svn commit: r315291 - in vendor-sys/illumos/dist/uts/common/fs/zfs: . sys

Josh Paetzel jpaetzel at FreeBSD.org
Wed Mar 15 04:18:42 UTC 2017


Author: jpaetzel
Date: Wed Mar 15 04:18:40 2017
New Revision: 315291
URL: https://svnweb.freebsd.org/changeset/base/315291

Log:
  7303 dynamic metaslab selection
  
  illumos/illumos-gate at 8363e80ae72609660f6090766ca8c2c18aa53f0c
  https://github.com/illumos/illumos-gate/commit/8363e80ae72609660f6090766ca8c2c18aa53f0
  
  https://www.illumos.org/issues/7303
  
    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).
  
    This also introduce a new allocation tracing facility and two new dcmds to help
    debug allocation problems. Each zio now contains a zio_alloc_list_t structure
    that is populated as the zio goes through the allocations stage. Here's an
    example of how to use the tracing facility:
  
  > c5ec000::print zio_t io_alloc_list | ::walk list | ::metaslab_trace
    MSID    DVA    ASIZE      WEIGHT             RESULT               VDEV
       -      0      400           0    NOT_ALLOCATABLE           ztest.0a
       -      0      400           0    NOT_ALLOCATABLE           ztest.0a
       -      0      400           0             ENOSPC           ztest.0a
       -      0      200           0    NOT_ALLOCATABLE           ztest.0a
       -      0      200           0    NOT_ALLOCATABLE           ztest.0a
       -      0      200           0             ENOSPC           ztest.0a
       1      0      400      1 x 8M            17b1a00           ztest.0a
  
  > 1ff2400::print zio_t io_alloc_list | ::walk list | ::metaslab_trace
    MSID    DVA    ASIZE      WEIGHT             RESULT               VDEV
       -      0      200           0    NOT_ALLOCATABLE           mirror-2
       -      0      200           0    NOT_ALLOCATABLE           mirror-0
       1      0      200      1 x 4M            112ae00           mirror-1
       -      1      200           0    NOT_ALLOCATABLE           mirror-2
       -      1      200           0    NOT_ALLOCATABLE           mirror-0
       1      1      200      1 x 4M            112b000           mirror-1
       -      2      200           0    NOT_ALLOCATABLE           mirror-2
  
    If the metaslab is using segment-based weighting then the WEIGHT column will
    display the number of segments available in the bucket where the allocation
    attempt was made.
  
  Author: George Wilson <george.wilson at delphix.com>
  Reviewed by: Alex Reece <alex at delphix.com>
  Reviewed by: Chris Siden <christopher.siden at delphix.com>
  Reviewed by: Dan Kimmel <dan.kimmel at delphix.com>
  Reviewed by: Matthew Ahrens <mahrens at delphix.com>
  Reviewed by: Paul Dagnelie <paul.dagnelie at delphix.com>
  Reviewed by: Pavel Zakharov <pavel.zakharov at delphix.com>
  Reviewed by: Prakash Surya <prakash.surya at delphix.com>
  Reviewed by: Don Brady <don.brady at intel.com>
  Approved by: Richard Lowe <richlowe at richlowe.net>

Modified:
  vendor-sys/illumos/dist/uts/common/fs/zfs/metaslab.c
  vendor-sys/illumos/dist/uts/common/fs/zfs/spa.c
  vendor-sys/illumos/dist/uts/common/fs/zfs/spa_misc.c
  vendor-sys/illumos/dist/uts/common/fs/zfs/space_map.c
  vendor-sys/illumos/dist/uts/common/fs/zfs/sys/metaslab.h
  vendor-sys/illumos/dist/uts/common/fs/zfs/sys/metaslab_impl.h
  vendor-sys/illumos/dist/uts/common/fs/zfs/sys/zfs_debug.h
  vendor-sys/illumos/dist/uts/common/fs/zfs/sys/zio.h
  vendor-sys/illumos/dist/uts/common/fs/zfs/zio.c

Modified: vendor-sys/illumos/dist/uts/common/fs/zfs/metaslab.c
==============================================================================
--- vendor-sys/illumos/dist/uts/common/fs/zfs/metaslab.c	Wed Mar 15 04:16:08 2017	(r315290)
+++ vendor-sys/illumos/dist/uts/common/fs/zfs/metaslab.c	Wed Mar 15 04:18:40 2017	(r315291)
@@ -38,18 +38,13 @@
 #define	GANG_ALLOCATION(flags) \
 	((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
 
-#define	METASLAB_WEIGHT_PRIMARY		(1ULL << 63)
-#define	METASLAB_WEIGHT_SECONDARY	(1ULL << 62)
-#define	METASLAB_ACTIVE_MASK		\
-	(METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
-
 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.
+ * 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;
@@ -122,7 +117,7 @@ uint64_t metaslab_df_alloc_threshold = S
 /*
  * The minimum free space, in percent, which must be available
  * in a space map to continue allocations in a first-fit fashion.
- * Once the space_map's free space drops below this level we dynamically
+ * Once the space map's free space drops below this level we dynamically
  * switch to using best-fit allocations.
  */
 int metaslab_df_free_pct = 4;
@@ -170,7 +165,38 @@ boolean_t metaslab_lba_weighting_enabled
  */
 boolean_t metaslab_bias_enabled = B_TRUE;
 
-static uint64_t metaslab_fragmentation(metaslab_t *);
+/*
+ * Enable/disable segment-based metaslab selection.
+ */
+boolean_t zfs_metaslab_segment_weight_enabled = B_TRUE;
+
+/*
+ * When using segment-based metaslab selection, we will continue
+ * allocating from the active metaslab until we have exhausted
+ * zfs_metaslab_switch_threshold of its buckets.
+ */
+int zfs_metaslab_switch_threshold = 2;
+
+/*
+ * Internal switch to enable/disable the metaslab allocation tracing
+ * facility.
+ */
+boolean_t metaslab_trace_enabled = B_TRUE;
+
+/*
+ * Maximum entries that the metaslab allocation tracing facility will keep
+ * in a given list when running in non-debug mode. We limit the number
+ * of entries in non-debug mode to prevent us from using up too much memory.
+ * The limit should be sufficiently large that we don't expect any allocation
+ * to every exceed this value. In debug mode, the system will panic if this
+ * limit is ever reached allowing for further investigation.
+ */
+uint64_t metaslab_trace_max_entries = 5000;
+
+static uint64_t metaslab_weight(metaslab_t *);
+static void metaslab_set_fragmentation(metaslab_t *);
+
+kmem_cache_t *metaslab_alloc_trace_cache;
 
 /*
  * ==========================================================================
@@ -388,11 +414,6 @@ metaslab_class_expandable_space(metaslab
 	return (space);
 }
 
-/*
- * ==========================================================================
- * Metaslab groups
- * ==========================================================================
- */
 static int
 metaslab_compare(const void *x1, const void *x2)
 {
@@ -418,6 +439,57 @@ metaslab_compare(const void *x1, const v
 }
 
 /*
+ * Verify that the space accounting on disk matches the in-core range_trees.
+ */
+void
+metaslab_verify_space(metaslab_t *msp, uint64_t txg)
+{
+	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
+	uint64_t allocated = 0;
+	uint64_t freed = 0;
+	uint64_t sm_free_space, msp_free_space;
+
+	ASSERT(MUTEX_HELD(&msp->ms_lock));
+
+	if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
+		return;
+
+	/*
+	 * We can only verify the metaslab space when we're called
+	 * from syncing context with a loaded metaslab that has an allocated
+	 * space map. Calling this in non-syncing context does not
+	 * provide a consistent view of the metaslab since we're performing
+	 * allocations in the future.
+	 */
+	if (txg != spa_syncing_txg(spa) || msp->ms_sm == NULL ||
+	    !msp->ms_loaded)
+		return;
+
+	sm_free_space = msp->ms_size - space_map_allocated(msp->ms_sm) -
+	    space_map_alloc_delta(msp->ms_sm);
+
+	/*
+	 * Account for future allocations since we would have already
+	 * deducted that space from the ms_freetree.
+	 */
+	for (int t = 0; t < TXG_CONCURRENT_STATES; t++) {
+		allocated +=
+		    range_tree_space(msp->ms_alloctree[(txg + t) & TXG_MASK]);
+	}
+	freed = range_tree_space(msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK]);
+
+	msp_free_space = range_tree_space(msp->ms_tree) + allocated +
+	    msp->ms_deferspace + freed;
+
+	VERIFY3U(sm_free_space, ==, msp_free_space);
+}
+
+/*
+ * ==========================================================================
+ * Metaslab groups
+ * ==========================================================================
+ */
+/*
  * Update the allocatable flag and the metaslab group's capacity.
  * The allocatable flag is set to true if the capacity is below
  * the zfs_mg_noalloc_threshold or has a fragmentation value that is
@@ -989,7 +1061,7 @@ static range_tree_ops_t metaslab_rt_ops 
 
 /*
  * ==========================================================================
- * Metaslab block operations
+ * Common allocator routines
  * ==========================================================================
  */
 
@@ -1008,31 +1080,22 @@ metaslab_block_maxsize(metaslab_t *msp)
 	return (rs->rs_end - rs->rs_start);
 }
 
-uint64_t
-metaslab_block_alloc(metaslab_t *msp, uint64_t size)
+static range_seg_t *
+metaslab_block_find(avl_tree_t *t, uint64_t start, uint64_t size)
 {
-	uint64_t start;
-	range_tree_t *rt = msp->ms_tree;
-
-	VERIFY(!msp->ms_condensing);
+	range_seg_t *rs, rsearch;
+	avl_index_t where;
 
-	start = msp->ms_ops->msop_alloc(msp, size);
-	if (start != -1ULL) {
-		vdev_t *vd = msp->ms_group->mg_vd;
+	rsearch.rs_start = start;
+	rsearch.rs_end = start + size;
 
-		VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
-		VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
-		VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
-		range_tree_remove(rt, start, size);
+	rs = avl_find(t, &rsearch, &where);
+	if (rs == NULL) {
+		rs = avl_nearest(t, where, AVL_AFTER);
 	}
-	return (start);
-}
 
-/*
- * ==========================================================================
- * Common allocator routines
- * ==========================================================================
- */
+	return (rs);
+}
 
 /*
  * This is a helper function that can be used by the allocator to find
@@ -1043,15 +1106,7 @@ static uint64_t
 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
     uint64_t align)
 {
-	range_seg_t *rs, rsearch;
-	avl_index_t where;
-
-	rsearch.rs_start = *cursor;
-	rsearch.rs_end = *cursor + size;
-
-	rs = avl_find(t, &rsearch, &where);
-	if (rs == NULL)
-		rs = avl_nearest(t, where, AVL_AFTER);
+	range_seg_t *rs = metaslab_block_find(t, *cursor, size);
 
 	while (rs != NULL) {
 		uint64_t offset = P2ROUNDUP(rs->rs_start, align);
@@ -1276,6 +1331,7 @@ int
 metaslab_load(metaslab_t *msp)
 {
 	int error = 0;
+	boolean_t success = B_FALSE;
 
 	ASSERT(MUTEX_HELD(&msp->ms_lock));
 	ASSERT(!msp->ms_loaded);
@@ -1293,14 +1349,18 @@ metaslab_load(metaslab_t *msp)
 	else
 		range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size);
 
-	msp->ms_loaded = (error == 0);
+	success = (error == 0);
 	msp->ms_loading = B_FALSE;
 
-	if (msp->ms_loaded) {
+	if (success) {
+		ASSERT3P(msp->ms_group, !=, NULL);
+		msp->ms_loaded = B_TRUE;
+
 		for (int t = 0; t < TXG_DEFER_SIZE; t++) {
 			range_tree_walk(msp->ms_defertree[t],
 			    range_tree_remove, msp->ms_tree);
 		}
+		msp->ms_max_size = metaslab_block_maxsize(msp);
 	}
 	cv_broadcast(&msp->ms_load_cv);
 	return (error);
@@ -1313,6 +1373,7 @@ metaslab_unload(metaslab_t *msp)
 	range_tree_vacate(msp->ms_tree, NULL, NULL);
 	msp->ms_loaded = B_FALSE;
 	msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
+	msp->ms_max_size = 0;
 }
 
 int
@@ -1357,21 +1418,23 @@ metaslab_init(metaslab_group_t *mg, uint
 	ms->ms_tree = range_tree_create(&metaslab_rt_ops, ms, &ms->ms_lock);
 	metaslab_group_add(mg, ms);
 
-	ms->ms_fragmentation = metaslab_fragmentation(ms);
-	ms->ms_ops = mg->mg_class->mc_ops;
+	metaslab_set_fragmentation(ms);
 
 	/*
 	 * If we're opening an existing pool (txg == 0) or creating
 	 * a new one (txg == TXG_INITIAL), all space is available now.
 	 * If we're adding space to an existing pool, the new space
 	 * does not become available until after this txg has synced.
+	 * The metaslab's weight will also be initialized when we sync
+	 * out this txg. This ensures that we don't attempt to allocate
+	 * from it before we have initialized it completely.
 	 */
 	if (txg <= TXG_INITIAL)
 		metaslab_sync_done(ms, 0);
 
 	/*
 	 * If metaslab_debug_load is set and we're initializing a metaslab
-	 * that has an allocated space_map object then load the its space
+	 * that has an allocated space map object then load the its space
 	 * map so that can verify frees.
 	 */
 	if (metaslab_debug_load && ms->ms_sm != NULL) {
@@ -1398,7 +1461,6 @@ metaslab_fini(metaslab_t *msp)
 	metaslab_group_remove(mg, msp);
 
 	mutex_enter(&msp->ms_lock);
-
 	VERIFY(msp->ms_group == NULL);
 	vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
 	    0, -msp->ms_size);
@@ -1471,8 +1533,8 @@ int zfs_frag_table[FRAGMENTATION_TABLE_S
  * not support this metric. Otherwise, the return value should be in the
  * range [0, 100].
  */
-static uint64_t
-metaslab_fragmentation(metaslab_t *msp)
+static void
+metaslab_set_fragmentation(metaslab_t *msp)
 {
 	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
 	uint64_t fragmentation = 0;
@@ -1480,18 +1542,22 @@ metaslab_fragmentation(metaslab_t *msp)
 	boolean_t feature_enabled = spa_feature_is_enabled(spa,
 	    SPA_FEATURE_SPACEMAP_HISTOGRAM);
 
-	if (!feature_enabled)
-		return (ZFS_FRAG_INVALID);
+	if (!feature_enabled) {
+		msp->ms_fragmentation = ZFS_FRAG_INVALID;
+		return;
+	}
 
 	/*
 	 * A null space map means that the entire metaslab is free
 	 * and thus is not fragmented.
 	 */
-	if (msp->ms_sm == NULL)
-		return (0);
+	if (msp->ms_sm == NULL) {
+		msp->ms_fragmentation = 0;
+		return;
+	}
 
 	/*
-	 * If this metaslab's space_map has not been upgraded, flag it
+	 * If this metaslab's space map has not been upgraded, flag it
 	 * so that we upgrade next time we encounter it.
 	 */
 	if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) {
@@ -1504,12 +1570,14 @@ metaslab_fragmentation(metaslab_t *msp)
 			spa_dbgmsg(spa, "txg %llu, requesting force condense: "
 			    "msp %p, vd %p", txg, msp, vd);
 		}
-		return (ZFS_FRAG_INVALID);
+		msp->ms_fragmentation = ZFS_FRAG_INVALID;
+		return;
 	}
 
 	for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
 		uint64_t space = 0;
 		uint8_t shift = msp->ms_sm->sm_shift;
+
 		int idx = MIN(shift - SPA_MINBLOCKSHIFT + i,
 		    FRAGMENTATION_TABLE_SIZE - 1);
 
@@ -1526,7 +1594,8 @@ metaslab_fragmentation(metaslab_t *msp)
 	if (total > 0)
 		fragmentation /= total;
 	ASSERT3U(fragmentation, <=, 100);
-	return (fragmentation);
+
+	msp->ms_fragmentation = fragmentation;
 }
 
 /*
@@ -1535,30 +1604,20 @@ metaslab_fragmentation(metaslab_t *msp)
  * the LBA range, and whether the metaslab is loaded.
  */
 static uint64_t
-metaslab_weight(metaslab_t *msp)
+metaslab_space_weight(metaslab_t *msp)
 {
 	metaslab_group_t *mg = msp->ms_group;
 	vdev_t *vd = mg->mg_vd;
 	uint64_t weight, space;
 
 	ASSERT(MUTEX_HELD(&msp->ms_lock));
-
-	/*
-	 * This vdev is in the process of being removed so there is nothing
-	 * for us to do here.
-	 */
-	if (vd->vdev_removing) {
-		ASSERT0(space_map_allocated(msp->ms_sm));
-		ASSERT0(vd->vdev_ms_shift);
-		return (0);
-	}
+	ASSERT(!vd->vdev_removing);
 
 	/*
 	 * The baseline weight is the metaslab's free space.
 	 */
 	space = msp->ms_size - space_map_allocated(msp->ms_sm);
 
-	msp->ms_fragmentation = metaslab_fragmentation(msp);
 	if (metaslab_fragmentation_factor_enabled &&
 	    msp->ms_fragmentation != ZFS_FRAG_INVALID) {
 		/*
@@ -1607,6 +1666,210 @@ metaslab_weight(metaslab_t *msp)
 		weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
 	}
 
+	WEIGHT_SET_SPACEBASED(weight);
+	return (weight);
+}
+
+/*
+ * Return the weight of the specified metaslab, according to the segment-based
+ * weighting algorithm. The metaslab must be loaded. This function can
+ * be called within a sync pass since it relies only on the metaslab's
+ * range tree which is always accurate when the metaslab is loaded.
+ */
+static uint64_t
+metaslab_weight_from_range_tree(metaslab_t *msp)
+{
+	uint64_t weight = 0;
+	uint32_t segments = 0;
+
+	ASSERT(msp->ms_loaded);
+
+	for (int i = RANGE_TREE_HISTOGRAM_SIZE - 1; i >= SPA_MINBLOCKSHIFT;
+	    i--) {
+		uint8_t shift = msp->ms_group->mg_vd->vdev_ashift;
+		int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
+
+		segments <<= 1;
+		segments += msp->ms_tree->rt_histogram[i];
+
+		/*
+		 * The range tree provides more precision than the space map
+		 * and must be downgraded so that all values fit within the
+		 * space map's histogram. This allows us to compare loaded
+		 * vs. unloaded metaslabs to determine which metaslab is
+		 * considered "best".
+		 */
+		if (i > max_idx)
+			continue;
+
+		if (segments != 0) {
+			WEIGHT_SET_COUNT(weight, segments);
+			WEIGHT_SET_INDEX(weight, i);
+			WEIGHT_SET_ACTIVE(weight, 0);
+			break;
+		}
+	}
+	return (weight);
+}
+
+/*
+ * Calculate the weight based on the on-disk histogram. This should only
+ * be called after a sync pass has completely finished since the on-disk
+ * information is updated in metaslab_sync().
+ */
+static uint64_t
+metaslab_weight_from_spacemap(metaslab_t *msp)
+{
+	uint64_t weight = 0;
+
+	for (int i = SPACE_MAP_HISTOGRAM_SIZE - 1; i >= 0; i--) {
+		if (msp->ms_sm->sm_phys->smp_histogram[i] != 0) {
+			WEIGHT_SET_COUNT(weight,
+			    msp->ms_sm->sm_phys->smp_histogram[i]);
+			WEIGHT_SET_INDEX(weight, i +
+			    msp->ms_sm->sm_shift);
+			WEIGHT_SET_ACTIVE(weight, 0);
+			break;
+		}
+	}
+	return (weight);
+}
+
+/*
+ * Compute a segment-based weight for the specified metaslab. The weight
+ * is determined by highest bucket in the histogram. The information
+ * for the highest bucket is encoded into the weight value.
+ */
+static uint64_t
+metaslab_segment_weight(metaslab_t *msp)
+{
+	metaslab_group_t *mg = msp->ms_group;
+	uint64_t weight = 0;
+	uint8_t shift = mg->mg_vd->vdev_ashift;
+
+	ASSERT(MUTEX_HELD(&msp->ms_lock));
+
+	/*
+	 * The metaslab is completely free.
+	 */
+	if (space_map_allocated(msp->ms_sm) == 0) {
+		int idx = highbit64(msp->ms_size) - 1;
+		int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
+
+		if (idx < max_idx) {
+			WEIGHT_SET_COUNT(weight, 1ULL);
+			WEIGHT_SET_INDEX(weight, idx);
+		} else {
+			WEIGHT_SET_COUNT(weight, 1ULL << (idx - max_idx));
+			WEIGHT_SET_INDEX(weight, max_idx);
+		}
+		WEIGHT_SET_ACTIVE(weight, 0);
+		ASSERT(!WEIGHT_IS_SPACEBASED(weight));
+
+		return (weight);
+	}
+
+	ASSERT3U(msp->ms_sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t));
+
+	/*
+	 * If the metaslab is fully allocated then just make the weight 0.
+	 */
+	if (space_map_allocated(msp->ms_sm) == msp->ms_size)
+		return (0);
+	/*
+	 * If the metaslab is already loaded, then use the range tree to
+	 * determine the weight. Otherwise, we rely on the space map information
+	 * to generate the weight.
+	 */
+	if (msp->ms_loaded) {
+		weight = metaslab_weight_from_range_tree(msp);
+	} else {
+		weight = metaslab_weight_from_spacemap(msp);
+	}
+
+	/*
+	 * If the metaslab was active the last time we calculated its weight
+	 * then keep it active. We want to consume the entire region that
+	 * is associated with this weight.
+	 */
+	if (msp->ms_activation_weight != 0 && weight != 0)
+		WEIGHT_SET_ACTIVE(weight, WEIGHT_GET_ACTIVE(msp->ms_weight));
+	return (weight);
+}
+
+/*
+ * Determine if we should attempt to allocate from this metaslab. If the
+ * metaslab has a maximum size then we can quickly determine if the desired
+ * allocation size can be satisfied. Otherwise, if we're using segment-based
+ * weighting then we can determine the maximum allocation that this metaslab
+ * can accommodate based on the index encoded in the weight. If we're using
+ * space-based weights then rely on the entire weight (excluding the weight
+ * type bit).
+ */
+boolean_t
+metaslab_should_allocate(metaslab_t *msp, uint64_t asize)
+{
+	boolean_t should_allocate;
+
+	if (msp->ms_max_size != 0)
+		return (msp->ms_max_size >= asize);
+
+	if (!WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
+		/*
+		 * The metaslab segment weight indicates segments in the
+		 * range [2^i, 2^(i+1)), where i is the index in the weight.
+		 * Since the asize might be in the middle of the range, we
+		 * should attempt the allocation if asize < 2^(i+1).
+		 */
+		should_allocate = (asize <
+		    1ULL << (WEIGHT_GET_INDEX(msp->ms_weight) + 1));
+	} else {
+		should_allocate = (asize <=
+		    (msp->ms_weight & ~METASLAB_WEIGHT_TYPE));
+	}
+	return (should_allocate);
+}
+
+static uint64_t
+metaslab_weight(metaslab_t *msp)
+{
+	vdev_t *vd = msp->ms_group->mg_vd;
+	spa_t *spa = vd->vdev_spa;
+	uint64_t weight;
+
+	ASSERT(MUTEX_HELD(&msp->ms_lock));
+
+	/*
+	 * This vdev is in the process of being removed so there is nothing
+	 * for us to do here.
+	 */
+	if (vd->vdev_removing) {
+		ASSERT0(space_map_allocated(msp->ms_sm));
+		ASSERT0(vd->vdev_ms_shift);
+		return (0);
+	}
+
+	metaslab_set_fragmentation(msp);
+
+	/*
+	 * Update the maximum size if the metaslab is loaded. This will
+	 * ensure that we get an accurate maximum size if newly freed space
+	 * has been added back into the free tree.
+	 */
+	if (msp->ms_loaded)
+		msp->ms_max_size = metaslab_block_maxsize(msp);
+
+	/*
+	 * Segment-based weighting requires space map histogram support.
+	 */
+	if (zfs_metaslab_segment_weight_enabled &&
+	    spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
+	    (msp->ms_sm == NULL || msp->ms_sm->sm_dbuf->db_size ==
+	    sizeof (space_map_phys_t))) {
+		weight = metaslab_segment_weight(msp);
+	} else {
+		weight = metaslab_space_weight(msp);
+	}
 	return (weight);
 }
 
@@ -1625,6 +1888,7 @@ metaslab_activate(metaslab_t *msp, uint6
 			}
 		}
 
+		msp->ms_activation_weight = msp->ms_weight;
 		metaslab_group_sort(msp->ms_group, msp,
 		    msp->ms_weight | activation_weight);
 	}
@@ -1635,18 +1899,56 @@ metaslab_activate(metaslab_t *msp, uint6
 }
 
 static void
-metaslab_passivate(metaslab_t *msp, uint64_t size)
+metaslab_passivate(metaslab_t *msp, uint64_t weight)
 {
+	uint64_t size = weight & ~METASLAB_WEIGHT_TYPE;
+
 	/*
 	 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
 	 * this metaslab again.  In that case, it had better be empty,
 	 * or we would be leaving space on the table.
 	 */
-	ASSERT(size >= SPA_MINBLOCKSIZE || range_tree_space(msp->ms_tree) == 0);
-	metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
+	ASSERT(size >= SPA_MINBLOCKSIZE ||
+	    range_tree_space(msp->ms_tree) == 0);
+	ASSERT0(weight & METASLAB_ACTIVE_MASK);
+
+	msp->ms_activation_weight = 0;
+	metaslab_group_sort(msp->ms_group, msp, weight);
 	ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
 }
 
+/*
+ * Segment-based metaslabs are activated once and remain active until
+ * we either fail an allocation attempt (similar to space-based metaslabs)
+ * or have exhausted the free space in zfs_metaslab_switch_threshold
+ * buckets since the metaslab was activated. This function checks to see
+ * if we've exhaused the zfs_metaslab_switch_threshold buckets in the
+ * metaslab and passivates it proactively. This will allow us to select a
+ * metaslabs with larger contiguous region if any remaining within this
+ * metaslab group. If we're in sync pass > 1, then we continue using this
+ * metaslab so that we don't dirty more block and cause more sync passes.
+ */
+void
+metaslab_segment_may_passivate(metaslab_t *msp)
+{
+	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
+
+	if (WEIGHT_IS_SPACEBASED(msp->ms_weight) || spa_sync_pass(spa) > 1)
+		return;
+
+	/*
+	 * Since we are in the middle of a sync pass, the most accurate
+	 * information that is accessible to us is the in-core range tree
+	 * histogram; calculate the new weight based on that information.
+	 */
+	uint64_t weight = metaslab_weight_from_range_tree(msp);
+	int activation_idx = WEIGHT_GET_INDEX(msp->ms_activation_weight);
+	int current_idx = WEIGHT_GET_INDEX(weight);
+
+	if (current_idx <= activation_idx - zfs_metaslab_switch_threshold)
+		metaslab_passivate(msp, weight);
+}
+
 static void
 metaslab_preload(void *arg)
 {
@@ -1659,11 +1961,7 @@ metaslab_preload(void *arg)
 	metaslab_load_wait(msp);
 	if (!msp->ms_loaded)
 		(void) metaslab_load(msp);
-
-	/*
-	 * Set the ms_access_txg value so that we don't unload it right away.
-	 */
-	msp->ms_access_txg = spa_syncing_txg(spa) + metaslab_unload_delay + 1;
+	msp->ms_selected_txg = spa_syncing_txg(spa);
 	mutex_exit(&msp->ms_lock);
 }
 
@@ -1684,10 +1982,7 @@ metaslab_group_preload(metaslab_group_t 
 	/*
 	 * Load the next potential metaslabs
 	 */
-	msp = avl_first(t);
-	while (msp != NULL) {
-		metaslab_t *msp_next = AVL_NEXT(t, msp);
-
+	for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) {
 		/*
 		 * We preload only the maximum number of metaslabs specified
 		 * by metaslab_preload_limit. If a metaslab is being forced
@@ -1695,27 +1990,11 @@ metaslab_group_preload(metaslab_group_t 
 		 * that force condensing happens in the next txg.
 		 */
 		if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
-			msp = msp_next;
 			continue;
 		}
 
-		/*
-		 * We must drop the metaslab group lock here to preserve
-		 * lock ordering with the ms_lock (when grabbing both
-		 * the mg_lock and the ms_lock, the ms_lock must be taken
-		 * first).  As a result, it is possible that the ordering
-		 * of the metaslabs within the avl tree may change before
-		 * we reacquire the lock. The metaslab cannot be removed from
-		 * the tree while we're in syncing context so it is safe to
-		 * drop the mg_lock here. If the metaslabs are reordered
-		 * nothing will break -- we just may end up loading a
-		 * less than optimal one.
-		 */
-		mutex_exit(&mg->mg_lock);
 		VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
 		    msp, TQ_SLEEP) != NULL);
-		mutex_enter(&mg->mg_lock);
-		msp = msp_next;
 	}
 	mutex_exit(&mg->mg_lock);
 }
@@ -1864,7 +2143,7 @@ metaslab_condense(metaslab_t *msp, uint6
 	mutex_enter(&msp->ms_lock);
 
 	/*
-	 * While we would ideally like to create a space_map representation
+	 * 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 tree can be
 	 * large, and therefore computationally expensive to subtract
@@ -1927,7 +2206,7 @@ metaslab_sync(metaslab_t *msp, uint64_t 
 	 * metaslab_sync() is the metaslab's ms_tree.  No other thread can
 	 * be modifying this txg's alloctree, freetree, freed_tree, or
 	 * space_map_phys_t. Therefore, we only hold ms_lock to satify
-	 * space_map ASSERTs. We drop it whenever we call into the DMU,
+	 * space map ASSERTs. 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.
 	 */
@@ -1949,7 +2228,7 @@ metaslab_sync(metaslab_t *msp, uint64_t 
 	mutex_enter(&msp->ms_lock);
 
 	/*
-	 * Note: metaslab_condense() clears the space_map's histogram.
+	 * Note: metaslab_condense() clears the space map's histogram.
 	 * Therefore we must verify and remove this histogram before
 	 * condensing.
 	 */
@@ -1974,16 +2253,38 @@ metaslab_sync(metaslab_t *msp, uint64_t 
 		 */
 		space_map_histogram_clear(msp->ms_sm);
 		space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx);
-	} else {
+
+		/*
+		 * Since we've cleared the histogram we need to add back
+		 * any free space that has already been processed, plus
+		 * any deferred space. This allows the on-disk histogram
+		 * to accurately reflect all free space even if some space
+		 * is not yet available for allocation (i.e. deferred).
+		 */
+		space_map_histogram_add(msp->ms_sm, *freed_tree, tx);
+
 		/*
-		 * Since the space map is not loaded we simply update the
-		 * exisiting histogram with what was freed in this txg. This
-		 * means that the on-disk histogram may not have an accurate
-		 * view of the free space but it's close enough to allow
-		 * us to make allocation decisions.
+		 * Add back any deferred free space that has not been
+		 * added back into the in-core free tree yet. This will
+		 * ensure that we don't end up with a space map histogram
+		 * that is completely empty unless the metaslab is fully
+		 * allocated.
 		 */
-		space_map_histogram_add(msp->ms_sm, *freetree, tx);
+		for (int t = 0; t < TXG_DEFER_SIZE; t++) {
+			space_map_histogram_add(msp->ms_sm,
+			    msp->ms_defertree[t], tx);
+		}
 	}
+
+	/*
+	 * Always add the free space from this sync pass to the space
+	 * map histogram. We want to make sure that the on-disk histogram
+	 * accounts for all free space. If the space map is not loaded,
+	 * then we will lose some accuracy but will correct it the next
+	 * time we load the space map.
+	 */
+	space_map_histogram_add(msp->ms_sm, *freetree, tx);
+
 	metaslab_group_histogram_add(mg, msp);
 	metaslab_group_histogram_verify(mg);
 	metaslab_class_histogram_verify(mg->mg_class);
@@ -2002,6 +2303,7 @@ metaslab_sync(metaslab_t *msp, uint64_t 
 	range_tree_vacate(alloctree, NULL, NULL);
 
 	ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
+	ASSERT0(range_tree_space(msp->ms_alloctree[TXG_CLEAN(txg) & TXG_MASK]));
 	ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
 
 	mutex_exit(&msp->ms_lock);
@@ -2023,9 +2325,11 @@ metaslab_sync_done(metaslab_t *msp, uint
 {
 	metaslab_group_t *mg = msp->ms_group;
 	vdev_t *vd = mg->mg_vd;
+	spa_t *spa = vd->vdev_spa;
 	range_tree_t **freed_tree;
 	range_tree_t **defer_tree;
 	int64_t alloc_delta, defer_delta;
+	boolean_t defer_allowed = B_TRUE;
 
 	ASSERT(!vd->vdev_ishole);
 
@@ -2060,9 +2364,20 @@ metaslab_sync_done(metaslab_t *msp, uint
 	freed_tree = &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
 	defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE];
 
+	uint64_t free_space = metaslab_class_get_space(spa_normal_class(spa)) -
+	    metaslab_class_get_alloc(spa_normal_class(spa));
+	if (free_space <= spa_get_slop_space(spa)) {
+		defer_allowed = B_FALSE;
+	}
+
+	defer_delta = 0;
 	alloc_delta = space_map_alloc_delta(msp->ms_sm);
-	defer_delta = range_tree_space(*freed_tree) -
-	    range_tree_space(*defer_tree);
+	if (defer_allowed) {
+		defer_delta = range_tree_space(*freed_tree) -
+		    range_tree_space(*defer_tree);
+	} else {
+		defer_delta -= range_tree_space(*defer_tree);
+	}
 
 	vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
 
@@ -2083,7 +2398,12 @@ metaslab_sync_done(metaslab_t *msp, uint
 	 */
 	range_tree_vacate(*defer_tree,
 	    msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
-	range_tree_swap(freed_tree, defer_tree);
+	if (defer_allowed) {
+		range_tree_swap(freed_tree, defer_tree);
+	} else {
+		range_tree_vacate(*freed_tree,
+		    msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
+	}
 
 	space_map_update(msp->ms_sm);
 
@@ -2098,7 +2418,18 @@ metaslab_sync_done(metaslab_t *msp, uint
 		vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
 	}
 
-	if (msp->ms_loaded && msp->ms_access_txg < txg) {
+	/*
+	 * Calculate the new weights before unloading any metaslabs.
+	 * This will give us the most accurate weighting.
+	 */
+	metaslab_group_sort(mg, msp, metaslab_weight(msp));
+
+	/*
+	 * If the metaslab is loaded and we've not tried to load or allocate
+	 * from it in 'metaslab_unload_delay' txgs, then unload it.
+	 */
+	if (msp->ms_loaded &&
+	    msp->ms_selected_txg + metaslab_unload_delay < txg) {
 		for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
 			VERIFY0(range_tree_space(
 			    msp->ms_alloctree[(txg + t) & TXG_MASK]));
@@ -2108,7 +2439,6 @@ metaslab_sync_done(metaslab_t *msp, uint
 			metaslab_unload(msp);
 	}
 
-	metaslab_group_sort(mg, msp, metaslab_weight(msp));
 	mutex_exit(&msp->ms_lock);
 }
 
@@ -2143,6 +2473,113 @@ metaslab_distance(metaslab_t *msp, dva_t
 
 /*
  * ==========================================================================
+ * Metaslab allocation tracing facility
+ * ==========================================================================
+ */
+kstat_t *metaslab_trace_ksp;
+kstat_named_t metaslab_trace_over_limit;
+
+void
+metaslab_alloc_trace_init(void)
+{
+	ASSERT(metaslab_alloc_trace_cache == NULL);
+	metaslab_alloc_trace_cache = kmem_cache_create(
+	    "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t),
+	    0, NULL, NULL, NULL, NULL, NULL, 0);
+	metaslab_trace_ksp = kstat_create("zfs", 0, "metaslab_trace_stats",
+	    "misc", KSTAT_TYPE_NAMED, 1, KSTAT_FLAG_VIRTUAL);
+	if (metaslab_trace_ksp != NULL) {
+		metaslab_trace_ksp->ks_data = &metaslab_trace_over_limit;
+		kstat_named_init(&metaslab_trace_over_limit,
+		    "metaslab_trace_over_limit", KSTAT_DATA_UINT64);
+		kstat_install(metaslab_trace_ksp);
+	}
+}
+
+void
+metaslab_alloc_trace_fini(void)
+{
+	if (metaslab_trace_ksp != NULL) {
+		kstat_delete(metaslab_trace_ksp);
+		metaslab_trace_ksp = NULL;
+	}
+	kmem_cache_destroy(metaslab_alloc_trace_cache);
+	metaslab_alloc_trace_cache = NULL;
+}
+
+/*
+ * Add an allocation trace element to the allocation tracing list.
+ */
+static void
+metaslab_trace_add(zio_alloc_list_t *zal, metaslab_group_t *mg,
+    metaslab_t *msp, uint64_t psize, uint32_t dva_id, uint64_t offset)
+{
+	if (!metaslab_trace_enabled)
+		return;
+
+	/*
+	 * When the tracing list reaches its maximum we remove
+	 * the second element in the list before adding a new one.
+	 * By removing the second element we preserve the original
+	 * entry as a clue to what allocations steps have already been
+	 * performed.
+	 */
+	if (zal->zal_size == metaslab_trace_max_entries) {
+		metaslab_alloc_trace_t *mat_next;
+#ifdef DEBUG
+		panic("too many entries in allocation list");
+#endif
+		atomic_inc_64(&metaslab_trace_over_limit.value.ui64);
+		zal->zal_size--;
+		mat_next = list_next(&zal->zal_list, list_head(&zal->zal_list));
+		list_remove(&zal->zal_list, mat_next);
+		kmem_cache_free(metaslab_alloc_trace_cache, mat_next);
+	}
+
+	metaslab_alloc_trace_t *mat =
+	    kmem_cache_alloc(metaslab_alloc_trace_cache, KM_SLEEP);
+	list_link_init(&mat->mat_list_node);
+	mat->mat_mg = mg;
+	mat->mat_msp = msp;
+	mat->mat_size = psize;
+	mat->mat_dva_id = dva_id;
+	mat->mat_offset = offset;
+	mat->mat_weight = 0;
+
+	if (msp != NULL)
+		mat->mat_weight = msp->ms_weight;
+
+	/*
+	 * The list is part of the zio so locking is not required. Only
+	 * a single thread will perform allocations for a given zio.
+	 */
+	list_insert_tail(&zal->zal_list, mat);
+	zal->zal_size++;
+
+	ASSERT3U(zal->zal_size, <=, metaslab_trace_max_entries);
+}
+
+void
+metaslab_trace_init(zio_alloc_list_t *zal)
+{
+	list_create(&zal->zal_list, sizeof (metaslab_alloc_trace_t),
+	    offsetof(metaslab_alloc_trace_t, mat_list_node));
+	zal->zal_size = 0;
+}
+
+void
+metaslab_trace_fini(zio_alloc_list_t *zal)
+{
+	metaslab_alloc_trace_t *mat;
+
+	while ((mat = list_remove_head(&zal->zal_list)) != NULL)
+		kmem_cache_free(metaslab_alloc_trace_cache, mat);
+	list_destroy(&zal->zal_list);
+	zal->zal_size = 0;
+}
+
+/*
+ * ==========================================================================
  * Metaslab block operations
  * ==========================================================================
  */
@@ -2191,13 +2628,48 @@ metaslab_group_alloc_verify(spa_t *spa, 
 }
 
 static uint64_t
-metaslab_group_alloc(metaslab_group_t *mg, uint64_t asize,
-    uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
+metaslab_block_alloc(metaslab_t *msp, uint64_t size, uint64_t txg)
+{
+	uint64_t start;
+	range_tree_t *rt = msp->ms_tree;
+	metaslab_class_t *mc = msp->ms_group->mg_class;
+
+	VERIFY(!msp->ms_condensing);
+
+	start = mc->mc_ops->msop_alloc(msp, size);
+	if (start != -1ULL) {

*** DIFF OUTPUT TRUNCATED AT 1000 LINES ***


More information about the svn-src-all mailing list