment in the in-core free map. If the tree is
+	 * empty then we should condense the map.
+	 */
+	ss = avl_last(sm->sm_pp_root);
+	if (ss == NULL)
+		return (B_TRUE);
+
+	/*
+	 * Calculate the number of 64-bit entries this segment would
+	 * require when written to disk. If this single segment would be
+	 * larger on-disk than the entire current on-disk structure, then
+	 * clearly condensing will increase the on-disk structure size.
+	 */
+	size = (ss->ss_end - ss->ss_start) >> sm->sm_shift;
+	entries = size / (MIN(size, SM_RUN_MAX));
+	segsz = entries * sizeof (uint64_t);
+
+	return (segsz <= smo->smo_objsize &&
+	    smo->smo_objsize >= (zfs_condense_pct *
+	    sizeof (uint64_t) * avl_numnodes(&sm->sm_root)) / 100);
+}
+
+/*
+ * Condense the on-disk space map representation to its minimized form.
+ * The minimized form consists of a small number of allocations followed by
+ * the in-core free map.
+ */
+static void
+metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
+{
+	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
+	space_map_t *freemap = msp->ms_freemap[txg & TXG_MASK];
+	space_map_t condense_map;
+	space_map_t *sm = msp->ms_map;
+	objset_t *mos = spa_meta_objset(spa);
+	space_map_obj_t *smo = &msp->ms_smo_syncing;
+	int t;
+
+	ASSERT(MUTEX_HELD(&msp->ms_lock));
+	ASSERT3U(spa_sync_pass(spa), ==, 1);
+	ASSERT(sm->sm_loaded);
+
+	spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, "
+	    "smo size %llu, segments %lu", txg,
+	    (msp->ms_map->sm_start / msp->ms_map->sm_size), msp,
+	    smo->smo_objsize, avl_numnodes(&sm->sm_root));
+
+	/*
+	 * Create an map that is a 100% allocated map. We remove segments
+	 * that have been freed in this txg, any deferred frees that exist,
+	 * and any allocation in the future. Removing segments should be
+	 * a relatively inexpensive operation since we expect these maps to
+	 * a small number of nodes.
+	 */
+	space_map_create(&condense_map, sm->sm_start, sm->sm_size,
+	    sm->sm_shift, sm->sm_lock);
+	space_map_add(&condense_map, condense_map.sm_start,
+	    condense_map.sm_size);
+
+	/*
+	 * Remove what's been freed in this txg from the condense_map.
+	 * Since we're in sync_pass 1, we know that all the frees from
+	 * this txg are in the freemap.
+	 */
+	space_map_walk(freemap, space_map_remove, &condense_map);
+
+	for (t = 0; t < TXG_DEFER_SIZE; t++)
+		space_map_walk(msp->ms_defermap[t],
+		    space_map_remove, &condense_map);
+
+	for (t = 1; t < TXG_CONCURRENT_STATES; t++)
+		space_map_walk(msp->ms_allocmap[(txg + t) & TXG_MASK],
+		    space_map_remove, &condense_map);
+
+	/*
+	 * We're about to drop the metaslab's lock thus allowing
+	 * other consumers to change it's content. Set the
+	 * space_map's sm_condensing flag to ensure that
+	 * allocations on this metaslab do not occur while we're
+	 * in the middle of committing it to disk. This is only critical
+	 * for the ms_map as all other space_maps use per txg
+	 * views of their content.
+	 */
+	sm->sm_condensing = B_TRUE;
+
+	mutex_exit(&msp->ms_lock);
+	space_map_truncate(smo, mos, tx);
+	mutex_enter(&msp->ms_lock);
+
+	/*
+	 * While we would ideally like to create a space_map representation
+	 * that consists only of allocation records, doing so can be
+	 * prohibitively expensive because the in-core free map can be
+	 * large, and therefore computationally expensive to subtract
+	 * from the condense_map. Instead we sync out two maps, a cheap
+	 * allocation only map followed by the in-core free map. While not
+	 * optimal, this is typically close to optimal, and much cheaper to
+	 * compute.
+	 */
+	space_map_sync(&condense_map, SM_ALLOC, smo, mos, tx);
+	space_map_vacate(&condense_map, NULL, NULL);
+	space_map_destroy(&condense_map);
+
+	space_map_sync(sm, SM_FREE, smo, mos, tx);
+	sm->sm_condensing = B_FALSE;
+
+	spa_dbgmsg(spa, "condensed: txg %llu, msp[%llu] %p, "
+	    "smo size %llu", txg,
+	    (msp->ms_map->sm_start / msp->ms_map->sm_size), msp,
+	    smo->smo_objsize);
+}
+
+/*
  * Write a metaslab to disk in the context of the specified transaction group.
  */
 void
@@ -624,17 +1113,31 @@ metaslab_sync(metaslab_t *msp, uint64_t txg)
 {
 	vdev_t *vd = msp->ms_group->mg_vd;
 	spa_t *spa = vd->vdev_spa;
-	objset_t *mos = spa->spa_meta_objset;
-	space_map_t *allocmap = &msp->ms_allocmap[txg & TXG_MASK];
-	space_map_t *freemap = &msp->ms_freemap[txg & TXG_MASK];
-	space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
-	space_map_t *sm = &msp->ms_map;
+	objset_t *mos = spa_meta_objset(spa);
+	space_map_t *allocmap = msp->ms_allocmap[txg & TXG_MASK];
+	space_map_t **freemap = &msp->ms_freemap[txg & TXG_MASK];
+	space_map_t **freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
+	space_map_t *sm = msp->ms_map;
 	space_map_obj_t *smo = &msp->ms_smo_syncing;
 	dmu_buf_t *db;
 	dmu_tx_t *tx;
-	int t;
 
-	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
+	ASSERT(!vd->vdev_ishole);
+
+	/*
+	 * This metaslab has just been added so there's no work to do now.
+	 */
+	if (*freemap == NULL) {
+		ASSERT3P(allocmap, ==, NULL);
+		return;
+	}
+
+	ASSERT3P(allocmap, !=, NULL);
+	ASSERT3P(*freemap, !=, NULL);
+	ASSERT3P(*freed_map, !=, NULL);
+
+	if (allocmap->sm_space == 0 && (*freemap)->sm_space == 0)
+		return;
 
 	/*
 	 * The only state that can actually be changing concurrently with
@@ -644,12 +1147,12 @@ metaslab_sync(metaslab_t *msp, uint64_t txg)
 	 * We drop it whenever we call into the DMU, because the DMU
 	 * can call down to us (e.g. via zio_free()) at any time.
 	 */
-	mutex_enter(&msp->ms_lock);
+
+	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
 
 	if (smo->smo_object == 0) {
 		ASSERT(smo->smo_objsize == 0);
 		ASSERT(smo->smo_alloc == 0);
-		mutex_exit(&msp->ms_lock);
 		smo->smo_object = dmu_object_alloc(mos,
 		    DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
 		    DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
@@ -657,47 +1160,40 @@ metaslab_sync(metaslab_t *msp, uint64_t txg)
 		dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
 		    (sm->sm_start >> vd->vdev_ms_shift),
 		    sizeof (uint64_t), &smo->smo_object, tx);
-		mutex_enter(&msp->ms_lock);
 	}
 
-	space_map_walk(freemap, space_map_add, freed_map);
-
-	if (sm->sm_loaded && spa_sync_pass(spa) == 1 && smo->smo_objsize >=
-	    2 * sizeof (uint64_t) * avl_numnodes(&sm->sm_root)) {
-		/*
-		 * The in-core space map representation is twice as compact
-		 * as the on-disk one, so it's time to condense the latter
-		 * by generating a pure allocmap from first principles.
-		 *
-		 * This metaslab is 100% allocated,
-		 * minus the content of the in-core map (sm),
-		 * minus what's been freed this txg (freed_map),
-		 * minus allocations from txgs in the future
-		 * (because they haven't been committed yet).
-		 */
-		space_map_vacate(allocmap, NULL, NULL);
-		space_map_vacate(freemap, NULL, NULL);
-
-		space_map_add(allocmap, allocmap->sm_start, allocmap->sm_size);
+	mutex_enter(&msp->ms_lock);
 
-		space_map_walk(sm, space_map_remove, allocmap);
-		space_map_walk(freed_map, space_map_remove, allocmap);
+	if (sm->sm_loaded && spa_sync_pass(spa) == 1 &&
+	    metaslab_should_condense(msp)) {
+		metaslab_condense(msp, txg, tx);
+	} else {
+		space_map_sync(allocmap, SM_ALLOC, smo, mos, tx);
+		space_map_sync(*freemap, SM_FREE, smo, mos, tx);
+	}
 
-		for (t = 1; t < TXG_CONCURRENT_STATES; t++)
-			space_map_walk(&msp->ms_allocmap[(txg + t) & TXG_MASK],
-			    space_map_remove, allocmap);
+	space_map_vacate(allocmap, NULL, NULL);
 
-		mutex_exit(&msp->ms_lock);
-		space_map_truncate(smo, mos, tx);
-		mutex_enter(&msp->ms_lock);
+	/*
+	 * For sync pass 1, we avoid walking the entire space map and
+	 * instead will just swap the pointers for freemap and
+	 * freed_map. We can safely do this since the freed_map is
+	 * guaranteed to be empty on the initial pass.
+	 */
+	if (spa_sync_pass(spa) == 1) {
+		ASSERT0((*freed_map)->sm_space);
+		ASSERT0(avl_numnodes(&(*freed_map)->sm_root));
+		space_map_swap(freemap, freed_map);
+	} else {
+		space_map_vacate(*freemap, space_map_add, *freed_map);
 	}
 
-	space_map_sync(allocmap, SM_ALLOC, smo, mos, tx);
-	space_map_sync(freemap, SM_FREE, smo, mos, tx);
+	ASSERT0(msp->ms_allocmap[txg & TXG_MASK]->sm_space);
+	ASSERT0(msp->ms_freemap[txg & TXG_MASK]->sm_space);
 
 	mutex_exit(&msp->ms_lock);
 
-	VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
+	VERIFY0(dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
 	dmu_buf_will_dirty(db, tx);
 	ASSERT3U(db->db_size, >=, sizeof (*smo));
 	bcopy(smo, db->db_data, sizeof (*smo));
@@ -715,43 +1211,79 @@ metaslab_sync_done(metaslab_t *msp, uint64_t txg)
 {
 	space_map_obj_t *smo = &msp->ms_smo;
 	space_map_obj_t *smosync = &msp->ms_smo_syncing;
-	space_map_t *sm = &msp->ms_map;
-	space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
+	space_map_t *sm = msp->ms_map;
+	space_map_t *freed_map = msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
+	space_map_t *defer_map = msp->ms_defermap[txg % TXG_DEFER_SIZE];
 	metaslab_group_t *mg = msp->ms_group;
 	vdev_t *vd = mg->mg_vd;
+	int64_t alloc_delta, defer_delta;
 	int t;
 
+	ASSERT(!vd->vdev_ishole);
+
 	mutex_enter(&msp->ms_lock);
 
 	/*
 	 * If this metaslab is just becoming available, initialize its
-	 * allocmaps and freemaps and add its capacity to the vdev.
+	 * allocmaps, freemaps, and defermap and add its capacity to the vdev.
 	 */
-	if (freed_map->sm_size == 0) {
+	if (freed_map == NULL) {
+		ASSERT(defer_map == NULL);
 		for (t = 0; t < TXG_SIZE; t++) {
-			space_map_create(&msp->ms_allocmap[t], sm->sm_start,
+			msp->ms_allocmap[t] = kmem_zalloc(sizeof (space_map_t),
+			    KM_PUSHPAGE);
+			space_map_create(msp->ms_allocmap[t], sm->sm_start,
+			    sm->sm_size, sm->sm_shift, sm->sm_lock);
+			msp->ms_freemap[t] = kmem_zalloc(sizeof (space_map_t),
+			    KM_PUSHPAGE);
+			space_map_create(msp->ms_freemap[t], sm->sm_start,
 			    sm->sm_size, sm->sm_shift, sm->sm_lock);
-			space_map_create(&msp->ms_freemap[t], sm->sm_start,
+		}
+
+		for (t = 0; t < TXG_DEFER_SIZE; t++) {
+			msp->ms_defermap[t] = kmem_zalloc(sizeof (space_map_t),
+			    KM_PUSHPAGE);
+			space_map_create(msp->ms_defermap[t], sm->sm_start,
 			    sm->sm_size, sm->sm_shift, sm->sm_lock);
 		}
-		vdev_space_update(vd, sm->sm_size, 0, B_TRUE);
+
+		freed_map = msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
+		defer_map = msp->ms_defermap[txg % TXG_DEFER_SIZE];
+
+		vdev_space_update(vd, 0, 0, sm->sm_size);
 	}
 
-	vdev_space_update(vd, 0, smosync->smo_alloc - smo->smo_alloc, B_TRUE);
+	alloc_delta = smosync->smo_alloc - smo->smo_alloc;
+	defer_delta = freed_map->sm_space - defer_map->sm_space;
 
-	ASSERT(msp->ms_allocmap[txg & TXG_MASK].sm_space == 0);
-	ASSERT(msp->ms_freemap[txg & TXG_MASK].sm_space == 0);
+	vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
+
+	ASSERT(msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0);
+	ASSERT(msp->ms_freemap[txg & TXG_MASK]->sm_space == 0);
 
 	/*
 	 * If there's a space_map_load() in progress, wait for it to complete
 	 * so that we have a consistent view of the in-core space map.
-	 * Then, add everything we freed in this txg to the map.
+	 * Then, add defer_map (oldest deferred frees) to this map and
+	 * transfer freed_map (this txg's frees) to defer_map.
 	 */
 	space_map_load_wait(sm);
-	space_map_vacate(freed_map, sm->sm_loaded ? space_map_free : NULL, sm);
+	space_map_vacate(defer_map, sm->sm_loaded ? space_map_free : NULL, sm);
+	space_map_vacate(freed_map, space_map_add, defer_map);
 
 	*smo = *smosync;
 
+	msp->ms_deferspace += defer_delta;
+	ASSERT3S(msp->ms_deferspace, >=, 0);
+	ASSERT3S(msp->ms_deferspace, <=, sm->sm_size);
+	if (msp->ms_deferspace != 0) {
+		/*
+		 * Keep syncing this metaslab until all deferred frees
+		 * are back in circulation.
+		 */
+		vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
+	}
+
 	/*
 	 * If the map is loaded but no longer active, evict it as soon as all
 	 * future allocations have synced.  (If we unloaded it now and then
@@ -761,10 +1293,10 @@ metaslab_sync_done(metaslab_t *msp, uint64_t txg)
 		int evictable = 1;
 
 		for (t = 1; t < TXG_CONCURRENT_STATES; t++)
-			if (msp->ms_allocmap[(txg + t) & TXG_MASK].sm_space)
+			if (msp->ms_allocmap[(txg + t) & TXG_MASK]->sm_space)
 				evictable = 0;
 
-		if (evictable)
+		if (evictable && !metaslab_debug)
 			space_map_unload(sm);
 	}
 
@@ -773,12 +1305,42 @@ metaslab_sync_done(metaslab_t *msp, uint64_t txg)
 	mutex_exit(&msp->ms_lock);
 }
 
+void
+metaslab_sync_reassess(metaslab_group_t *mg)
+{
+	vdev_t *vd = mg->mg_vd;
+	int64_t failures = mg->mg_alloc_failures;
+	int m;
+
+	/*
+	 * Re-evaluate all metaslabs which have lower offsets than the
+	 * bonus area.
+	 */
+	for (m = 0; m < vd->vdev_ms_count; m++) {
+		metaslab_t *msp = vd->vdev_ms[m];
+
+		if (msp->ms_map->sm_start > mg->mg_bonus_area)
+			break;
+
+		mutex_enter(&msp->ms_lock);
+		metaslab_group_sort(mg, msp, metaslab_weight(msp));
+		mutex_exit(&msp->ms_lock);
+	}
+
+	atomic_add_64(&mg->mg_alloc_failures, -failures);
+
+	/*
+	 * Prefetch the next potential metaslabs
+	 */
+	metaslab_prefetch(mg);
+}
+
 static uint64_t
 metaslab_distance(metaslab_t *msp, dva_t *dva)
 {
 	uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
 	uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
-	uint64_t start = msp->ms_map.sm_start >> ms_shift;
+	uint64_t start = msp->ms_map->sm_start >> ms_shift;
 
 	if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
 		return (1ULL << 63);
@@ -791,9 +1353,10 @@ metaslab_distance(metaslab_t *msp, dva_t *dva)
 }
 
 static uint64_t
-metaslab_group_alloc(metaslab_group_t *mg, uint64_t size, uint64_t txg,
-    uint64_t min_distance, dva_t *dva, int d)
+metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
+    uint64_t txg, uint64_t min_distance, dva_t *dva, int d, int flags)
 {
+	spa_t *spa = mg->mg_vd->vdev_spa;
 	metaslab_t *msp = NULL;
 	uint64_t offset = -1ULL;
 	avl_tree_t *t = &mg->mg_metaslab_tree;
@@ -814,11 +1377,17 @@ metaslab_group_alloc(metaslab_group_t *mg, uint64_t size, uint64_t txg,
 
 		mutex_enter(&mg->mg_lock);
 		for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
-			if (msp->ms_weight < size) {
+			if (msp->ms_weight < asize) {
+				spa_dbgmsg(spa, "%s: failed to meet weight "
+				    "requirement: vdev %llu, txg %llu, mg %p, "
+				    "msp %p, psize %llu, asize %llu, "
+				    "failures %llu, weight %llu",
+				    spa_name(spa), mg->mg_vd->vdev_id, txg,
+				    mg, msp, psize, asize,
+				    mg->mg_alloc_failures, msp->ms_weight);
 				mutex_exit(&mg->mg_lock);
 				return (-1ULL);
 			}
-
 			was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
 			if (activation_weight == METASLAB_WEIGHT_PRIMARY)
 				break;
@@ -837,15 +1406,44 @@ metaslab_group_alloc(metaslab_group_t *mg, uint64_t size, uint64_t txg,
 		if (msp == NULL)
 			return (-1ULL);
 
+		/*
+		 * If we've already reached the allowable number of failed
+		 * allocation attempts on this metaslab group then we
+		 * consider skipping it. We skip it only if we're allowed
+		 * to "fast" gang, the physical size is larger than
+		 * a gang block, and we're attempting to allocate from
+		 * the primary metaslab.
+		 */
+		if (mg->mg_alloc_failures > zfs_mg_alloc_failures &&
+		    CAN_FASTGANG(flags) && psize > SPA_GANGBLOCKSIZE &&
+		    activation_weight == METASLAB_WEIGHT_PRIMARY) {
+			spa_dbgmsg(spa, "%s: skipping metaslab group: "
+			    "vdev %llu, txg %llu, mg %p, psize %llu, "
+			    "asize %llu, failures %llu", spa_name(spa),
+			    mg->mg_vd->vdev_id, txg, mg, psize, asize,
+			    mg->mg_alloc_failures);
+			return (-1ULL);
+		}
+
 		mutex_enter(&msp->ms_lock);
 
 		/*
+		 * If this metaslab is currently condensing then pick again as
+		 * we can't manipulate this metaslab until it's committed
+		 * to disk.
+		 */
+		if (msp->ms_map->sm_condensing) {
+			mutex_exit(&msp->ms_lock);
+			continue;
+		}
+
+		/*
 		 * Ensure that the metaslab we have selected is still
 		 * capable of handling our request. It's possible that
 		 * another thread may have changed the weight while we
 		 * were blocked on the metaslab lock.
 		 */
-		if (msp->ms_weight < size || (was_active &&
+		if (msp->ms_weight < asize || (was_active &&
 		    !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
 		    activation_weight == METASLAB_WEIGHT_PRIMARY)) {
 			mutex_exit(&msp->ms_lock);
@@ -860,23 +1458,25 @@ metaslab_group_alloc(metaslab_group_t *mg, uint64_t size, uint64_t txg,
 			continue;
 		}
 
-		if (metaslab_activate(msp, activation_weight, size) != 0) {
+		if (metaslab_activate(msp, activation_weight) != 0) {
 			mutex_exit(&msp->ms_lock);
 			continue;
 		}
 
-		if ((offset = space_map_alloc(&msp->ms_map, size)) != -1ULL)
+		if ((offset = space_map_alloc(msp->ms_map, asize)) != -1ULL)
 			break;
 
-		metaslab_passivate(msp, size - 1);
+		atomic_inc_64(&mg->mg_alloc_failures);
+
+		metaslab_passivate(msp, space_map_maxsize(msp->ms_map));
 
 		mutex_exit(&msp->ms_lock);
 	}
 
-	if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
+	if (msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0)
 		vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
 
-	space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, size);
+	space_map_add(msp->ms_allocmap[txg & TXG_MASK], offset, asize);
 
 	mutex_exit(&msp->ms_lock);
 
@@ -890,7 +1490,7 @@ static int
 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
     dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
 {
-	metaslab_group_t *mg, *rotor;
+	metaslab_group_t *mg, *fast_mg, *rotor;
 	vdev_t *vd;
 	int dshift = 3;
 	int all_zero;
@@ -905,12 +1505,15 @@ metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
 	/*
 	 * For testing, make some blocks above a certain size be gang blocks.
 	 */
-	if (psize >= metaslab_gang_bang && (lbolt & 3) == 0)
+	if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
 		return (ENOSPC);
 
+	if (flags & METASLAB_FASTWRITE)
+		mutex_enter(&mc->mc_fastwrite_lock);
+
 	/*
 	 * Start at the rotor and loop through all mgs until we find something.
-	 * Note that there's no locking on mc_rotor or mc_allocated because
+	 * Note that there's no locking on mc_rotor or mc_aliquot because
 	 * nothing actually breaks if we miss a few updates -- we just won't
 	 * allocate quite as evenly.  It all balances out over time.
 	 *
@@ -932,27 +1535,50 @@ metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
 	 */
 	if (hintdva) {
 		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
-		if (flags & METASLAB_HINTBP_AVOID)
-			mg = vd->vdev_mg->mg_next;
-		else
+
+		/*
+		 * It's possible the vdev we're using as the hint no
+		 * longer exists (i.e. removed). Consult the rotor when
+		 * all else fails.
+		 */
+		if (vd != NULL) {
 			mg = vd->vdev_mg;
+
+			if (flags & METASLAB_HINTBP_AVOID &&
+			    mg->mg_next != NULL)
+				mg = mg->mg_next;
+		} else {
+			mg = mc->mc_rotor;
+		}
 	} else if (d != 0) {
 		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
 		mg = vd->vdev_mg->mg_next;
+	} else if (flags & METASLAB_FASTWRITE) {
+		mg = fast_mg = mc->mc_rotor;
+
+		do {
+			if (fast_mg->mg_vd->vdev_pending_fastwrite <
+			    mg->mg_vd->vdev_pending_fastwrite)
+				mg = fast_mg;
+		} while ((fast_mg = fast_mg->mg_next) != mc->mc_rotor);
+
 	} else {
 		mg = mc->mc_rotor;
 	}
 
 	/*
-	 * If the hint put us into the wrong class, just follow the rotor.
+	 * If the hint put us into the wrong metaslab class, or into a
+	 * metaslab group that has been passivated, just follow the rotor.
 	 */
-	if (mg->mg_class != mc)
+	if (mg->mg_class != mc || mg->mg_activation_count <= 0)
 		mg = mc->mc_rotor;
 
 	rotor = mg;
 top:
 	all_zero = B_TRUE;
 	do {
+		ASSERT(mg->mg_activation_count == 1);
+
 		vd = mg->mg_vd;
 
 		/*
@@ -989,7 +1615,8 @@ top:
 		asize = vdev_psize_to_asize(vd, psize);
 		ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
 
-		offset = metaslab_group_alloc(mg, asize, txg, distance, dva, d);
+		offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
+		    dva, d, flags);
 		if (offset != -1ULL) {
 			/*
 			 * If we've just selected this metaslab group,
@@ -997,32 +1624,35 @@ top:
 			 * over- or under-used relative to the pool,
 			 * and set an allocation bias to even it out.
 			 */
-			if (mc->mc_allocated == 0) {
+			if (mc->mc_aliquot == 0) {
 				vdev_stat_t *vs = &vd->vdev_stat;
-				uint64_t alloc, space;
-				int64_t vu, su;
+				int64_t vu, cu;
 
-				alloc = spa_get_alloc(spa);
-				space = spa_get_space(spa);
-
-				/*
-				 * Determine percent used in units of 0..1024.
-				 * (This is just to avoid floating point.)
-				 */
-				vu = (vs->vs_alloc << 10) / (vs->vs_space + 1);
-				su = (alloc << 10) / (space + 1);
+				vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
+				cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
 
 				/*
-				 * Bias by at most +/- 25% of the aliquot.
+				 * Calculate how much more or less we should
+				 * try to allocate from this device during
+				 * this iteration around the rotor.
+				 * For example, if a device is 80% full
+				 * and the pool is 20% full then we should
+				 * reduce allocations by 60% on this device.
+				 *
+				 * mg_bias = (20 - 80) * 512K / 100 = -307K
+				 *
+				 * This reduces allocations by 307K for this
+				 * iteration.
 				 */
-				mg->mg_bias = ((su - vu) *
-				    (int64_t)mg->mg_aliquot) / (1024 * 4);
+				mg->mg_bias = ((cu - vu) *
+				    (int64_t)mg->mg_aliquot) / 100;
 			}
 
-			if (atomic_add_64_nv(&mc->mc_allocated, asize) >=
+			if ((flags & METASLAB_FASTWRITE) ||
+			    atomic_add_64_nv(&mc->mc_aliquot, asize) >=
 			    mg->mg_aliquot + mg->mg_bias) {
 				mc->mc_rotor = mg->mg_next;
-				mc->mc_allocated = 0;
+				mc->mc_aliquot = 0;
 			}
 
 			DVA_SET_VDEV(&dva[d], vd->vdev_id);
@@ -1030,11 +1660,17 @@ top:
 			DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
 			DVA_SET_ASIZE(&dva[d], asize);
 
+			if (flags & METASLAB_FASTWRITE) {
+				atomic_add_64(&vd->vdev_pending_fastwrite,
+				    psize);
+				mutex_exit(&mc->mc_fastwrite_lock);
+			}
+
 			return (0);
 		}
 next:
 		mc->mc_rotor = mg->mg_next;
-		mc->mc_allocated = 0;
+		mc->mc_aliquot = 0;
 	} while ((mg = mg->mg_next) != rotor);
 
 	if (!all_zero) {
@@ -1051,6 +1687,8 @@ next:
 
 	bzero(&dva[d], sizeof (dva_t));
 
+	if (flags & METASLAB_FASTWRITE)
+		mutex_exit(&mc->mc_fastwrite_lock);
 	return (ENOSPC);
 }
 
@@ -1088,13 +1726,13 @@ metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
 	mutex_enter(&msp->ms_lock);
 
 	if (now) {
-		space_map_remove(&msp->ms_allocmap[txg & TXG_MASK],
+		space_map_remove(msp->ms_allocmap[txg & TXG_MASK],
 		    offset, size);
-		space_map_free(&msp->ms_map, offset, size);
+		space_map_free(msp->ms_map, offset, size);
 	} else {
-		if (msp->ms_freemap[txg & TXG_MASK].sm_space == 0)
+		if (msp->ms_freemap[txg & TXG_MASK]->sm_space == 0)
 			vdev_dirty(vd, VDD_METASLAB, msp, txg);
-		space_map_add(&msp->ms_freemap[txg & TXG_MASK], offset, size);
+		space_map_add(msp->ms_freemap[txg & TXG_MASK], offset, size);
 	}
 
 	mutex_exit(&msp->ms_lock);
@@ -1114,7 +1752,7 @@ metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
 	uint64_t size = DVA_GET_ASIZE(dva);
 	vdev_t *vd;
 	metaslab_t *msp;
-	int error;
+	int error = 0;
 
 	ASSERT(DVA_IS_VALID(dva));
 
@@ -1129,18 +1767,23 @@ metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
 
 	mutex_enter(&msp->ms_lock);
 
-	error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY, 0);
+	if ((txg != 0 && spa_writeable(spa)) || !msp->ms_map->sm_loaded)
+		error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
+
+	if (error == 0 && !space_map_contains(msp->ms_map, offset, size))
+		error = ENOENT;
+
 	if (error || txg == 0) {	/* txg == 0 indicates dry run */
 		mutex_exit(&msp->ms_lock);
 		return (error);
 	}
 
-	space_map_claim(&msp->ms_map, offset, size);
+	space_map_claim(msp->ms_map, offset, size);
 
 	if (spa_writeable(spa)) {	/* don't dirty if we're zdb(1M) */
-		if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
+		if (msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0)
 			vdev_dirty(vd, VDD_METASLAB, msp, txg);
-		space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, size);
+		space_map_add(msp->ms_allocmap[txg & TXG_MASK], offset, size);
 	}
 
 	mutex_exit(&msp->ms_lock);
@@ -1154,9 +1797,10 @@ metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
 {
 	dva_t *dva = bp->blk_dva;
 	dva_t *hintdva = hintbp->blk_dva;
-	int error = 0;
+	int d, error = 0;
 
 	ASSERT(bp->blk_birth == 0);
+	ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
 
 	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
 
@@ -1169,7 +1813,7 @@ metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
 	ASSERT(BP_GET_NDVAS(bp) == 0);
 	ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
 
-	for (int d = 0; d < ndvas; d++) {
+	for (d = 0; d < ndvas; d++) {
 		error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
 		    txg, flags);
 		if (error) {
@@ -1186,7 +1830,7 @@ metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
 
 	spa_config_exit(spa, SCL_ALLOC, FTAG);
 
-	bp->blk_birth = txg;
+	BP_SET_BIRTH(bp, txg, txg);
 
 	return (0);
 }
@@ -1195,14 +1839,14 @@ void
 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
 {
 	const dva_t *dva = bp->blk_dva;
-	int ndvas = BP_GET_NDVAS(bp);
+	int d, ndvas = BP_GET_NDVAS(bp);
 
 	ASSERT(!BP_IS_HOLE(bp));
-	ASSERT(!now || bp->blk_birth >= spa->spa_syncing_txg);
+	ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
 
 	spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
 
-	for (int d = 0; d < ndvas; d++)
+	for (d = 0; d < ndvas; d++)
 		metaslab_free_dva(spa, &dva[d], txg, now);
 
 	spa_config_exit(spa, SCL_FREE, FTAG);
@@ -1213,7 +1857,7 @@ metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
 {
 	const dva_t *dva = bp->blk_dva;
 	int ndvas = BP_GET_NDVAS(bp);
-	int error = 0;
+	int d, error = 0;
 
 	ASSERT(!BP_IS_HOLE(bp));
 
@@ -1228,7 +1872,7 @@ metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
 
 	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
 
-	for (int d = 0; d < ndvas; d++)
+	for (d = 0; d < ndvas; d++)
 		if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
 			break;
 
@@ -1238,3 +1882,53 @@ metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
 
 	return (error);
 }
+
+void metaslab_fastwrite_mark(spa_t *spa, const blkptr_t *bp)
+{
+	const dva_t *dva = bp->blk_dva;
+	int ndvas = BP_GET_NDVAS(bp);
+	uint64_t psize = BP_GET_PSIZE(bp);
+	int d;
+	vdev_t *vd;
+
+	ASSERT(!BP_IS_HOLE(bp));
+	ASSERT(psize > 0);
+
+	spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
+
+	for (d = 0; d < ndvas; d++) {
+		if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL)
+			continue;
+		atomic_add_64(&vd->vdev_pending_fastwrite, psize);
+	}
+
+	spa_config_exit(spa, SCL_VDEV, FTAG);
+}
+
+void metaslab_fastwrite_unmark(spa_t *spa, const blkptr_t *bp)
+{
+	const dva_t *dva = bp->blk_dva;
+	int ndvas = BP_GET_NDVAS(bp);
+	uint64_t psize = BP_GET_PSIZE(bp);
+	int d;
+	vdev_t *vd;
+
+	ASSERT(!BP_IS_HOLE(bp));
+	ASSERT(psize > 0);
+
+	spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
+
+	for (d = 0; d < ndvas; d++) {
+		if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL)
+			continue;
+		ASSERT3U(vd->vdev_pending_fastwrite, >=, psize);
+		atomic_sub_64(&vd->vdev_pending_fastwrite, psize);
+	}
+
+	spa_config_exit(spa, SCL_VDEV, FTAG);
+}
+
+#if defined(_KERNEL) && defined(HAVE_SPL)
+module_param(metaslab_debug, int, 0644);
+MODULE_PARM_DESC(metaslab_debug, "keep space maps in core to verify frees");
+#endif /* _KERNEL && HAVE_SPL */