Optimize allocation throttling

Remove mc_lock use from metaslab_class_throttle_*().  The math there
is based on refcounts and so atomic, so the only race possible there
is between zfs_refcount_count() and zfs_refcount_add().  But in most
cases metaslab_class_throttle_reserve() is called with the allocator
lock held, which covers the race.  In cases where the lock is not
held, GANG_ALLOCATION() or METASLAB_MUST_RESERVE are set, and so we
do not use zfs_refcount_count().  And even if we assume some other
non-existing scenario, the worst that may happen from this race is
few more I/Os get to allocation earlier, that is not a problem.

Move locks and data of different allocators into different cache
lines to avoid false sharing.  Group spa_alloc_* arrays together
into single array of aligned struct spa_alloc spa_allocs.  Align
struct metaslab_class_allocator.

Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Don Brady <don.brady@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes #12314

include/sys/metaslab_impl.h

@@ -157,7 +157,7 @@ typedef struct metaslab_class_allocator {
 	 */
 	uint64_t		mca_alloc_max_slots;
 	zfs_refcount_t		mca_alloc_slots;
-} metaslab_class_allocator_t;
+} ____cacheline_aligned metaslab_class_allocator_t;
 
 /*
  * A metaslab class encompasses a category of allocatable top-level vdevs.

include/sys/spa_impl.h

@@ -57,6 +57,11 @@
 extern "C" {
 #endif
 
+typedef struct spa_alloc {
+	kmutex_t	spaa_lock;
+	avl_tree_t	spaa_tree;
+} ____cacheline_aligned spa_alloc_t;
+
 typedef struct spa_error_entry {
 	zbookmark_phys_t	se_bookmark;
 	char			*se_name;
@@ -250,13 +255,11 @@ struct spa {
 	list_t		spa_config_dirty_list;	/* vdevs with dirty config */
 	list_t		spa_state_dirty_list;	/* vdevs with dirty state */
 	/*
-	 * spa_alloc_locks and spa_alloc_trees are arrays, whose lengths are
-	 * stored in spa_alloc_count. There is one tree and one lock for each
-	 * allocator, to help improve allocation performance in write-heavy
-	 * workloads.
+	 * spa_allocs is an array, whose lengths is stored in spa_alloc_count.
+	 * There is one tree and one lock for each allocator, to help improve
+	 * allocation performance in write-heavy workloads.
 	 */
-	kmutex_t	*spa_alloc_locks;
-	avl_tree_t	*spa_alloc_trees;
+	spa_alloc_t	*spa_allocs;
 	int		spa_alloc_count;
 
 	spa_aux_vdev_t	spa_spares;		/* hot spares */

module/zfs/metaslab.c

@@ -5611,19 +5611,11 @@ metaslab_class_throttle_reserve(metaslab_class_t *mc, int slots, int allocator,
     zio_t *zio, int flags)
 {
 	metaslab_class_allocator_t *mca = &mc->mc_allocator[allocator];
-	uint64_t available_slots = 0;
-	boolean_t slot_reserved = B_FALSE;
 	uint64_t max = mca->mca_alloc_max_slots;
 
 	ASSERT(mc->mc_alloc_throttle_enabled);
-	mutex_enter(&mc->mc_lock);
-
-	uint64_t reserved_slots = zfs_refcount_count(&mca->mca_alloc_slots);
-	if (reserved_slots < max)
-		available_slots = max - reserved_slots;
-
-	if (slots <= available_slots || GANG_ALLOCATION(flags) ||
-	    flags & METASLAB_MUST_RESERVE) {
+	if (GANG_ALLOCATION(flags) || (flags & METASLAB_MUST_RESERVE) ||
+	    zfs_refcount_count(&mca->mca_alloc_slots) + slots <= max) {
 		/*
 		 * We reserve the slots individually so that we can unreserve
 		 * them individually when an I/O completes.
@@ -5631,11 +5623,9 @@ metaslab_class_throttle_reserve(metaslab_class_t *mc, int slots, int allocator,
 		for (int d = 0; d < slots; d++)
 			zfs_refcount_add(&mca->mca_alloc_slots, zio);
 		zio->io_flags |= ZIO_FLAG_IO_ALLOCATING;
-		slot_reserved = B_TRUE;
+		return (B_TRUE);
 	}
-
-	mutex_exit(&mc->mc_lock);
-	return (slot_reserved);
+	return (B_FALSE);
 }
 
 void
@@ -5645,10 +5635,8 @@ metaslab_class_throttle_unreserve(metaslab_class_t *mc, int slots,
 	metaslab_class_allocator_t *mca = &mc->mc_allocator[allocator];
 
 	ASSERT(mc->mc_alloc_throttle_enabled);
-	mutex_enter(&mc->mc_lock);
 	for (int d = 0; d < slots; d++)
 		zfs_refcount_remove(&mca->mca_alloc_slots, zio);
-	mutex_exit(&mc->mc_lock);
 }
 
 static int

module/zfs/spa.c

@@ -9197,9 +9197,9 @@ spa_sync(spa_t *spa, uint64_t txg)
 	spa->spa_sync_pass = 0;
 
 	for (int i = 0; i < spa->spa_alloc_count; i++) {
-		mutex_enter(&spa->spa_alloc_locks[i]);
-		VERIFY0(avl_numnodes(&spa->spa_alloc_trees[i]));
-		mutex_exit(&spa->spa_alloc_locks[i]);
+		mutex_enter(&spa->spa_allocs[i].spaa_lock);
+		VERIFY0(avl_numnodes(&spa->spa_allocs[i].spaa_tree));
+		mutex_exit(&spa->spa_allocs[i].spaa_lock);
 	}
 
 	/*
@@ -9309,9 +9309,9 @@ spa_sync(spa_t *spa, uint64_t txg)
 	dsl_pool_sync_done(dp, txg);
 
 	for (int i = 0; i < spa->spa_alloc_count; i++) {
-		mutex_enter(&spa->spa_alloc_locks[i]);
-		VERIFY0(avl_numnodes(&spa->spa_alloc_trees[i]));
-		mutex_exit(&spa->spa_alloc_locks[i]);
+		mutex_enter(&spa->spa_allocs[i].spaa_lock);
+		VERIFY0(avl_numnodes(&spa->spa_allocs[i].spaa_tree));
+		mutex_exit(&spa->spa_allocs[i].spaa_lock);
 	}
 
 	/*

module/zfs/spa_misc.c

@@ -700,13 +700,12 @@ spa_add(const char *name, nvlist_t *config, const char *altroot)
 		spa->spa_root = spa_strdup(altroot);
 
 	spa->spa_alloc_count = spa_allocators;
-	spa->spa_alloc_locks = kmem_zalloc(spa->spa_alloc_count *
-	    sizeof (kmutex_t), KM_SLEEP);
-	spa->spa_alloc_trees = kmem_zalloc(spa->spa_alloc_count *
-	    sizeof (avl_tree_t), KM_SLEEP);
+	spa->spa_allocs = kmem_zalloc(spa->spa_alloc_count *
+	    sizeof (spa_alloc_t), KM_SLEEP);
 	for (int i = 0; i < spa->spa_alloc_count; i++) {
-		mutex_init(&spa->spa_alloc_locks[i], NULL, MUTEX_DEFAULT, NULL);
-		avl_create(&spa->spa_alloc_trees[i], zio_bookmark_compare,
+		mutex_init(&spa->spa_allocs[i].spaa_lock, NULL, MUTEX_DEFAULT,
+		    NULL);
+		avl_create(&spa->spa_allocs[i].spaa_tree, zio_bookmark_compare,
 		    sizeof (zio_t), offsetof(zio_t, io_alloc_node));
 	}
 	avl_create(&spa->spa_metaslabs_by_flushed, metaslab_sort_by_flushed,
@@ -799,13 +798,11 @@ spa_remove(spa_t *spa)
 	}
 
 	for (int i = 0; i < spa->spa_alloc_count; i++) {
-		avl_destroy(&spa->spa_alloc_trees[i]);
-		mutex_destroy(&spa->spa_alloc_locks[i]);
+		avl_destroy(&spa->spa_allocs[i].spaa_tree);
+		mutex_destroy(&spa->spa_allocs[i].spaa_lock);
 	}
-	kmem_free(spa->spa_alloc_locks, spa->spa_alloc_count *
-	    sizeof (kmutex_t));
-	kmem_free(spa->spa_alloc_trees, spa->spa_alloc_count *
-	    sizeof (avl_tree_t));
+	kmem_free(spa->spa_allocs, spa->spa_alloc_count *
+	    sizeof (spa_alloc_t));
 
 	avl_destroy(&spa->spa_metaslabs_by_flushed);
 	avl_destroy(&spa->spa_sm_logs_by_txg);

module/zfs/zio.c

@@ -877,8 +877,7 @@ zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
 		zio->io_bookmark = *zb;
 
 	if (pio != NULL) {
-		if (zio->io_metaslab_class == NULL)
-			zio->io_metaslab_class = pio->io_metaslab_class;
+		zio->io_metaslab_class = pio->io_metaslab_class;
 		if (zio->io_logical == NULL)
 			zio->io_logical = pio->io_logical;
 		if (zio->io_child_type == ZIO_CHILD_GANG)
@@ -3380,9 +3379,9 @@ zio_io_to_allocate(spa_t *spa, int allocator)
 {
 	zio_t *zio;
 
-	ASSERT(MUTEX_HELD(&spa->spa_alloc_locks[allocator]));
+	ASSERT(MUTEX_HELD(&spa->spa_allocs[allocator].spaa_lock));
 
-	zio = avl_first(&spa->spa_alloc_trees[allocator]);
+	zio = avl_first(&spa->spa_allocs[allocator].spaa_tree);
 	if (zio == NULL)
 		return (NULL);
 
@@ -3394,11 +3393,11 @@ zio_io_to_allocate(spa_t *spa, int allocator)
 	 */
 	ASSERT3U(zio->io_allocator, ==, allocator);
 	if (!metaslab_class_throttle_reserve(zio->io_metaslab_class,
-	    zio->io_prop.zp_copies, zio->io_allocator, zio, 0)) {
+	    zio->io_prop.zp_copies, allocator, zio, 0)) {
 		return (NULL);
 	}
 
-	avl_remove(&spa->spa_alloc_trees[allocator], zio);
+	avl_remove(&spa->spa_allocs[allocator].spaa_tree, zio);
 	ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
 
 	return (zio);
@@ -3422,8 +3421,8 @@ zio_dva_throttle(zio_t *zio)
 		return (zio);
 	}
 
+	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
 	ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
-
 	ASSERT3U(zio->io_queued_timestamp, >, 0);
 	ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
 
@@ -3435,14 +3434,14 @@ zio_dva_throttle(zio_t *zio)
 	 * into 2^20 block regions, and then hash based on the objset, object,
 	 * level, and region to accomplish both of these goals.
 	 */
-	zio->io_allocator = cityhash4(bm->zb_objset, bm->zb_object,
+	int allocator = (uint_t)cityhash4(bm->zb_objset, bm->zb_object,
 	    bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count;
-	mutex_enter(&spa->spa_alloc_locks[zio->io_allocator]);
-	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
+	zio->io_allocator = allocator;
 	zio->io_metaslab_class = mc;
-	avl_add(&spa->spa_alloc_trees[zio->io_allocator], zio);
-	nio = zio_io_to_allocate(spa, zio->io_allocator);
-	mutex_exit(&spa->spa_alloc_locks[zio->io_allocator]);
+	mutex_enter(&spa->spa_allocs[allocator].spaa_lock);
+	avl_add(&spa->spa_allocs[allocator].spaa_tree, zio);
+	nio = zio_io_to_allocate(spa, allocator);
+	mutex_exit(&spa->spa_allocs[allocator].spaa_lock);
 	return (nio);
 }
 
@@ -3451,9 +3450,9 @@ zio_allocate_dispatch(spa_t *spa, int allocator)
 {
 	zio_t *zio;
 
-	mutex_enter(&spa->spa_alloc_locks[allocator]);
+	mutex_enter(&spa->spa_allocs[allocator].spaa_lock);
 	zio = zio_io_to_allocate(spa, allocator);
-	mutex_exit(&spa->spa_alloc_locks[allocator]);
+	mutex_exit(&spa->spa_allocs[allocator].spaa_lock);
 	if (zio == NULL)
 		return;
 
@@ -3643,8 +3642,8 @@ zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
 	 * some parallelism.
 	 */
 	int flags = METASLAB_FASTWRITE | METASLAB_ZIL;
-	int allocator = cityhash4(0, 0, 0, os->os_dsl_dataset->ds_object) %
-	    spa->spa_alloc_count;
+	int allocator = (uint_t)cityhash4(0, 0, 0,
+	    os->os_dsl_dataset->ds_object) % spa->spa_alloc_count;
 	error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
 	    txg, NULL, flags, &io_alloc_list, NULL, allocator);
 	*slog = (error == 0);