Compact dbuf/buf hashes and lock arrays

With default dbuf cache size of 1/32 of ARC, it makes no sense to have
hash table of the same size (or even bigger on Linux).  Reduce it to
1/8 of ARC's one, still leaving some slack, assuming higher I/O rate
via dbuf cache than via ARC.

Remove padding from ARC hash locks array.  The idea behind padding
is to avoid false sharing between locks.  It would have sense if
there would be a limited number of very busy locks.  But since we
have no limit on the number, using the same memory for more locks we
can achieve even lower lock contention with the same false sharing,
or we can use less memory for the same contention level.

Reduce number of hash locks from 8192 to 2048.  The number is still
big enough to not cause contention, but reduced memory size improves
cache hit rate for mutex_tryenter() in ARC eviction thread, saving
about 1% of the thread time.

Reviewed-by: Matthew Ahrens <[email protected]>
Reviewed-by: Brian Behlendorf <[email protected]>
Signed-off-by: Alexander Motin <[email protected]>
Sponsored-By: iXsystems, Inc.
Closes #12289

3 files changed, 11 insertions, 24 deletions

diff --git a/include/sys/dbuf.h b/include/sys/dbuf.h
index d221eac4c..6ae079c6a 100644
--- a/include/sys/dbuf.h
+++ b/include/sys/dbuf.h
@@ -322,12 +322,12 @@ typedef struct dmu_buf_impl {
 } dmu_buf_impl_t;
 
 /* Note: the dbuf hash table is exposed only for the mdb module */
-#define	DBUF_MUTEXES 8192
+#define	DBUF_MUTEXES 2048
 #define	DBUF_HASH_MUTEX(h, idx) (&(h)->hash_mutexes[(idx) & (DBUF_MUTEXES-1)])
 typedef struct dbuf_hash_table {
 	uint64_t hash_table_mask;
 	dmu_buf_impl_t **hash_table;
-	kmutex_t hash_mutexes[DBUF_MUTEXES];
+	kmutex_t hash_mutexes[DBUF_MUTEXES] ____cacheline_aligned;
 } dbuf_hash_table_t;
 
 typedef void (*dbuf_prefetch_fn)(void *, boolean_t);
diff --git a/module/zfs/arc.c b/module/zfs/arc.c
index 3484fff3b..394ca1bfe 100644
--- a/module/zfs/arc.c
+++ b/module/zfs/arc.c
@@ -740,29 +740,18 @@ taskq_t *arc_prune_taskq;
  * Hash table routines
  */
 
-#define	HT_LOCK_ALIGN	64
-#define	HT_LOCK_PAD	(P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
-
-struct ht_lock {
-	kmutex_t	ht_lock;
-#ifdef _KERNEL
-	unsigned char	pad[HT_LOCK_PAD];
-#endif
-};
-
-#define	BUF_LOCKS 8192
+#define	BUF_LOCKS 2048
 typedef struct buf_hash_table {
 	uint64_t ht_mask;
 	arc_buf_hdr_t **ht_table;
-	struct ht_lock ht_locks[BUF_LOCKS];
+	kmutex_t ht_locks[BUF_LOCKS] ____cacheline_aligned;
 } buf_hash_table_t;
 
 static buf_hash_table_t buf_hash_table;
 
 #define	BUF_HASH_INDEX(spa, dva, birth) \
 	(buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
-#define	BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
-#define	BUF_HASH_LOCK(idx)	(&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
+#define	BUF_HASH_LOCK(idx)	(&buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
 #define	HDR_LOCK(hdr) \
 	(BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
 
@@ -1111,7 +1100,7 @@ buf_fini(void)
 	    (buf_hash_table.ht_mask + 1) * sizeof (void *));
 #endif
 	for (i = 0; i < BUF_LOCKS; i++)
-		mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
+		mutex_destroy(BUF_HASH_LOCK(i));
 	kmem_cache_destroy(hdr_full_cache);
 	kmem_cache_destroy(hdr_full_crypt_cache);
 	kmem_cache_destroy(hdr_l2only_cache);
@@ -1276,10 +1265,8 @@ retry:
 		for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
 			*ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
 
-	for (i = 0; i < BUF_LOCKS; i++) {
-		mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
-		    NULL, MUTEX_DEFAULT, NULL);
-	}
+	for (i = 0; i < BUF_LOCKS; i++)
+		mutex_init(BUF_HASH_LOCK(i), NULL, MUTEX_DEFAULT, NULL);
 }
 
 #define	ARC_MINTIME	(hz>>4) /* 62 ms */
diff --git a/module/zfs/dbuf.c b/module/zfs/dbuf.c
index 9ce091b80..289247c6e 100644
--- a/module/zfs/dbuf.c
+++ b/module/zfs/dbuf.c
@@ -826,12 +826,12 @@ dbuf_init(void)
 	int i;
 
 	/*
-	 * The hash table is big enough to fill all of physical memory
+	 * The hash table is big enough to fill one eighth of physical memory
 	 * with an average block size of zfs_arc_average_blocksize (default 8K).
 	 * By default, the table will take up
 	 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
 	 */
-	while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
+	while (hsize * zfs_arc_average_blocksize < arc_all_memory() / 8)
 		hsize <<= 1;
 
 retry:
@@ -3055,8 +3055,8 @@ dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
 	db->db_state = DB_EVICTING; /* not worth logging this state change */
 	if ((odb = dbuf_hash_insert(db)) != NULL) {
 		/* someone else inserted it first */
-		kmem_cache_free(dbuf_kmem_cache, db);
 		mutex_exit(&dn->dn_dbufs_mtx);
+		kmem_cache_free(dbuf_kmem_cache, db);
 		DBUF_STAT_BUMP(hash_insert_race);
 		return (odb);
 	}