6322 ZFS indirect block predictive prefetch

Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Paul Dagnelie <pcd@delphix.com>
Author: Alexander Motin <mav@FreeBSD.org>

Improve speculative prefetch of indirect blocks.

Scalability of many operations on wide ZFS pool can be limited by
requirement to prefetch indirect blocks first.  Recently added
asynchronous indirect block read partially helped, but did not
solve the problem completely.  This patch extends existing prefetcher
functionality to explicitly work with indirect blocks.

Before this change prefetcher issued reads for up to 8MB of data in
advance.  With this change it also issues indirect block reads
for up to 64MB of data in advance, so that when it will be time to
actually read those data, it can be done immediately.  Alike effect
can be achieved by just increasing maximal data prefetch distance,
but at higher memory cost.

Also this change introduces indirect block prefetch for rewrite
operations, that was never done before.  Previously ARC miss for
Indirect blocks regularly blocked rewrites, converting perfectly
aligned asynchronous operations into synchronous read-write pairs,
significantly reducing maximal rewrite speed.

While being there this issue was also fixed:
 - prefetch was done always, even if caching for the dataset was
completely disabled.

Testing on FreeBSD with zvol on top of 6x striped 2x mirrored pool
of 12 assorted HDDs shown me such performance numbers:
------- BEFORE --------
Write       491363677 bytes/sec
Read        312430631 bytes/sec
Rewrite      97680464 bytes/sec
-------- AFTER --------
Write       493524146 bytes/sec
Read        438598079 bytes/sec
Rewrite     277506044 bytes/sec

Closes #65
Closes #80

openzfs/openzfs@792fd28ac0

uts/common/fs/zfs/dbuf.c

@@ -721,7 +721,7 @@ dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
 	if (db->db_state == DB_CACHED) {
 		mutex_exit(&db->db_mtx);
 		if (prefetch)
-			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1);
+			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
 		if ((flags & DB_RF_HAVESTRUCT) == 0)
 			rw_exit(&dn->dn_struct_rwlock);
 		DB_DNODE_EXIT(db);
@@ -735,7 +735,7 @@ dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
 		/* dbuf_read_impl has dropped db_mtx for us */
 
 		if (prefetch)
-			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1);
+			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
 
 		if ((flags & DB_RF_HAVESTRUCT) == 0)
 			rw_exit(&dn->dn_struct_rwlock);
@@ -754,7 +754,7 @@ dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
 		 */
 		mutex_exit(&db->db_mtx);
 		if (prefetch)
-			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1);
+			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
 		if ((flags & DB_RF_HAVESTRUCT) == 0)
 			rw_exit(&dn->dn_struct_rwlock);
 		DB_DNODE_EXIT(db);

uts/common/fs/zfs/dmu.c

@@ -441,9 +441,10 @@ dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
 		dbp[i] = &db->db;
 	}
 
-	if ((flags & DMU_READ_NO_PREFETCH) == 0 && read &&
-	    length <= zfetch_array_rd_sz) {
-		dmu_zfetch(&dn->dn_zfetch, blkid, nblks);
+	if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
+	    DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
+		dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
+		    read && DNODE_IS_CACHEABLE(dn));
 	}
 	rw_exit(&dn->dn_struct_rwlock);
 

uts/common/fs/zfs/dmu_zfetch.c

@@ -49,6 +49,8 @@ uint32_t	zfetch_max_streams = 8;
 uint32_t	zfetch_min_sec_reap = 2;
 /* max bytes to prefetch per stream (default 8MB) */
 uint32_t	zfetch_max_distance = 8 * 1024 * 1024;
+/* max bytes to prefetch indirects for per stream (default 64MB) */
+uint32_t	zfetch_max_idistance = 64 * 1024 * 1024;
 /* max number of bytes in an array_read in which we allow prefetching (1MB) */
 uint64_t	zfetch_array_rd_sz = 1024 * 1024;
 
@@ -186,6 +188,7 @@ dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid)
 	zstream_t *zs = kmem_zalloc(sizeof (*zs), KM_SLEEP);
 	zs->zs_blkid = blkid;
 	zs->zs_pf_blkid = blkid;
+	zs->zs_ipf_blkid = blkid;
 	zs->zs_atime = gethrtime();
 	mutex_init(&zs->zs_lock, NULL, MUTEX_DEFAULT, NULL);
 
@@ -193,13 +196,21 @@ dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid)
 }
 
 /*
- * This is the prefetch entry point.  It calls all of the other dmu_zfetch
- * routines to create, delete, find, or operate upon prefetch streams.
+ * This is the predictive prefetch entry point.  It associates dnode access
+ * specified with blkid and nblks arguments with prefetch stream, predicts
+ * further accesses based on that stats and initiates speculative prefetch.
+ * fetch_data argument specifies whether actual data blocks should be fetched:
+ *   FALSE -- prefetch only indirect blocks for predicted data blocks;
+ *   TRUE -- prefetch predicted data blocks plus following indirect blocks.
  */
 void
-dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks)
+dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data)
 {
 	zstream_t *zs;
+	int64_t pf_start, ipf_start, ipf_istart, ipf_iend;
+	int64_t pf_ahead_blks, max_blks;
+	int epbs, max_dist_blks, pf_nblks, ipf_nblks;
+	uint64_t end_of_access_blkid = blkid + nblks;
 
 	if (zfs_prefetch_disable)
 		return;
@@ -236,7 +247,7 @@ dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks)
 		 */
 		ZFETCHSTAT_BUMP(zfetchstat_misses);
 		if (rw_tryupgrade(&zf->zf_rwlock))
-			dmu_zfetch_stream_create(zf, blkid + nblks);
+			dmu_zfetch_stream_create(zf, end_of_access_blkid);
 		rw_exit(&zf->zf_rwlock);
 		return;
 	}
@@ -248,35 +259,74 @@ dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks)
 	 * Normally, we start prefetching where we stopped
 	 * prefetching last (zs_pf_blkid).  But when we get our first
 	 * hit on this stream, zs_pf_blkid == zs_blkid, we don't
-	 * want to prefetch to block we just accessed.  In this case,
+	 * want to prefetch the block we just accessed.  In this case,
 	 * start just after the block we just accessed.
 	 */
-	int64_t pf_start = MAX(zs->zs_pf_blkid, blkid + nblks);
+	pf_start = MAX(zs->zs_pf_blkid, end_of_access_blkid);
 
 	/*
 	 * Double our amount of prefetched data, but don't let the
 	 * prefetch get further ahead than zfetch_max_distance.
 	 */
-	int pf_nblks =
-	    MIN((int64_t)zs->zs_pf_blkid - zs->zs_blkid + nblks,
-	    zs->zs_blkid + nblks +
-	    (zfetch_max_distance >> zf->zf_dnode->dn_datablkshift) - pf_start);
+	if (fetch_data) {
+		max_dist_blks =
+		    zfetch_max_distance >> zf->zf_dnode->dn_datablkshift;
+		/*
+		 * Previously, we were (zs_pf_blkid - blkid) ahead.  We
+		 * want to now be double that, so read that amount again,
+		 * plus the amount we are catching up by (i.e. the amount
+		 * read just now).
+		 */
+		pf_ahead_blks = zs->zs_pf_blkid - blkid + nblks;
+		max_blks = max_dist_blks - (pf_start - end_of_access_blkid);
+		pf_nblks = MIN(pf_ahead_blks, max_blks);
+	} else {
+		pf_nblks = 0;
+	}
 
 	zs->zs_pf_blkid = pf_start + pf_nblks;
-	zs->zs_atime = gethrtime();
-	zs->zs_blkid = blkid + nblks;
 
 	/*
-	 * dbuf_prefetch() issues the prefetch i/o
-	 * asynchronously, but it may need to wait for an
-	 * indirect block to be read from disk.  Therefore
-	 * we do not want to hold any locks while we call it.
+	 * Do the same for indirects, starting from where we stopped last,
+	 * or where we will stop reading data blocks (and the indirects
+	 * that point to them).
 	 */
+	ipf_start = MAX(zs->zs_ipf_blkid, zs->zs_pf_blkid);
+	max_dist_blks = zfetch_max_idistance >> zf->zf_dnode->dn_datablkshift;
+	/*
+	 * We want to double our distance ahead of the data prefetch
+	 * (or reader, if we are not prefetching data).  Previously, we
+	 * were (zs_ipf_blkid - blkid) ahead.  To double that, we read
+	 * that amount again, plus the amount we are catching up by
+	 * (i.e. the amount read now + the amount of data prefetched now).
+	 */
+	pf_ahead_blks = zs->zs_ipf_blkid - blkid + nblks + pf_nblks;
+	max_blks = max_dist_blks - (ipf_start - end_of_access_blkid);
+	ipf_nblks = MIN(pf_ahead_blks, max_blks);
+	zs->zs_ipf_blkid = ipf_start + ipf_nblks;
+
+	epbs = zf->zf_dnode->dn_indblkshift - SPA_BLKPTRSHIFT;
+	ipf_istart = P2ROUNDUP(ipf_start, 1 << epbs) >> epbs;
+	ipf_iend = P2ROUNDUP(zs->zs_ipf_blkid, 1 << epbs) >> epbs;
+
+	zs->zs_atime = gethrtime();
+	zs->zs_blkid = end_of_access_blkid;
 	mutex_exit(&zs->zs_lock);
 	rw_exit(&zf->zf_rwlock);
+
+	/*
+	 * dbuf_prefetch() is asynchronous (even when it needs to read
+	 * indirect blocks), but we still prefer to drop our locks before
+	 * calling it to reduce the time we hold them.
+	 */
+
 	for (int i = 0; i < pf_nblks; i++) {
 		dbuf_prefetch(zf->zf_dnode, 0, pf_start + i,
 		    ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
 	}
+	for (int64_t iblk = ipf_istart; iblk < ipf_iend; iblk++) {
+		dbuf_prefetch(zf->zf_dnode, 1, iblk,
+		    ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
+	}
 	ZFETCHSTAT_BUMP(zfetchstat_hits);
 }

uts/common/fs/zfs/sys/dmu_zfetch.h

@@ -43,6 +43,13 @@ struct dnode;				/* so we can reference dnode */
 typedef struct zstream {
 	uint64_t	zs_blkid;	/* expect next access at this blkid */
 	uint64_t	zs_pf_blkid;	/* next block to prefetch */
+
+	/*
+	 * We will next prefetch the L1 indirect block of this level-0
+	 * block id.
+	 */
+	uint64_t	zs_ipf_blkid;
+
 	kmutex_t	zs_lock;	/* protects stream */
 	hrtime_t	zs_atime;	/* time last prefetch issued */
 	list_node_t	zs_node;	/* link for zf_stream */
@@ -59,7 +66,7 @@ void		zfetch_fini(void);
 
 void		dmu_zfetch_init(zfetch_t *, struct dnode *);
 void		dmu_zfetch_fini(zfetch_t *);
-void		dmu_zfetch(zfetch_t *, uint64_t, uint64_t);
+void		dmu_zfetch(zfetch_t *, uint64_t, uint64_t, boolean_t);
 
 
 #ifdef	__cplusplus

uts/common/fs/zfs/sys/dnode.h

@@ -305,6 +305,15 @@ int dnode_next_offset(dnode_t *dn, int flags, uint64_t *off,
 void dnode_evict_dbufs(dnode_t *dn);
 void dnode_evict_bonus(dnode_t *dn);
 
+#define	DNODE_IS_CACHEABLE(_dn)						\
+	((_dn)->dn_objset->os_primary_cache == ZFS_CACHE_ALL ||		\
+	(DMU_OT_IS_METADATA((_dn)->dn_type) &&				\
+	(_dn)->dn_objset->os_primary_cache == ZFS_CACHE_METADATA))
+
+#define	DNODE_META_IS_CACHEABLE(_dn)					\
+	((_dn)->dn_objset->os_primary_cache == ZFS_CACHE_ALL ||		\
+	(_dn)->dn_objset->os_primary_cache == ZFS_CACHE_METADATA)
+
 #ifdef ZFS_DEBUG
 
 /*