OpenZFS 7303 - dynamic metaslab selection

This change introduces a new weighting algorithm to improve
metaslab selection. The new weighting algorithm relies on the
SPACEMAP_HISTOGRAM feature. As a result, the metaslab weight
now encodes the type of weighting algorithm used (size-based
vs segment-based).

Porting Notes: The metaslab allocation tracing code is conditionally
removed on linux (dependent on mdb debugger).

Authored by: George Wilson <george.wilson@delphix.com>
Reviewed by: Alex Reece <alex@delphix.com>
Reviewed by: Chris Siden <christopher.siden@delphix.com>
Reviewed by: Dan Kimmel <dan.kimmel@delphix.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Paul Dagnelie <paul.dagnelie@delphix.com>
Reviewed by: Pavel Zakharov pavel.zakharov@delphix.com
Reviewed by: Prakash Surya <prakash.surya@delphix.com>
Reviewed by: Don Brady <don.brady@intel.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Ported-by: Don Brady <don.brady@intel.com>

OpenZFS-issue: https://www.illumos.org/issues/7303
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/d5190931bd
Closes #5404

cmd/zdb/zdb.c

@@ -2643,10 +2643,21 @@ zdb_leak_init(spa_t *spa, zdb_cb_t *zcb)
 
 	if (!dump_opt['L']) {
 		vdev_t *rvd = spa->spa_root_vdev;
+
+		/*
+		 * We are going to be changing the meaning of the metaslab's
+		 * ms_tree.  Ensure that the allocator doesn't try to
+		 * use the tree.
+		 */
+		spa->spa_normal_class->mc_ops = &zdb_metaslab_ops;
+		spa->spa_log_class->mc_ops = &zdb_metaslab_ops;
+
 		for (c = 0; c < rvd->vdev_children; c++) {
 			vdev_t *vd = rvd->vdev_child[c];
+			metaslab_group_t *mg = vd->vdev_mg;
 			for (m = 0; m < vd->vdev_ms_count; m++) {
 				metaslab_t *msp = vd->vdev_ms[m];
+				ASSERT3P(msp->ms_group, ==, mg);
 				mutex_enter(&msp->ms_lock);
 				metaslab_unload(msp);
 
@@ -2667,8 +2678,6 @@ zdb_leak_init(spa_t *spa, zdb_cb_t *zcb)
 					    (longlong_t)m,
 					    (longlong_t)vd->vdev_ms_count);
 
-					msp->ms_ops = &zdb_metaslab_ops;
-
 					/*
 					 * We don't want to spend the CPU
 					 * manipulating the size-ordered
@@ -2678,7 +2687,9 @@ zdb_leak_init(spa_t *spa, zdb_cb_t *zcb)
 					msp->ms_tree->rt_ops = NULL;
 					VERIFY0(space_map_load(msp->ms_sm,
 					    msp->ms_tree, SM_ALLOC));
-					msp->ms_loaded = B_TRUE;
+
+					if (!msp->ms_loaded)
+						msp->ms_loaded = B_TRUE;
 				}
 				mutex_exit(&msp->ms_lock);
 			}
@@ -2702,8 +2713,10 @@ zdb_leak_fini(spa_t *spa)
 		vdev_t *rvd = spa->spa_root_vdev;
 		for (c = 0; c < rvd->vdev_children; c++) {
 			vdev_t *vd = rvd->vdev_child[c];
+			metaslab_group_t *mg = vd->vdev_mg;
 			for (m = 0; m < vd->vdev_ms_count; m++) {
 				metaslab_t *msp = vd->vdev_ms[m];
+				ASSERT3P(mg, ==, msp->ms_group);
 				mutex_enter(&msp->ms_lock);
 
 				/*
@@ -2717,7 +2730,9 @@ zdb_leak_fini(spa_t *spa)
 				 * from the ms_tree.
 				 */
 				range_tree_vacate(msp->ms_tree, zdb_leak, vd);
-				msp->ms_loaded = B_FALSE;
+
+				if (msp->ms_loaded)
+					msp->ms_loaded = B_FALSE;
 
 				mutex_exit(&msp->ms_lock);
 			}

cmd/ztest/ztest.c

@@ -178,7 +178,7 @@ static const ztest_shared_opts_t ztest_opts_defaults = {
 	.zo_mirrors = 2,
 	.zo_raidz = 4,
 	.zo_raidz_parity = 1,
-	.zo_vdev_size = SPA_MINDEVSIZE * 2,
+	.zo_vdev_size = SPA_MINDEVSIZE * 4,	/* 256m default size */
 	.zo_datasets = 7,
 	.zo_threads = 23,
 	.zo_passtime = 60,		/* 60 seconds */

include/sys/metaslab.h

@@ -20,7 +20,7 @@
  */
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
- * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
+ * Copyright (c) 2011, 2016 by Delphix. All rights reserved.
  */
 
 #ifndef _SYS_METASLAB_H
@@ -36,10 +36,12 @@
 extern "C" {
 #endif
 
+
 typedef struct metaslab_ops {
-	uint64_t (*msop_alloc)(metaslab_t *msp, uint64_t size);
+	uint64_t (*msop_alloc)(metaslab_t *, uint64_t);
 } metaslab_ops_t;
 
+
 extern metaslab_ops_t *zfs_metaslab_ops;
 
 int metaslab_init(metaslab_group_t *, uint64_t, uint64_t, uint64_t,
@@ -64,13 +66,18 @@ uint64_t metaslab_block_maxsize(metaslab_t *);
 #define	METASLAB_FASTWRITE		0x20
 
 int metaslab_alloc(spa_t *, metaslab_class_t *, uint64_t,
-    blkptr_t *, int, uint64_t, blkptr_t *, int, zio_t *);
+    blkptr_t *, int, uint64_t, blkptr_t *, int, zio_alloc_list_t *, zio_t *);
 void metaslab_free(spa_t *, const blkptr_t *, uint64_t, boolean_t);
 int metaslab_claim(spa_t *, const blkptr_t *, uint64_t);
 void metaslab_check_free(spa_t *, const blkptr_t *);
 void metaslab_fastwrite_mark(spa_t *, const blkptr_t *);
 void metaslab_fastwrite_unmark(spa_t *, const blkptr_t *);
 
+void metaslab_alloc_trace_init(void);
+void metaslab_alloc_trace_fini(void);
+void metaslab_trace_init(zio_alloc_list_t *);
+void metaslab_trace_fini(zio_alloc_list_t *);
+
 metaslab_class_t *metaslab_class_create(spa_t *, metaslab_ops_t *);
 void metaslab_class_destroy(metaslab_class_t *);
 int metaslab_class_validate(metaslab_class_t *);

include/sys/metaslab_impl.h

@@ -24,7 +24,7 @@
  */
 
 /*
- * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
+ * Copyright (c) 2011, 2016 by Delphix. All rights reserved.
  */
 
 #ifndef _SYS_METASLAB_IMPL_H
@@ -41,6 +41,94 @@
 extern "C" {
 #endif
 
+/*
+ * Metaslab allocation tracing record.
+ */
+typedef struct metaslab_alloc_trace {
+	list_node_t			mat_list_node;
+	metaslab_group_t		*mat_mg;
+	metaslab_t			*mat_msp;
+	uint64_t			mat_size;
+	uint64_t			mat_weight;
+	uint32_t			mat_dva_id;
+	uint64_t			mat_offset;
+} metaslab_alloc_trace_t;
+
+/*
+ * Used by the metaslab allocation tracing facility to indicate
+ * error conditions. These errors are stored to the offset member
+ * of the metaslab_alloc_trace_t record and displayed by mdb.
+ */
+typedef enum trace_alloc_type {
+	TRACE_ALLOC_FAILURE	= -1ULL,
+	TRACE_TOO_SMALL		= -2ULL,
+	TRACE_FORCE_GANG	= -3ULL,
+	TRACE_NOT_ALLOCATABLE	= -4ULL,
+	TRACE_GROUP_FAILURE	= -5ULL,
+	TRACE_ENOSPC		= -6ULL,
+	TRACE_CONDENSING	= -7ULL,
+	TRACE_VDEV_ERROR	= -8ULL
+} trace_alloc_type_t;
+
+#define	METASLAB_WEIGHT_PRIMARY		(1ULL << 63)
+#define	METASLAB_WEIGHT_SECONDARY	(1ULL << 62)
+#define	METASLAB_WEIGHT_TYPE		(1ULL << 61)
+#define	METASLAB_ACTIVE_MASK		\
+	(METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
+
+/*
+ * The metaslab weight is used to encode the amount of free space in a
+ * metaslab, such that the "best" metaslab appears first when sorting the
+ * metaslabs by weight. The weight (and therefore the "best" metaslab) can
+ * be determined in two different ways: by computing a weighted sum of all
+ * the free space in the metaslab (a space based weight) or by counting only
+ * the free segments of the largest size (a segment based weight). We prefer
+ * the segment based weight because it reflects how the free space is
+ * comprised, but we cannot always use it -- legacy pools do not have the
+ * space map histogram information necessary to determine the largest
+ * contiguous regions. Pools that have the space map histogram determine
+ * the segment weight by looking at each bucket in the histogram and
+ * determining the free space whose size in bytes is in the range:
+ *	[2^i, 2^(i+1))
+ * We then encode the largest index, i, that contains regions into the
+ * segment-weighted value.
+ *
+ * Space-based weight:
+ *
+ *      64      56      48      40      32      24      16      8       0
+ *      +-------+-------+-------+-------+-------+-------+-------+-------+
+ *      |PS1|                   weighted-free space                     |
+ *      +-------+-------+-------+-------+-------+-------+-------+-------+
+ *
+ *	PS - indicates primary and secondary activation
+ *	space - the fragmentation-weighted space
+ *
+ * Segment-based weight:
+ *
+ *      64      56      48      40      32      24      16      8       0
+ *      +-------+-------+-------+-------+-------+-------+-------+-------+
+ *      |PS0| idx|             count of segments in region              |
+ *      +-------+-------+-------+-------+-------+-------+-------+-------+
+ *
+ *	PS - indicates primary and secondary activation
+ *	idx - index for the highest bucket in the histogram
+ *	count - number of segments in the specified bucket
+ */
+#define	WEIGHT_GET_ACTIVE(weight)		BF64_GET((weight), 62, 2)
+#define	WEIGHT_SET_ACTIVE(weight, x)		BF64_SET((weight), 62, 2, x)
+
+#define	WEIGHT_IS_SPACEBASED(weight)		\
+	((weight) == 0 || BF64_GET((weight), 61, 1))
+#define	WEIGHT_SET_SPACEBASED(weight)		BF64_SET((weight), 61, 1, 1)
+
+/*
+ * These macros are only applicable to segment-based weighting.
+ */
+#define	WEIGHT_GET_INDEX(weight)		BF64_GET((weight), 55, 6)
+#define	WEIGHT_SET_INDEX(weight, x)		BF64_SET((weight), 55, 6, x)
+#define	WEIGHT_GET_COUNT(weight)		BF64_GET((weight), 0, 55)
+#define	WEIGHT_SET_COUNT(weight, x)		BF64_SET((weight), 0, 55, x)
+
 /*
  * A metaslab class encompasses a category of allocatable top-level vdevs.
  * Each top-level vdev is associated with a metaslab group which defines
@@ -220,7 +308,6 @@ struct metaslab {
 	kmutex_t	ms_lock;
 	kcondvar_t	ms_load_cv;
 	space_map_t	*ms_sm;
-	metaslab_ops_t	*ms_ops;
 	uint64_t	ms_id;
 	uint64_t	ms_start;
 	uint64_t	ms_size;
@@ -233,12 +320,27 @@ struct metaslab {
 
 	boolean_t	ms_condensing;	/* condensing? */
 	boolean_t	ms_condense_wanted;
+
+	/*
+	 * We must hold both ms_lock and ms_group->mg_lock in order to
+	 * modify ms_loaded.
+	 */
 	boolean_t	ms_loaded;
 	boolean_t	ms_loading;
 
 	int64_t		ms_deferspace;	/* sum of ms_defermap[] space	*/
 	uint64_t	ms_weight;	/* weight vs. others in group	*/
-	uint64_t	ms_access_txg;
+	uint64_t	ms_activation_weight;	/* activation weight	*/
+
+	/*
+	 * Track of whenever a metaslab is selected for loading or allocation.
+	 * We use this value to determine how long the metaslab should
+	 * stay cached.
+	 */
+	uint64_t	ms_selected_txg;
+
+	uint64_t	ms_alloc_txg;	/* last successful alloc (debug only) */
+	uint64_t	ms_max_size;	/* maximum allocatable size	*/
 
 	/*
 	 * The metaslab block allocators can optionally use a size-ordered

include/sys/zfs_context.h

@@ -394,6 +394,7 @@ extern void cv_broadcast(kcondvar_t *cv);
  */
 extern kstat_t *kstat_create(const char *, int,
     const char *, const char *, uchar_t, ulong_t, uchar_t);
+extern void kstat_named_init(kstat_named_t *, const char *, uchar_t);
 extern void kstat_install(kstat_t *);
 extern void kstat_delete(kstat_t *);
 extern void kstat_waitq_enter(kstat_io_t *);

include/sys/zfs_debug.h

@@ -42,14 +42,15 @@ extern int zfs_flags;
 extern int zfs_recover;
 extern int zfs_free_leak_on_eio;
 
-#define	ZFS_DEBUG_DPRINTF		(1<<0)
+#define	ZFS_DEBUG_DPRINTF		(1 << 0)
-#define	ZFS_DEBUG_DBUF_VERIFY		(1<<1)
+#define	ZFS_DEBUG_DBUF_VERIFY		(1 << 1)
-#define	ZFS_DEBUG_DNODE_VERIFY		(1<<2)
+#define	ZFS_DEBUG_DNODE_VERIFY		(1 << 2)
-#define	ZFS_DEBUG_SNAPNAMES		(1<<3)
+#define	ZFS_DEBUG_SNAPNAMES		(1 << 3)
-#define	ZFS_DEBUG_MODIFY		(1<<4)
+#define	ZFS_DEBUG_MODIFY		(1 << 4)
-#define	ZFS_DEBUG_SPA			(1<<5)
+#define	ZFS_DEBUG_SPA			(1 << 5)
-#define	ZFS_DEBUG_ZIO_FREE		(1<<6)
+#define	ZFS_DEBUG_ZIO_FREE		(1 << 6)
-#define	ZFS_DEBUG_HISTOGRAM_VERIFY	(1<<7)
+#define	ZFS_DEBUG_HISTOGRAM_VERIFY	(1 << 7)
+#define	ZFS_DEBUG_METASLAB_VERIFY	(1 << 8)
 
 extern void __dprintf(const char *file, const char *func,
     int line, const char *fmt, ...);

include/sys/zio.h

@@ -358,6 +358,11 @@ typedef int zio_pipe_stage_t(zio_t *zio);
 #define	ZIO_REEXECUTE_NOW	0x01
 #define	ZIO_REEXECUTE_SUSPEND	0x02
 
+typedef struct zio_alloc_list {
+	list_t  zal_list;
+	uint64_t zal_size;
+} zio_alloc_list_t;
+
 typedef struct zio_link {
 	zio_t		*zl_parent;
 	zio_t		*zl_child;
@@ -417,6 +422,7 @@ struct zio {
 	avl_node_t	io_queue_node;
 	avl_node_t	io_offset_node;
 	avl_node_t	io_alloc_node;
+	zio_alloc_list_t 	io_alloc_list;
 
 	/* Internal pipeline state */
 	enum zio_flag	io_flags;

man/man5/zfs-module-parameters.5

@@ -179,6 +179,30 @@ relative to the pool.
 Use \fB1\fR for yes (default) and \fB0\fR for no.
 .RE
 
+.sp
+.ne 2
+.na
+\fBzfs_metaslab_segment_weight_enabled\fR (int)
+.ad
+.RS 12n
+Enable/disable segment-based metaslab selection.
+.sp
+Use \fB1\fR for yes (default) and \fB0\fR for no.
+.RE
+
+.sp
+.ne 2
+.na
+\fBzfs_metaslab_switch_threshold\fR (int)
+.ad
+.RS 12n
+When using segment-based metaslab selection, continue allocating
+from the active metaslab until \fBlzfs_metaslab_switch_threshold\fR
+worth of buckets have been exhausted.
+.sp
+Default value: \fB2\fR.
+.RE
+
 .sp
 .ne 2
 .na

module/zfs/spa.c

@@ -1313,7 +1313,7 @@ spa_config_parse(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent,
 static void
 spa_unload(spa_t *spa)
 {
-	int i;
+	int i, c;
 
 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
 
@@ -1330,6 +1330,19 @@ spa_unload(spa_t *spa)
 		spa->spa_sync_on = B_FALSE;
 	}
 
+	/*
+	 * Even though vdev_free() also calls vdev_metaslab_fini, we need
+	 * to call it earlier, before we wait for async i/o to complete.
+	 * This ensures that there is no async metaslab prefetching, by
+	 * calling taskq_wait(mg_taskq).
+	 */
+	if (spa->spa_root_vdev != NULL) {
+		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
+		for (c = 0; c < spa->spa_root_vdev->vdev_children; c++)
+			vdev_metaslab_fini(spa->spa_root_vdev->vdev_child[c]);
+		spa_config_exit(spa, SCL_ALL, FTAG);
+	}
+
 	/*
 	 * Wait for any outstanding async I/O to complete.
 	 */

module/zfs/spa_misc.c

@@ -1833,6 +1833,7 @@ spa_init(int mode)
 	refcount_init();
 	unique_init();
 	range_tree_init();
+	metaslab_alloc_trace_init();
 	ddt_init();
 	zio_init();
 	dmu_init();
@@ -1861,6 +1862,7 @@ spa_fini(void)
 	dmu_fini();
 	zio_fini();
 	ddt_fini();
+	metaslab_alloc_trace_fini();
 	range_tree_fini();
 	unique_fini();
 	refcount_fini();

module/zfs/space_map.c

@@ -173,7 +173,6 @@ space_map_histogram_add(space_map_t *sm, range_tree_t *rt, dmu_tx_t *tx)
 	dmu_buf_will_dirty(sm->sm_dbuf, tx);
 
 	ASSERT(space_map_histogram_verify(sm, rt));
-
 	/*
 	 * Transfer the content of the range tree histogram to the space
 	 * map histogram. The space map histogram contains 32 buckets ranging

module/zfs/zio.c

@@ -596,6 +596,7 @@ zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
 	    offsetof(zio_link_t, zl_parent_node));
 	list_create(&zio->io_child_list, sizeof (zio_link_t),
 	    offsetof(zio_link_t, zl_child_node));
+	metaslab_trace_init(&zio->io_alloc_list);
 
 	if (vd != NULL)
 		zio->io_child_type = ZIO_CHILD_VDEV;
@@ -657,6 +658,7 @@ zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
 static void
 zio_destroy(zio_t *zio)
 {
+	metaslab_trace_fini(&zio->io_alloc_list);
 	list_destroy(&zio->io_parent_list);
 	list_destroy(&zio->io_child_list);
 	mutex_destroy(&zio->io_lock);
@@ -2299,7 +2301,8 @@ zio_write_gang_block(zio_t *pio)
 	}
 
 	error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
-	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, pio);
+	    bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
+	    &pio->io_alloc_list, pio);
 	if (error) {
 		if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
 			ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
@@ -3011,7 +3014,8 @@ zio_dva_allocate(zio_t *zio)
 		flags |= METASLAB_ASYNC_ALLOC;
 
 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
-	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags, zio);
+	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
+	    &zio->io_alloc_list, zio);
 
 	if (error != 0) {
 		spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
@@ -3077,18 +3081,24 @@ zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, uint64_t size,
     boolean_t use_slog)
 {
 	int error = 1;
+	zio_alloc_list_t io_alloc_list;
 
 	ASSERT(txg > spa_syncing_txg(spa));
 
+	metaslab_trace_init(&io_alloc_list);
+
 	if (use_slog) {
 		error = metaslab_alloc(spa, spa_log_class(spa), size,
-		    new_bp, 1, txg, NULL, METASLAB_FASTWRITE, NULL);
+		    new_bp, 1, txg, NULL, METASLAB_FASTWRITE,
+		    &io_alloc_list, NULL);
 	}
 
 	if (error) {
 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
-		    new_bp, 1, txg, NULL, METASLAB_FASTWRITE, NULL);
+		    new_bp, 1, txg, NULL, METASLAB_FASTWRITE,
+		    &io_alloc_list, NULL);
 	}
+	metaslab_trace_fini(&io_alloc_list);
 
 	if (error == 0) {
 		BP_SET_LSIZE(new_bp, size);