Limit the amount of dnode metadata in the ARC

Metadata-intensive workloads can cause the ARC to become permanently
filled with dnode_t objects as they're pinned by the VFS layer.
Subsequent data-intensive workloads may only benefit from about
25% of the potential ARC (arc_c_max - arc_meta_limit).

In order to help track metadata usage more precisely, the other_size
metadata arcstat has replaced with dbuf_size, dnode_size and bonus_size.

The new zfs_arc_dnode_limit tunable, which defaults to 10% of
zfs_arc_meta_limit, defines the minimum number of bytes which is desirable
to be consumed by dnodes.  Attempts to evict non-metadata will trigger
async prune tasks if the space used by dnodes exceeds this limit.

The new zfs_arc_dnode_reduce_percent tunable specifies the amount by
which the excess dnode space is attempted to be pruned as a percentage of
the amount by which zfs_arc_dnode_limit is being exceeded.  By default,
it tries to unpin 10% of the dnodes.

The problem of dnode metadata pinning was observed with the following
testing procedure (in this example, zfs_arc_max is set to 4GiB):

    - Create a large number of small files until arc_meta_used exceeds
      arc_meta_limit (3GiB with default tuning) and arc_prune
      starts increasing.

    - Create a 3GiB file with dd.  Observe arc_mata_used.  It will still
      be around 3GiB.

    - Repeatedly read the 3GiB file and observe arc_meta_limit as before.
      It will continue to stay around 3GiB.

With this modification, space for the 3GiB file is gradually made
available as subsequent demands on the ARC are made.

include/sys/arc.h

@@ -130,7 +130,9 @@ typedef enum arc_space_type {
 	ARC_SPACE_META,
 	ARC_SPACE_HDRS,
 	ARC_SPACE_L2HDRS,
-	ARC_SPACE_OTHER,
+	ARC_SPACE_DBUF,
+	ARC_SPACE_DNODE,
+	ARC_SPACE_BONUS,
 	ARC_SPACE_NUMTYPES
 } arc_space_type_t;
 

man/man5/zfs-module-parameters.5

@@ -361,6 +361,37 @@ Min time before an active prefetch stream can be reclaimed
 Default value: \fB2\fR.
 .RE
 
+.sp
+.ne 2
+.na
+\fBzfs_arc_dnode_limit\fR (ulong)
+.ad
+.RS 12n
+When the number of bytes consumed by dnodes in the ARC exceeds this number of
+bytes, try to unpin some of it in response to demand for non-metadata.  This
+value acts as a floor to the amount of dnode metadata.
+
+See also \fBzfs_arc_meta_prune\fR which serves a similar purpose but is used
+when the amount of metadata in the ARC exceeds \fBzfs_arc_meta_limit\fR rather
+than in response to overall demand for non-metadata.
+
+.sp
+Default value: \fB10% of zfs_arc_meta_limit\fR.
+.RE
+
+.sp
+.ne 2
+.na
+\fBzfs_arc_dnode_reduce_percent\fR (ulong)
+.ad
+.RS 12n
+Percentage of ARC dnodes to try to scan in response to demand for non-metadata
+when the number of bytes consumed by dnodes exceeds \fBzfs_arc_dnode_limit\fB.
+
+.sp
+Default value: \fB10% of the number of dnodes in the ARC\fR.
+.RE
+
 .sp
 .ne 2
 .na

module/zfs/arc.c

@@ -231,6 +231,8 @@ unsigned long zfs_arc_max = 0;
 unsigned long zfs_arc_min = 0;
 unsigned long zfs_arc_meta_limit = 0;
 unsigned long zfs_arc_meta_min = 0;
+unsigned long zfs_arc_dnode_limit = 0;
+unsigned long zfs_arc_dnode_reduce_percent = 10;
 int zfs_arc_grow_retry = 0;
 int zfs_arc_shrink_shift = 0;
 int zfs_arc_p_min_shift = 0;
@@ -328,13 +330,17 @@ typedef struct arc_stats {
 	 */
 	kstat_named_t arcstat_metadata_size;
 	/*
-	 * Number of bytes consumed by various buffers and structures
-	 * not actually backed with ARC buffers. This includes bonus
-	 * buffers (allocated directly via zio_buf_* functions),
-	 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
-	 * cache), and dnode_t structures (allocated via dnode_t cache).
+	 * Number of bytes consumed by dmu_buf_impl_t objects.
 	 */
-	kstat_named_t arcstat_other_size;
+	kstat_named_t arcstat_dbuf_size;
+	/*
+	 * Number of bytes consumed by dnode_t objects.
+	 */
+	kstat_named_t arcstat_dnode_size;
+	/*
+	 * Number of bytes consumed by bonus buffers.
+	 */
+	kstat_named_t arcstat_bonus_size;
 	/*
 	 * Total number of bytes consumed by ARC buffers residing in the
 	 * arc_anon state. This includes *all* buffers in the arc_anon
@@ -472,6 +478,7 @@ typedef struct arc_stats {
 	kstat_named_t arcstat_prune;
 	kstat_named_t arcstat_meta_used;
 	kstat_named_t arcstat_meta_limit;
+	kstat_named_t arcstat_dnode_limit;
 	kstat_named_t arcstat_meta_max;
 	kstat_named_t arcstat_meta_min;
 	kstat_named_t arcstat_need_free;
@@ -514,7 +521,9 @@ static arc_stats_t arc_stats = {
 	{ "hdr_size",			KSTAT_DATA_UINT64 },
 	{ "data_size",			KSTAT_DATA_UINT64 },
 	{ "metadata_size",		KSTAT_DATA_UINT64 },
-	{ "other_size",			KSTAT_DATA_UINT64 },
+	{ "dbuf_size",			KSTAT_DATA_UINT64 },
+	{ "dnode_size",			KSTAT_DATA_UINT64 },
+	{ "bonus_size",			KSTAT_DATA_UINT64 },
 	{ "anon_size",			KSTAT_DATA_UINT64 },
 	{ "anon_evictable_data",	KSTAT_DATA_UINT64 },
 	{ "anon_evictable_metadata",	KSTAT_DATA_UINT64 },
@@ -566,6 +575,7 @@ static arc_stats_t arc_stats = {
 	{ "arc_prune",			KSTAT_DATA_UINT64 },
 	{ "arc_meta_used",		KSTAT_DATA_UINT64 },
 	{ "arc_meta_limit",		KSTAT_DATA_UINT64 },
+	{ "arc_dnode_limit",		KSTAT_DATA_UINT64 },
 	{ "arc_meta_max",		KSTAT_DATA_UINT64 },
 	{ "arc_meta_min",		KSTAT_DATA_UINT64 },
 	{ "arc_need_free",		KSTAT_DATA_UINT64 },
@@ -635,9 +645,13 @@ static arc_state_t	*arc_l2c_only;
 #define	arc_tempreserve	ARCSTAT(arcstat_tempreserve)
 #define	arc_loaned_bytes	ARCSTAT(arcstat_loaned_bytes)
 #define	arc_meta_limit	ARCSTAT(arcstat_meta_limit) /* max size for metadata */
+#define	arc_dnode_limit	ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */
 #define	arc_meta_min	ARCSTAT(arcstat_meta_min) /* min size for metadata */
 #define	arc_meta_used	ARCSTAT(arcstat_meta_used) /* size of metadata */
 #define	arc_meta_max	ARCSTAT(arcstat_meta_max) /* max size of metadata */
+#define	arc_dbuf_size	ARCSTAT(arcstat_dbuf_size) /* dbuf metadata */
+#define	arc_dnode_size	ARCSTAT(arcstat_dnode_size) /* dnode metadata */
+#define	arc_bonus_size	ARCSTAT(arcstat_bonus_size) /* bonus buffer metadata */
 #define	arc_need_free	ARCSTAT(arcstat_need_free) /* bytes to be freed */
 #define	arc_sys_free	ARCSTAT(arcstat_sys_free) /* target system free bytes */
 
@@ -793,6 +807,7 @@ static void arc_access(arc_buf_hdr_t *, kmutex_t *);
 static boolean_t arc_is_overflowing(void);
 static void arc_buf_watch(arc_buf_t *);
 static void arc_tuning_update(void);
+static void arc_prune_async(int64_t);
 
 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
@@ -1666,8 +1681,14 @@ arc_space_consume(uint64_t space, arc_space_type_t type)
 	case ARC_SPACE_META:
 		ARCSTAT_INCR(arcstat_metadata_size, space);
 		break;
-	case ARC_SPACE_OTHER:
-		ARCSTAT_INCR(arcstat_other_size, space);
+	case ARC_SPACE_BONUS:
+		ARCSTAT_INCR(arcstat_bonus_size, space);
+		break;
+	case ARC_SPACE_DNODE:
+		ARCSTAT_INCR(arcstat_dnode_size, space);
+		break;
+	case ARC_SPACE_DBUF:
+		ARCSTAT_INCR(arcstat_dbuf_size, space);
 		break;
 	case ARC_SPACE_HDRS:
 		ARCSTAT_INCR(arcstat_hdr_size, space);
@@ -1697,8 +1718,14 @@ arc_space_return(uint64_t space, arc_space_type_t type)
 	case ARC_SPACE_META:
 		ARCSTAT_INCR(arcstat_metadata_size, -space);
 		break;
-	case ARC_SPACE_OTHER:
-		ARCSTAT_INCR(arcstat_other_size, -space);
+	case ARC_SPACE_BONUS:
+		ARCSTAT_INCR(arcstat_bonus_size, -space);
+		break;
+	case ARC_SPACE_DNODE:
+		ARCSTAT_INCR(arcstat_dnode_size, -space);
+		break;
+	case ARC_SPACE_DBUF:
+		ARCSTAT_INCR(arcstat_dbuf_size, -space);
 		break;
 	case ARC_SPACE_HDRS:
 		ARCSTAT_INCR(arcstat_hdr_size, -space);
@@ -2585,16 +2612,25 @@ arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
 	 * we're evicting all available buffers.
 	 */
 	while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
+		int sublist_idx = multilist_get_random_index(ml);
+		uint64_t scan_evicted = 0;
+
+		/*
+		 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
+		 * Request that 10% of the LRUs be scanned by the superblock
+		 * shrinker.
+		 */
+		if (type == ARC_BUFC_DATA && arc_dnode_size > arc_dnode_limit)
+			arc_prune_async((arc_dnode_size - arc_dnode_limit) /
+			    sizeof (dnode_t) / zfs_arc_dnode_reduce_percent);
+
 		/*
 		 * Start eviction using a randomly selected sublist,
 		 * this is to try and evenly balance eviction across all
 		 * sublists. Always starting at the same sublist
 		 * (e.g. index 0) would cause evictions to favor certain
 		 * sublists over others.
 		 */
-		int sublist_idx = multilist_get_random_index(ml);
-		uint64_t scan_evicted = 0;
-
 		for (i = 0; i < num_sublists; i++) {
 			uint64_t bytes_remaining;
 			uint64_t bytes_evicted;
@@ -5250,7 +5286,8 @@ arc_tuning_update(void)
 		arc_c_max = zfs_arc_max;
 		arc_c = arc_c_max;
 		arc_p = (arc_c >> 1);
-		arc_meta_limit = MIN(arc_meta_limit, arc_c_max);
+		arc_meta_limit = MIN(arc_meta_limit, (3 * arc_c_max) / 4);
+		arc_dnode_limit = arc_meta_limit / 10;
 	}
 
 	/* Valid range: 32M - <arc_c_max> */
@@ -5267,6 +5304,7 @@ arc_tuning_update(void)
 	    (zfs_arc_meta_min <= arc_c_max)) {
 		arc_meta_min = zfs_arc_meta_min;
 		arc_meta_limit = MAX(arc_meta_limit, arc_meta_min);
+		arc_dnode_limit = arc_meta_limit / 10;
 	}
 
 	/* Valid range: <arc_meta_min> - <arc_c_max> */
@@ -5275,6 +5313,12 @@ arc_tuning_update(void)
 	    (zfs_arc_meta_limit <= arc_c_max))
 		arc_meta_limit = zfs_arc_meta_limit;
 
+	/* Valid range: <arc_meta_min> - <arc_c_max> */
+	if ((zfs_arc_dnode_limit) && (zfs_arc_dnode_limit != arc_dnode_limit) &&
+	    (zfs_arc_dnode_limit >= zfs_arc_meta_min) &&
+	    (zfs_arc_dnode_limit <= arc_c_max))
+		arc_dnode_limit = zfs_arc_dnode_limit;
+
 	/* Valid range: 1 - N */
 	if (zfs_arc_grow_retry)
 		arc_grow_retry = zfs_arc_grow_retry;
@@ -5363,6 +5407,8 @@ arc_init(void)
 	arc_meta_max = 0;
 	/* Set limit to 3/4 of arc_c_max with a floor of arc_meta_min */
 	arc_meta_limit = MAX((3 * arc_c_max) / 4, arc_meta_min);
+	/* Default dnode limit is 10% of overall meta limit */
+	arc_dnode_limit = arc_meta_limit / 10;
 
 	/* Apply user specified tunings */
 	arc_tuning_update();
@@ -7090,4 +7136,11 @@ MODULE_PARM_DESC(zfs_arc_lotsfree_percent,
 module_param(zfs_arc_sys_free, ulong, 0644);
 MODULE_PARM_DESC(zfs_arc_sys_free, "System free memory target size in bytes");
 
+module_param(zfs_arc_dnode_limit, ulong, 0644);
+MODULE_PARM_DESC(zfs_arc_dnode_limit, "Minimum bytes of dnodes in arc");
+
+module_param(zfs_arc_dnode_reduce_percent, ulong, 0644);
+MODULE_PARM_DESC(zfs_arc_dnode_reduce_percent,
+	"Percentage of excess dnodes to try to unpin");
+
 #endif

module/zfs/dbuf.c

@@ -671,7 +671,7 @@ dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
 
 		ASSERT3U(bonuslen, <=, db->db.db_size);
 		db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
-		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
+		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_BONUS);
 		if (bonuslen < DN_MAX_BONUSLEN)
 			bzero(db->db.db_data, DN_MAX_BONUSLEN);
 		if (bonuslen)
@@ -884,7 +884,7 @@ dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
 	if (db->db_blkid == DMU_BONUS_BLKID) {
 		/* Note that the data bufs here are zio_bufs */
 		dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
-		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
+		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_BONUS);
 		bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
 	} else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
 		int size = db->db.db_size;
@@ -1733,7 +1733,7 @@ dbuf_clear(dmu_buf_impl_t *db)
 		ASSERT(db->db.db_data != NULL);
 		if (db->db_blkid == DMU_BONUS_BLKID) {
 			zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
-			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
+			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_BONUS);
 		}
 		db->db.db_data = NULL;
 		db->db_state = DB_UNCACHED;
@@ -1892,7 +1892,7 @@ dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
 		db->db.db_offset = DMU_BONUS_BLKID;
 		db->db_state = DB_UNCACHED;
 		/* the bonus dbuf is not placed in the hash table */
-		arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
+		arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
 		return (db);
 	} else if (blkid == DMU_SPILL_BLKID) {
 		db->db.db_size = (blkptr != NULL) ?
@@ -1926,7 +1926,7 @@ dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
 		dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
 	db->db_state = DB_UNCACHED;
 	mutex_exit(&dn->dn_dbufs_mtx);
-	arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
+	arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
 
 	if (parent && parent != dn->dn_dbuf)
 		dbuf_add_ref(parent, db);
@@ -2003,7 +2003,7 @@ dbuf_destroy(dmu_buf_impl_t *db)
 	ASSERT(db->db_data_pending == NULL);
 
 	kmem_cache_free(dbuf_cache, db);
-	arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
+	arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
 }
 
 void
@@ -2649,7 +2649,7 @@ dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
 
 		if (*datap != db->db.db_data) {
 			zio_buf_free(*datap, DN_MAX_BONUSLEN);
-			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
+			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_BONUS);
 		}
 		db->db_data_pending = NULL;
 		drp = &db->db_last_dirty;

module/zfs/dnode.c

@@ -470,7 +470,7 @@ dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
 	dnh->dnh_dnode = dn;
 	mutex_exit(&os->os_lock);
 
-	arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
+	arc_space_consume(sizeof (dnode_t), ARC_SPACE_DNODE);
 	return (dn);
 }
 
@@ -526,7 +526,7 @@ dnode_destroy(dnode_t *dn)
 
 	dmu_zfetch_rele(&dn->dn_zfetch);
 	kmem_cache_free(dnode_cache, dn);
-	arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
+	arc_space_return(sizeof (dnode_t), ARC_SPACE_DNODE);
 
 	if (complete_os_eviction)
 		dmu_objset_evict_done(os);