Improved dnode allocation and dmu_hold_impl()

Refactor dmu_object_alloc_dnsize() and dnode_hold_impl() to simplify the
code, fix errors introduced by commit dbeb879 (PR #6117) interacting
badly with large dnodes, and improve performance.

* When allocating a new dnode in dmu_object_alloc_dnsize(), update the
percpu object ID for the core's metadnode chunk immediately.  This
eliminates most lock contention when taking the hold and creating the
dnode.

* Correct detection of the chunk boundary to work properly with large
dnodes.

* Separate the dmu_hold_impl() code for the FREE case from the code for
the ALLOCATED case to make it easier to read.

* Fully populate the dnode handle array immediately after reading a
block of the metadnode from disk.  Subsequently the dnode handle array
provides enough information to determine which dnode slots are in use
and which are free.

* Add several kstats to allow the behavior of the code to be examined.

* Verify dnode packing in large_dnode_008_pos.ksh.  Since the test is
purely creates, it should leave very few holes in the metadnode.

* Add test large_dnode_009_pos.ksh, which performs concurrent creates
and deletes, to complement existing test which does only creates.

With the above fixes, there is very little contention in a test of about
200,000 racing dnode allocations produced by tests 'large_dnode_008_pos'
and 'large_dnode_009_pos'.

name                            type data
dnode_hold_dbuf_hold            4    0
dnode_hold_dbuf_read            4    0
dnode_hold_alloc_hits           4    3804690
dnode_hold_alloc_misses         4    216
dnode_hold_alloc_interior       4    3
dnode_hold_alloc_lock_retry     4    0
dnode_hold_alloc_lock_misses    4    0
dnode_hold_alloc_type_none      4    0
dnode_hold_free_hits            4    203105
dnode_hold_free_misses          4    4
dnode_hold_free_lock_misses     4    0
dnode_hold_free_lock_retry      4    0
dnode_hold_free_overflow        4    0
dnode_hold_free_refcount        4    57
dnode_hold_free_txg             4    0
dnode_allocate                  4    203154
dnode_reallocate                4    0
dnode_buf_evict                 4    23918
dnode_alloc_next_chunk          4    4887
dnode_alloc_race                4    0
dnode_alloc_next_block          4    18

The performance is slightly improved for concurrent creates with
16+ threads, and unchanged for low thread counts.

Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Olaf Faaland <faaland1@llnl.gov>
Closes #5396 
Closes #6522 
Closes #6414 
Closes #6564

Author: ofaaland

cmd/zdb/zdb.c

@@ -1934,7 +1934,8 @@ static object_viewer_t *object_viewer[DMU_OT_NUMTYPES + 1] = {
 };
 
 static void
-dump_object(objset_t *os, uint64_t object, int verbosity, int *print_header)
+dump_object(objset_t *os, uint64_t object, int verbosity, int *print_header,
+    uint64_t *dnode_slots_used)
 {
 	dmu_buf_t *db = NULL;
 	dmu_object_info_t doi;
@@ -1984,6 +1985,9 @@ dump_object(objset_t *os, uint64_t object, int verbosity, int *print_header)
 		}
 	}
 
+	if (dnode_slots_used)
+		*dnode_slots_used = doi.doi_dnodesize / DNODE_MIN_SIZE;
+
 	zdb_nicenum(doi.doi_metadata_block_size, iblk);
 	zdb_nicenum(doi.doi_data_block_size, dblk);
 	zdb_nicenum(doi.doi_max_offset, lsize);
@@ -2104,6 +2108,9 @@ dump_dir(objset_t *os)
 	int verbosity = dump_opt['d'];
 	int print_header = 1;
 	int i, error;
+	uint64_t total_slots_used = 0;
+	uint64_t max_slot_used = 0;
+	uint64_t dnode_slots;
 
 	dsl_pool_config_enter(dmu_objset_pool(os), FTAG);
 	dmu_objset_fast_stat(os, &dds);
@@ -2144,7 +2151,7 @@ dump_dir(objset_t *os)
 	if (zopt_objects != 0) {
 		for (i = 0; i < zopt_objects; i++)
 			dump_object(os, zopt_object[i], verbosity,
-			    &print_header);
+			    &print_header, NULL);
 		(void) printf("\n");
 		return;
 	}
@@ -2161,24 +2168,39 @@ dump_dir(objset_t *os)
 	if (BP_IS_HOLE(os->os_rootbp))
 		return;
 
-	dump_object(os, 0, verbosity, &print_header);
+	dump_object(os, 0, verbosity, &print_header, NULL);
 	object_count = 0;
 	if (DMU_USERUSED_DNODE(os) != NULL &&
 	    DMU_USERUSED_DNODE(os)->dn_type != 0) {
-		dump_object(os, DMU_USERUSED_OBJECT, verbosity, &print_header);
-		dump_object(os, DMU_GROUPUSED_OBJECT, verbosity, &print_header);
+		dump_object(os, DMU_USERUSED_OBJECT, verbosity, &print_header,
+		    NULL);
+		dump_object(os, DMU_GROUPUSED_OBJECT, verbosity, &print_header,
+		    NULL);
 	}
 
 	object = 0;
 	while ((error = dmu_object_next(os, &object, B_FALSE, 0)) == 0) {
-		dump_object(os, object, verbosity, &print_header);
+		dump_object(os, object, verbosity, &print_header, &dnode_slots);
 		object_count++;
+		total_slots_used += dnode_slots;
+		max_slot_used = object + dnode_slots - 1;
 	}
 
 	ASSERT3U(object_count, ==, usedobjs);
 
 	(void) printf("\n");
 
+	(void) printf("    Dnode slots:\n");
+	(void) printf("\tTotal used:    %10llu\n",
+	    (u_longlong_t)total_slots_used);
+	(void) printf("\tMax used:      %10llu\n",
+	    (u_longlong_t)max_slot_used);
+	(void) printf("\tPercent empty: %10lf\n",
+	    (double)(max_slot_used - total_slots_used)*100 /
+	    (double)max_slot_used);
+
+	(void) printf("\n");
+
 	if (error != ESRCH) {
 		(void) fprintf(stderr, "dmu_object_next() = %d\n", error);
 		abort();
@@ -2642,7 +2664,7 @@ dump_path_impl(objset_t *os, uint64_t obj, char *name)
 			return (dump_path_impl(os, child_obj, s + 1));
 		/*FALLTHROUGH*/
 	case DMU_OT_PLAIN_FILE_CONTENTS:
-		dump_object(os, child_obj, dump_opt['v'], &header);
+		dump_object(os, child_obj, dump_opt['v'], &header, NULL);
 		return (0);
 	default:
 		(void) fprintf(stderr, "object %llu has non-file/directory "

include/sys/dnode.h

@@ -98,6 +98,13 @@ extern "C" {
 #define	DN_ZERO_BONUSLEN	(DN_BONUS_SIZE(DNODE_MAX_SIZE) + 1)
 #define	DN_KILL_SPILLBLK (1)
 
+#define	DN_SLOT_UNINIT		((void *)NULL)	/* Uninitialized */
+#define	DN_SLOT_FREE		((void *)1UL)	/* Free slot */
+#define	DN_SLOT_ALLOCATED	((void *)2UL)	/* Allocated slot */
+#define	DN_SLOT_INTERIOR	((void *)3UL)	/* Interior allocated slot */
+#define	DN_SLOT_IS_PTR(dn)	((void *)dn > DN_SLOT_INTERIOR)
+#define	DN_SLOT_IS_VALID(dn)	((void *)dn != NULL)
+
 #define	DNODES_PER_BLOCK_SHIFT	(DNODE_BLOCK_SHIFT - DNODE_SHIFT)
 #define	DNODES_PER_BLOCK	(1ULL << DNODES_PER_BLOCK_SHIFT)
 
@@ -419,6 +426,135 @@ void dnode_evict_bonus(dnode_t *dn);
 	((_dn)->dn_objset->os_primary_cache == ZFS_CACHE_ALL ||		\
 	(_dn)->dn_objset->os_primary_cache == ZFS_CACHE_METADATA)
 
+/*
+ * Used for dnodestats kstat.
+ */
+typedef struct dnode_stats {
+	/*
+	 * Number of failed attempts to hold a meta dnode dbuf.
+	 */
+	kstat_named_t dnode_hold_dbuf_hold;
+	/*
+	 * Number of failed attempts to read a meta dnode dbuf.
+	 */
+	kstat_named_t dnode_hold_dbuf_read;
+	/*
+	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) was able
+	 * to hold the requested object number which was allocated.  This is
+	 * the common case when looking up any allocated object number.
+	 */
+	kstat_named_t dnode_hold_alloc_hits;
+	/*
+	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) was not
+	 * able to hold the request object number because it was not allocated.
+	 */
+	kstat_named_t dnode_hold_alloc_misses;
+	/*
+	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) was not
+	 * able to hold the request object number because the object number
+	 * refers to an interior large dnode slot.
+	 */
+	kstat_named_t dnode_hold_alloc_interior;
+	/*
+	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) needed
+	 * to retry acquiring slot zrl locks due to contention.
+	 */
+	kstat_named_t dnode_hold_alloc_lock_retry;
+	/*
+	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) did not
+	 * need to create the dnode because another thread did so after
+	 * dropping the read lock but before acquiring the write lock.
+	 */
+	kstat_named_t dnode_hold_alloc_lock_misses;
+	/*
+	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) found
+	 * a free dnode instantiated by dnode_create() but not yet allocated
+	 * by dnode_allocate().
+	 */
+	kstat_named_t dnode_hold_alloc_type_none;
+	/*
+	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) was able
+	 * to hold the requested range of free dnode slots.
+	 */
+	kstat_named_t dnode_hold_free_hits;
+	/*
+	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) was not
+	 * able to hold the requested range of free dnode slots because
+	 * at least one slot was allocated.
+	 */
+	kstat_named_t dnode_hold_free_misses;
+	/*
+	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) was not
+	 * able to hold the requested range of free dnode slots because
+	 * after acquiring the zrl lock at least one slot was allocated.
+	 */
+	kstat_named_t dnode_hold_free_lock_misses;
+	/*
+	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) needed
+	 * to retry acquiring slot zrl locks due to contention.
+	 */
+	kstat_named_t dnode_hold_free_lock_retry;
+	/*
+	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) requested
+	 * a range of dnode slots which were held by another thread.
+	 */
+	kstat_named_t dnode_hold_free_refcount;
+	/*
+	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) requested
+	 * a range of dnode slots which would overflow the dnode_phys_t.
+	 */
+	kstat_named_t dnode_hold_free_overflow;
+	/*
+	 * Number of times a dnode_hold(...) was attempted on a dnode
+	 * which had already been unlinked in an earlier txg.
+	 */
+	kstat_named_t dnode_hold_free_txg;
+	/*
+	 * Number of new dnodes allocated by dnode_allocate().
+	 */
+	kstat_named_t dnode_allocate;
+	/*
+	 * Number of dnodes re-allocated by dnode_reallocate().
+	 */
+	kstat_named_t dnode_reallocate;
+	/*
+	 * Number of meta dnode dbufs evicted.
+	 */
+	kstat_named_t dnode_buf_evict;
+	/*
+	 * Number of times dmu_object_alloc*() reached the end of the existing
+	 * object ID chunk and advanced to a new one.
+	 */
+	kstat_named_t dnode_alloc_next_chunk;
+	/*
+	 * Number of times multiple threads attempted to allocate a dnode
+	 * from the same block of free dnodes.
+	 */
+	kstat_named_t dnode_alloc_race;
+	/*
+	 * Number of times dmu_object_alloc*() was forced to advance to the
+	 * next meta dnode dbuf due to an error from  dmu_object_next().
+	 */
+	kstat_named_t dnode_alloc_next_block;
+	/*
+	 * Statistics for tracking dnodes which have been moved.
+	 */
+	kstat_named_t dnode_move_invalid;
+	kstat_named_t dnode_move_recheck1;
+	kstat_named_t dnode_move_recheck2;
+	kstat_named_t dnode_move_special;
+	kstat_named_t dnode_move_handle;
+	kstat_named_t dnode_move_rwlock;
+	kstat_named_t dnode_move_active;
+} dnode_stats_t;
+
+extern dnode_stats_t dnode_stats;
+
+#define	DNODE_STAT_INCR(stat, val) \
+    atomic_add_64(&dnode_stats.stat.value.ui64, (val));
+#define	DNODE_STAT_BUMP(stat) \
+    DNODE_STAT_INCR(stat, 1);
+
 #ifdef ZFS_DEBUG
 
 /*

module/zfs/dbuf_stats.c

@@ -72,8 +72,7 @@ __dbuf_stats_hash_table_data(char *buf, size_t size, dmu_buf_impl_t *db)
 	if (db->db_buf)
 		arc_buf_info(db->db_buf, &abi, zfs_dbuf_state_index);
 
-	if (dn)
-		__dmu_object_info_from_dnode(dn, &doi);
+	__dmu_object_info_from_dnode(dn, &doi);
 
 	nwritten = snprintf(buf, size,
 	    "%-16s %-8llu %-8lld %-8lld %-8lld %-8llu %-8llu %-5d %-5d %-5lu | "

module/zfs/dmu_object.c

@@ -93,7 +93,10 @@ dmu_object_alloc_dnsize(objset_t *os, dmu_object_type_t ot, int blocksize,
 		 * If we finished a chunk of dnodes, get a new one from
 		 * the global allocator.
 		 */
-		if (P2PHASE(object, dnodes_per_chunk) == 0) {
+		if ((P2PHASE(object, dnodes_per_chunk) == 0) ||
+		    (P2PHASE(object + dn_slots - 1, dnodes_per_chunk) <
+		    dn_slots)) {
+			DNODE_STAT_BUMP(dnode_alloc_next_chunk);
 			mutex_enter(&os->os_obj_lock);
 			ASSERT0(P2PHASE(os->os_obj_next_chunk,
 			    dnodes_per_chunk));
@@ -157,6 +160,13 @@ dmu_object_alloc_dnsize(objset_t *os, dmu_object_type_t ot, int blocksize,
 			mutex_exit(&os->os_obj_lock);
 		}
 
+		/*
+		 * The value of (*cpuobj) before adding dn_slots is the object
+		 * ID assigned to us.  The value afterwards is the object ID
+		 * assigned to whoever wants to do an allocation next.
+		 */
+		object = atomic_add_64_nv(cpuobj, dn_slots) - dn_slots;
+
 		/*
 		 * XXX We should check for an i/o error here and return
 		 * up to our caller.  Actually we should pre-read it in
@@ -177,21 +187,20 @@ dmu_object_alloc_dnsize(objset_t *os, dmu_object_type_t ot, int blocksize,
 				rw_exit(&dn->dn_struct_rwlock);
 				dmu_tx_add_new_object(tx, dn);
 				dnode_rele(dn, FTAG);
-
-				(void) atomic_swap_64(cpuobj,
-				    object + dn_slots);
 				return (object);
 			}
 			rw_exit(&dn->dn_struct_rwlock);
 			dnode_rele(dn, FTAG);
+			DNODE_STAT_BUMP(dnode_alloc_race);
 		}
 
+		/*
+		 * Skip to next known valid starting point on error.  This
+		 * is the start of the next block of dnodes.
+		 */
 		if (dmu_object_next(os, &object, B_TRUE, 0) != 0) {
-			/*
-			 * Skip to next known valid starting point for a
-			 * dnode.
-			 */
 			object = P2ROUNDUP(object + 1, DNODES_PER_BLOCK);
+			DNODE_STAT_BUMP(dnode_alloc_next_block);
 		}
 		(void) atomic_swap_64(cpuobj, object);
 	}
@@ -304,24 +313,37 @@ dmu_object_next(objset_t *os, uint64_t *objectp, boolean_t hole, uint64_t txg)
 	if (*objectp == 0) {
 		start_obj = 1;
 	} else if (ds && ds->ds_feature_inuse[SPA_FEATURE_LARGE_DNODE]) {
+		uint64_t i = *objectp + 1;
+		uint64_t last_obj = *objectp | (DNODES_PER_BLOCK - 1);
+		dmu_object_info_t doi;
+
 		/*
-		 * For large_dnode datasets, scan from the beginning of the
-		 * dnode block to find the starting offset. This is needed
-		 * because objectp could be part of a large dnode so we can't
-		 * assume it's a hole even if dmu_object_info() returns ENOENT.
+		 * Scan through the remaining meta dnode block.  The contents
+		 * of each slot in the block are known so it can be quickly
+		 * checked.  If the block is exhausted without a match then
+		 * hand off to dnode_next_offset() for further scanning.
 		 */
-		int epb = DNODE_BLOCK_SIZE >> DNODE_SHIFT;
-		int skip;
-		uint64_t i;
-
-		for (i = *objectp & ~(epb - 1); i <= *objectp; i += skip) {
-			dmu_object_info_t doi;
-
+		while (i <= last_obj) {
 			error = dmu_object_info(os, i, &doi);
-			if (error != 0)
-				skip = 1;
-			else
-				skip = doi.doi_dnodesize >> DNODE_SHIFT;
+			if (error == ENOENT) {
+				if (hole) {
+					*objectp = i;
+					return (0);
+				} else {
+					i++;
+				}
+			} else if (error == EEXIST) {
+				i++;
+			} else if (error == 0) {
+				if (hole) {
+					i += doi.doi_dnodesize >> DNODE_SHIFT;
+				} else {
+					*objectp = i;
+					return (0);
+				}
+			} else {
+				return (error);
+			}
 		}
 
 		start_obj = i;