Illumos 5987 - zfs prefetch code needs work

5987 zfs prefetch code needs work
Reviewed by: Adam Leventhal <[email protected]>
Reviewed by: George Wilson <[email protected]>
Reviewed by: Paul Dagnelie <[email protected]>
Approved by: Gordon Ross <[email protected]>

References:
  https://www.illumos.org/issues/5987 zfs prefetch code needs work
  illumos/illumos-gate@cf6106c 5987 zfs prefetch code needs work

Porting notes:
- [module/zfs/dbuf.c]
  - 5f6d0b6 Handle block pointers with a corrupt logical size
- [module/zfs/dmu_zfetch.c]
  - c65aa5b Fix gcc missing parenthesis warnings
  - 428870f Update core ZFS code from build 121 to build 141.
  - 79c76d5 Change KM_PUSHPAGE -> KM_SLEEP
  - b8d06fc Switch KM_SLEEP to KM_PUSHPAGE
  - Account for ISO C90 - mixed declarations and code - warnings
  - Module parameters (new/changed):
    - Replaced zfetch_block_cap with zfetch_max_distance
      (Max bytes to prefetch per stream (default 8MB; 8 * 1024 * 1024))
    - Preserved zfs_prefetch_disable as 'int' for consistency with
      existing Linux module options.
- [include/sys/trace_arc.h]
  - Added new tracepoints
    - DEFINE_ARC_BUF_HDR_EVENT(zfs_arc__sync__wait__for__async);
    - DEFINE_ARC_BUF_HDR_EVENT(zfs_arc__demand__hit__predictive__prefetch);
- [man/man5/zfs-module-parameters.5]
  - Updated man page

Ported-by: kernelOfTruth [email protected]
Signed-off-by: Brian Behlendorf <[email protected]>

10 files changed, 280 insertions, 658 deletions

diff --git a/include/sys/arc.h b/include/sys/arc.h
index db7a64aa2..5780554b5 100644
--- a/include/sys/arc.h
+++ b/include/sys/arc.h
@@ -84,27 +84,31 @@ typedef enum arc_flags
 	ARC_FLAG_CACHED			= 1 << 4,	/* I/O was in cache */
 	ARC_FLAG_L2CACHE		= 1 << 5,	/* cache in L2ARC */
 	ARC_FLAG_L2COMPRESS		= 1 << 6,	/* compress in L2ARC */
+	ARC_FLAG_PREDICTIVE_PREFETCH	= 1 << 7,	/* I/O from zfetch */
 
 	/*
 	 * Private ARC flags.  These flags are private ARC only flags that
 	 * will show up in b_flags in the arc_hdr_buf_t. These flags should
 	 * only be set by ARC code.
 	 */
-	ARC_FLAG_IN_HASH_TABLE		= 1 << 7,	/* buffer is hashed */
-	ARC_FLAG_IO_IN_PROGRESS		= 1 << 8,	/* I/O in progress */
-	ARC_FLAG_IO_ERROR		= 1 << 9,	/* I/O failed for buf */
-	ARC_FLAG_FREED_IN_READ		= 1 << 10,	/* freed during read */
-	ARC_FLAG_BUF_AVAILABLE		= 1 << 11,	/* block not in use */
-	ARC_FLAG_INDIRECT		= 1 << 12,	/* indirect block */
-	ARC_FLAG_L2_WRITING		= 1 << 13,	/* write in progress */
-	ARC_FLAG_L2_EVICTED		= 1 << 14,	/* evicted during I/O */
-	ARC_FLAG_L2_WRITE_HEAD		= 1 << 15,	/* head of write list */
+	ARC_FLAG_IN_HASH_TABLE		= 1 << 8,	/* buffer is hashed */
+	ARC_FLAG_IO_IN_PROGRESS		= 1 << 9,	/* I/O in progress */
+	ARC_FLAG_IO_ERROR		= 1 << 10,	/* I/O failed for buf */
+	ARC_FLAG_FREED_IN_READ		= 1 << 11,	/* freed during read */
+	ARC_FLAG_BUF_AVAILABLE		= 1 << 12,	/* block not in use */
+	ARC_FLAG_INDIRECT		= 1 << 13,	/* indirect block */
+	/* Indicates that block was read with ASYNC priority. */
+	ARC_FLAG_PRIO_ASYNC_READ	= 1 << 14,
+	ARC_FLAG_L2_WRITING		= 1 << 15,	/* write in progress */
+	ARC_FLAG_L2_EVICTED		= 1 << 16,	/* evicted during I/O */
+	ARC_FLAG_L2_WRITE_HEAD		= 1 << 17,	/* head of write list */
 	/* indicates that the buffer contains metadata (otherwise, data) */
-	ARC_FLAG_BUFC_METADATA		= 1 << 16,
+	ARC_FLAG_BUFC_METADATA		= 1 << 18,
 
 	/* Flags specifying whether optional hdr struct fields are defined */
-	ARC_FLAG_HAS_L1HDR		= 1 << 17,
-	ARC_FLAG_HAS_L2HDR		= 1 << 18,
+	ARC_FLAG_HAS_L1HDR		= 1 << 19,
+	ARC_FLAG_HAS_L2HDR		= 1 << 20,
+
 } arc_flags_t;
 
 struct arc_buf {
diff --git a/include/sys/dmu.h b/include/sys/dmu.h
index 75d85951a..0a5b2809b 100644
--- a/include/sys/dmu.h
+++ b/include/sys/dmu.h
@@ -487,7 +487,8 @@ uint64_t dmu_buf_refcount(dmu_buf_t *db);
  * individually with dmu_buf_rele.
  */
 int dmu_buf_hold_array_by_bonus(dmu_buf_t *db, uint64_t offset,
-    uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp);
+    uint64_t length, boolean_t read, void *tag,
+    int *numbufsp, dmu_buf_t ***dbpp);
 void dmu_buf_rele_array(dmu_buf_t **, int numbufs, void *tag);
 
 typedef void dmu_buf_evict_func_t(void *user_ptr);
diff --git a/include/sys/dmu_zfetch.h b/include/sys/dmu_zfetch.h
index 38ed1d872..df33f182b 100644
--- a/include/sys/dmu_zfetch.h
+++ b/include/sys/dmu_zfetch.h
@@ -23,8 +23,12 @@
  * Use is subject to license terms.
  */
 
-#ifndef	_DFETCH_H
-#define	_DFETCH_H
+/*
+ * Copyright (c) 2014 by Delphix. All rights reserved.
+ */
+
+#ifndef	_DMU_ZFETCH_H
+#define	_DMU_ZFETCH_H
 
 #include <sys/zfs_context.h>
 
@@ -36,41 +40,30 @@ extern unsigned long	zfetch_array_rd_sz;
 
 struct dnode;				/* so we can reference dnode */
 
-typedef enum zfetch_dirn {
-	ZFETCH_FORWARD = 1,		/* prefetch increasing block numbers */
-	ZFETCH_BACKWARD	= -1		/* prefetch decreasing block numbers */
-} zfetch_dirn_t;
-
 typedef struct zstream {
-	uint64_t	zst_offset;	/* offset of starting block in range */
-	uint64_t	zst_len;	/* length of range, in blocks */
-	zfetch_dirn_t	zst_direction;	/* direction of prefetch */
-	uint64_t	zst_stride;	/* length of stride, in blocks */
-	uint64_t	zst_ph_offset;	/* prefetch offset, in blocks */
-	uint64_t	zst_cap;	/* prefetch limit (cap), in blocks */
-	kmutex_t	zst_lock;	/* protects stream */
-	clock_t		zst_last;	/* lbolt of last prefetch */
-	list_node_t	zst_node;	/* next zstream here */
+	uint64_t	zs_blkid;	/* expect next access at this blkid */
+	uint64_t	zs_pf_blkid;	/* next block to prefetch */
+	kmutex_t	zs_lock;	/* protects stream */
+	hrtime_t	zs_atime;	/* time last prefetch issued */
+	list_node_t	zs_node;	/* link for zf_stream */
 } zstream_t;
 
 typedef struct zfetch {
 	krwlock_t	zf_rwlock;	/* protects zfetch structure */
-	list_t		zf_stream;	/* AVL tree of zstream_t's */
+	list_t		zf_stream;	/* list of zstream_t's */
 	struct dnode	*zf_dnode;	/* dnode that owns this zfetch */
-	uint32_t	zf_stream_cnt;	/* # of active streams */
-	uint64_t	zf_alloc_fail;	/* # of failed attempts to alloc strm */
 } zfetch_t;
 
 void		zfetch_init(void);
 void		zfetch_fini(void);
 
 void		dmu_zfetch_init(zfetch_t *, struct dnode *);
-void		dmu_zfetch_rele(zfetch_t *);
-void		dmu_zfetch(zfetch_t *, uint64_t, uint64_t, int);
+void		dmu_zfetch_fini(zfetch_t *);
+void		dmu_zfetch(zfetch_t *, uint64_t, uint64_t);
 
 
 #ifdef	__cplusplus
 }
 #endif
 
-#endif	/* _DFETCH_H */
+#endif	/* _DMU_ZFETCH_H */
diff --git a/include/sys/trace_arc.h b/include/sys/trace_arc.h
index 31c3cdcb9..55dbdf19b 100644
--- a/include/sys/trace_arc.h
+++ b/include/sys/trace_arc.h
@@ -102,6 +102,8 @@ DEFINE_ARC_BUF_HDR_EVENT(zfs_arc__evict);
 DEFINE_ARC_BUF_HDR_EVENT(zfs_arc__delete);
 DEFINE_ARC_BUF_HDR_EVENT(zfs_new_state__mru);
 DEFINE_ARC_BUF_HDR_EVENT(zfs_new_state__mfu);
+DEFINE_ARC_BUF_HDR_EVENT(zfs_arc__sync__wait__for__async);
+DEFINE_ARC_BUF_HDR_EVENT(zfs_arc__demand__hit__predictive__prefetch);
 DEFINE_ARC_BUF_HDR_EVENT(zfs_l2arc__hit);
 DEFINE_ARC_BUF_HDR_EVENT(zfs_l2arc__miss);
 
diff --git a/man/man5/zfs-module-parameters.5 b/man/man5/zfs-module-parameters.5
index 6be382eac..eabd8ebe9 100644
--- a/man/man5/zfs-module-parameters.5
+++ b/man/man5/zfs-module-parameters.5
@@ -331,12 +331,12 @@ Default value: \fB1,048,576\fR.
 .sp
 .ne 2
 .na
-\fBzfetch_block_cap\fR (uint)
+\fBzfetch_max_distance\fR (uint)
 .ad
 .RS 12n
-Max number of blocks to prefetch at a time
+Max bytes to prefetch per stream (default 8MB).
 .sp
-Default value: \fB256\fR.
+Default value: \fB8,388,608\fR.
 .RE
 
 .sp
@@ -1246,7 +1246,10 @@ Default value: \fB52,428,800\fR.
 \fBzfs_prefetch_disable\fR (int)
 .ad
 .RS 12n
-Disable all ZFS prefetching
+This tunable disables predictive prefetch.  Note that it leaves "prescient"
+prefetch (e.g. prefetch for zfs send) intact.  Unlike predictive prefetch,
+prescient prefetch never issues i/os that end up not being needed, so it
+can't hurt performance.
 .sp
 Use \fB1\fR for yes and \fB0\fR for no (default).
 .RE
diff --git a/module/zfs/arc.c b/module/zfs/arc.c
index c3d88679f..b8707474d 100644
--- a/module/zfs/arc.c
+++ b/module/zfs/arc.c
@@ -474,6 +474,8 @@ typedef struct arc_stats {
 	kstat_named_t arcstat_meta_limit;
 	kstat_named_t arcstat_meta_max;
 	kstat_named_t arcstat_meta_min;
+	kstat_named_t arcstat_sync_wait_for_async;
+	kstat_named_t arcstat_demand_hit_predictive_prefetch;
 	kstat_named_t arcstat_need_free;
 	kstat_named_t arcstat_sys_free;
 } arc_stats_t;
@@ -568,6 +570,8 @@ static arc_stats_t arc_stats = {
 	{ "arc_meta_limit",		KSTAT_DATA_UINT64 },
 	{ "arc_meta_max",		KSTAT_DATA_UINT64 },
 	{ "arc_meta_min",		KSTAT_DATA_UINT64 },
+	{ "sync_wait_for_async",	KSTAT_DATA_UINT64 },
+	{ "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
 	{ "arc_need_free",		KSTAT_DATA_UINT64 },
 	{ "arc_sys_free",		KSTAT_DATA_UINT64 }
 };
@@ -4244,6 +4248,36 @@ top:
 
 		if (HDR_IO_IN_PROGRESS(hdr)) {
 
+			if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
+			    priority == ZIO_PRIORITY_SYNC_READ) {
+				/*
+				 * This sync read must wait for an
+				 * in-progress async read (e.g. a predictive
+				 * prefetch).  Async reads are queued
+				 * separately at the vdev_queue layer, so
+				 * this is a form of priority inversion.
+				 * Ideally, we would "inherit" the demand
+				 * i/o's priority by moving the i/o from
+				 * the async queue to the synchronous queue,
+				 * but there is currently no mechanism to do
+				 * so.  Track this so that we can evaluate
+				 * the magnitude of this potential performance
+				 * problem.
+				 *
+				 * Note that if the prefetch i/o is already
+				 * active (has been issued to the device),
+				 * the prefetch improved performance, because
+				 * we issued it sooner than we would have
+				 * without the prefetch.
+				 */
+				DTRACE_PROBE1(arc__sync__wait__for__async,
+				    arc_buf_hdr_t *, hdr);
+				ARCSTAT_BUMP(arcstat_sync_wait_for_async);
+			}
+			if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
+				hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH;
+			}
+
 			if (*arc_flags & ARC_FLAG_WAIT) {
 				cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
 				mutex_exit(hash_lock);
@@ -4252,7 +4286,7 @@ top:
 			ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
 
 			if (done) {
-				arc_callback_t	*acb = NULL;
+				arc_callback_t *acb = NULL;
 
 				acb = kmem_zalloc(sizeof (arc_callback_t),
 				    KM_SLEEP);
@@ -4277,6 +4311,19 @@ top:
 		    hdr->b_l1hdr.b_state == arc_mfu);
 
 		if (done) {
+			if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
+				/*
+				 * This is a demand read which does not have to
+				 * wait for i/o because we did a predictive
+				 * prefetch i/o for it, which has completed.
+				 */
+				DTRACE_PROBE1(
+				    arc__demand__hit__predictive__prefetch,
+				    arc_buf_hdr_t *, hdr);
+				ARCSTAT_BUMP(
+				    arcstat_demand_hit_predictive_prefetch);
+				hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH;
+			}
 			add_reference(hdr, hash_lock, private);
 			/*
 			 * If this block is already in use, create a new
@@ -4349,12 +4396,16 @@ top:
 				goto top; /* restart the IO request */
 			}
 
-			/* if this is a prefetch, we don't have a reference */
-			if (*arc_flags & ARC_FLAG_PREFETCH) {
+			/*
+			 * If there is a callback, we pass our reference to
+			 * it; otherwise we remove our reference.
+			 */
+			if (done == NULL) {
 				(void) remove_reference(hdr, hash_lock,
 				    private);
-				hdr->b_flags |= ARC_FLAG_PREFETCH;
 			}
+			if (*arc_flags & ARC_FLAG_PREFETCH)
+				hdr->b_flags |= ARC_FLAG_PREFETCH;
 			if (*arc_flags & ARC_FLAG_L2CACHE)
 				hdr->b_flags |= ARC_FLAG_L2CACHE;
 			if (*arc_flags & ARC_FLAG_L2COMPRESS)
@@ -4377,11 +4428,13 @@ top:
 			ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
 			ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
 
-			/* if this is a prefetch, we don't have a reference */
+			/*
+			 * If there is a callback, we pass a reference to it.
+			 */
+			if (done != NULL)
+				add_reference(hdr, hash_lock, private);
 			if (*arc_flags & ARC_FLAG_PREFETCH)
 				hdr->b_flags |= ARC_FLAG_PREFETCH;
-			else
-				add_reference(hdr, hash_lock, private);
 			if (*arc_flags & ARC_FLAG_L2CACHE)
 				hdr->b_flags |= ARC_FLAG_L2CACHE;
 			if (*arc_flags & ARC_FLAG_L2COMPRESS)
@@ -4399,6 +4452,8 @@ top:
 			arc_access(hdr, hash_lock);
 		}
 
+		if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
+			hdr->b_flags |= ARC_FLAG_PREDICTIVE_PREFETCH;
 		ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
 
 		acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
@@ -4438,6 +4493,11 @@ top:
 		    demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
 		    data, metadata, misses);
 
+		if (priority == ZIO_PRIORITY_ASYNC_READ)
+			hdr->b_flags |= ARC_FLAG_PRIO_ASYNC_READ;
+		else
+			hdr->b_flags &= ~ARC_FLAG_PRIO_ASYNC_READ;
+
 		if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
 			/*
 			 * Read from the L2ARC if the following are true:
diff --git a/module/zfs/dbuf.c b/module/zfs/dbuf.c
index e08dcc4a3..d53d8607a 100644
--- a/module/zfs/dbuf.c
+++ b/module/zfs/dbuf.c
@@ -676,7 +676,7 @@ dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
 }
 
 static int
-dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
+dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
 {
 	dnode_t *dn;
 	zbookmark_phys_t zb;
@@ -723,7 +723,6 @@ dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
 		    db->db.db_size, db, type));
 		bzero(db->db.db_data, db->db.db_size);
 		db->db_state = DB_CACHED;
-		*flags |= DB_RF_CACHED;
 		mutex_exit(&db->db_mtx);
 		return (0);
 	}
@@ -746,10 +745,8 @@ dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
 
 	err = arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
 	    dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
-	    (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
+	    (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
 	    &aflags, &zb);
-	if (aflags & ARC_FLAG_CACHED)
-		*flags |= DB_RF_CACHED;
 
 	return (SET_ERROR(err));
 }
@@ -784,8 +781,7 @@ 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.db_offset,
-			    db->db.db_size, TRUE);
+			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1);
 		if ((flags & DB_RF_HAVESTRUCT) == 0)
 			rw_exit(&dn->dn_struct_rwlock);
 		DB_DNODE_EXIT(db);
@@ -795,13 +791,12 @@ dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
 		if (zio == NULL)
 			zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
 
-		err = dbuf_read_impl(db, zio, &flags);
+		err = dbuf_read_impl(db, zio, flags);
 
 		/* dbuf_read_impl has dropped db_mtx for us */
 
 		if (!err && prefetch)
-			dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
-			    db->db.db_size, flags & DB_RF_CACHED);
+			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1);
 
 		if ((flags & DB_RF_HAVESTRUCT) == 0)
 			rw_exit(&dn->dn_struct_rwlock);
@@ -820,8 +815,7 @@ 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.db_offset,
-			    db->db.db_size, TRUE);
+			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1);
 		if ((flags & DB_RF_HAVESTRUCT) == 0)
 			rw_exit(&dn->dn_struct_rwlock);
 		DB_DNODE_EXIT(db);
@@ -2143,6 +2137,9 @@ dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
 	ASSERT(blkid != DMU_BONUS_BLKID);
 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
 
+	if (blkid > dn->dn_maxblkid)
+		return;
+
 	if (dnode_block_freed(dn, blkid))
 		return;
 
diff --git a/module/zfs/dmu.c b/module/zfs/dmu.c
index b4133f0e4..1bf108bfe 100644
--- a/module/zfs/dmu.c
+++ b/module/zfs/dmu.c
@@ -20,7 +20,7 @@
  */
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
- * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
+ * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
  * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
  * Copyright (c) 2014, Nexenta Systems, Inc. All rights reserved.
  * Copyright (c) 2015 by Chunwei Chen. All rights reserved.
@@ -386,7 +386,7 @@ dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
  */
 static int
 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
-    int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
+    boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
 {
 	dmu_buf_t **dbp;
 	uint64_t blkid, nblks, i;
@@ -396,15 +396,19 @@ dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
 
 	ASSERT(length <= DMU_MAX_ACCESS);
 
-	dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
-	if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
-		dbuf_flags |= DB_RF_NOPREFETCH;
+	/*
+	 * Note: We directly notify the prefetch code of this read, so that
+	 * we can tell it about the multi-block read.  dbuf_read() only knows
+	 * about the one block it is accessing.
+	 */
+	dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
+	    DB_RF_NOPREFETCH;
 
 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
 	if (dn->dn_datablkshift) {
 		int blkshift = dn->dn_datablkshift;
-		nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
-		    P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
+		nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
+		    P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
 	} else {
 		if (offset + length > dn->dn_datablksz) {
 			zfs_panic_recover("zfs: accessing past end of object "
@@ -423,19 +427,24 @@ dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
 	zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
 	blkid = dbuf_whichblock(dn, 0, offset);
 	for (i = 0; i < nblks; i++) {
-		dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
+		dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
 		if (db == NULL) {
 			rw_exit(&dn->dn_struct_rwlock);
 			dmu_buf_rele_array(dbp, nblks, tag);
 			zio_nowait(zio);
 			return (SET_ERROR(EIO));
 		}
+
 		/* initiate async i/o */
-		if (read) {
+		if (read)
 			(void) dbuf_read(db, zio, dbuf_flags);
-		}
 		dbp[i] = &db->db;
 	}
+
+	if ((flags & DMU_READ_NO_PREFETCH) == 0 && read &&
+	    length < zfetch_array_rd_sz) {
+		dmu_zfetch(&dn->dn_zfetch, blkid, nblks);
+	}
 	rw_exit(&dn->dn_struct_rwlock);
 
 	/* wait for async i/o */
@@ -489,7 +498,8 @@ dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
 
 int
 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
-    uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
+    uint64_t length, boolean_t read, void *tag, int *numbufsp,
+    dmu_buf_t ***dbpp)
 {
 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
 	dnode_t *dn;
@@ -537,9 +547,6 @@ dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
 	uint64_t blkid;
 	int nblks, err;
 
-	if (zfs_prefetch_disable)
-		return;
-
 	if (len == 0) {  /* they're interested in the bonus buffer */
 		dn = DMU_META_DNODE(os);
 
diff --git a/module/zfs/dmu_zfetch.c b/module/zfs/dmu_zfetch.c
index 9d41bdc0f..b25b0e007 100644
--- a/module/zfs/dmu_zfetch.c
+++ b/module/zfs/dmu_zfetch.c
@@ -24,7 +24,7 @@
  */
 
 /*
- * Copyright (c) 2013 by Delphix. All rights reserved.
+ * Copyright (c) 2013, 2014 by Delphix. All rights reserved.
  */
 
 #include <sys/zfs_context.h>
@@ -36,209 +36,43 @@
 #include <sys/kstat.h>
 
 /*
- * I'm against tune-ables, but these should probably exist as tweakable globals
- * until we can get this working the way we want it to.
+ * This tunable disables predictive prefetch.  Note that it leaves "prescient"
+ * prefetch (e.g. prefetch for zfs send) intact.  Unlike predictive prefetch,
+ * prescient prefetch never issues i/os that end up not being needed,
+ * so it can't hurt performance.
  */
 
-int zfs_prefetch_disable = 0;
+int zfs_prefetch_disable = B_FALSE;
 
 /* max # of streams per zfetch */
 unsigned int	zfetch_max_streams = 8;
 /* min time before stream reclaim */
 unsigned int	zfetch_min_sec_reap = 2;
-/* max number of blocks to fetch at a time */
-unsigned int	zfetch_block_cap = 256;
-/* number of bytes in a array_read at which we stop prefetching (1Mb) */
+/* max bytes to prefetch per stream (default 8MB) */
+unsigned int	zfetch_max_distance = 8 * 1024 * 1024;
+/* number of bytes in a array_read at which we stop prefetching (1MB) */
 unsigned long	zfetch_array_rd_sz = 1024 * 1024;
 
-/* forward decls for static routines */
-static boolean_t	dmu_zfetch_colinear(zfetch_t *, zstream_t *);
-static void		dmu_zfetch_dofetch(zfetch_t *, zstream_t *);
-static uint64_t		dmu_zfetch_fetch(dnode_t *, uint64_t, uint64_t);
-static uint64_t		dmu_zfetch_fetchsz(dnode_t *, uint64_t, uint64_t);
-static boolean_t	dmu_zfetch_find(zfetch_t *, zstream_t *, int);
-static int		dmu_zfetch_stream_insert(zfetch_t *, zstream_t *);
-static zstream_t	*dmu_zfetch_stream_reclaim(zfetch_t *);
-static void		dmu_zfetch_stream_remove(zfetch_t *, zstream_t *);
-static int		dmu_zfetch_streams_equal(zstream_t *, zstream_t *);
-
 typedef struct zfetch_stats {
 	kstat_named_t zfetchstat_hits;
 	kstat_named_t zfetchstat_misses;
-	kstat_named_t zfetchstat_colinear_hits;
-	kstat_named_t zfetchstat_colinear_misses;
-	kstat_named_t zfetchstat_stride_hits;
-	kstat_named_t zfetchstat_stride_misses;
-	kstat_named_t zfetchstat_reclaim_successes;
-	kstat_named_t zfetchstat_reclaim_failures;
-	kstat_named_t zfetchstat_stream_resets;
-	kstat_named_t zfetchstat_stream_noresets;
-	kstat_named_t zfetchstat_bogus_streams;
+	kstat_named_t zfetchstat_max_streams;
 } zfetch_stats_t;
 
 static zfetch_stats_t zfetch_stats = {
 	{ "hits",			KSTAT_DATA_UINT64 },
 	{ "misses",			KSTAT_DATA_UINT64 },
-	{ "colinear_hits",		KSTAT_DATA_UINT64 },
-	{ "colinear_misses",		KSTAT_DATA_UINT64 },
-	{ "stride_hits",		KSTAT_DATA_UINT64 },
-	{ "stride_misses",		KSTAT_DATA_UINT64 },
-	{ "reclaim_successes",		KSTAT_DATA_UINT64 },
-	{ "reclaim_failures",		KSTAT_DATA_UINT64 },
-	{ "streams_resets",		KSTAT_DATA_UINT64 },
-	{ "streams_noresets",		KSTAT_DATA_UINT64 },
-	{ "bogus_streams",		KSTAT_DATA_UINT64 },
+	{ "max_streams",		KSTAT_DATA_UINT64 },
 };
 
-#define	ZFETCHSTAT_INCR(stat, val) \
-	atomic_add_64(&zfetch_stats.stat.value.ui64, (val));
-
-#define	ZFETCHSTAT_BUMP(stat)		ZFETCHSTAT_INCR(stat, 1);
+#define	ZFETCHSTAT_BUMP(stat) \
+	atomic_inc_64(&zfetch_stats.stat.value.ui64);
 
 kstat_t		*zfetch_ksp;
 
-/*
- * Given a zfetch structure and a zstream structure, determine whether the
- * blocks to be read are part of a co-linear pair of existing prefetch
- * streams.  If a set is found, coalesce the streams, removing one, and
- * configure the prefetch so it looks for a strided access pattern.
- *
- * In other words: if we find two sequential access streams that are
- * the same length and distance N appart, and this read is N from the
- * last stream, then we are probably in a strided access pattern.  So
- * combine the two sequential streams into a single strided stream.
- *
- * Returns whether co-linear streams were found.
- */
-static boolean_t
-dmu_zfetch_colinear(zfetch_t *zf, zstream_t *zh)
-{
-	zstream_t	*z_walk;
-	zstream_t	*z_comp;
-
-	if (! rw_tryenter(&zf->zf_rwlock, RW_WRITER))
-		return (0);
-
-	if (zh == NULL) {
-		rw_exit(&zf->zf_rwlock);
-		return (0);
-	}
-
-	for (z_walk = list_head(&zf->zf_stream); z_walk;
-	    z_walk = list_next(&zf->zf_stream, z_walk)) {
-		for (z_comp = list_next(&zf->zf_stream, z_walk); z_comp;
-		    z_comp = list_next(&zf->zf_stream, z_comp)) {
-			int64_t		diff;
-
-			if (z_walk->zst_len != z_walk->zst_stride ||
-			    z_comp->zst_len != z_comp->zst_stride) {
-				continue;
-			}
-
-			diff = z_comp->zst_offset - z_walk->zst_offset;
-			if (z_comp->zst_offset + diff == zh->zst_offset) {
-				z_walk->zst_offset = zh->zst_offset;
-				z_walk->zst_direction = diff < 0 ?
-				    ZFETCH_BACKWARD : ZFETCH_FORWARD;
-				z_walk->zst_stride =
-				    diff * z_walk->zst_direction;
-				z_walk->zst_ph_offset =
-				    zh->zst_offset + z_walk->zst_stride;
-				dmu_zfetch_stream_remove(zf, z_comp);
-				mutex_destroy(&z_comp->zst_lock);
-				kmem_free(z_comp, sizeof (zstream_t));
-
-				dmu_zfetch_dofetch(zf, z_walk);
-
-				rw_exit(&zf->zf_rwlock);
-				return (1);
-			}
-
-			diff = z_walk->zst_offset - z_comp->zst_offset;
-			if (z_walk->zst_offset + diff == zh->zst_offset) {
-				z_walk->zst_offset = zh->zst_offset;
-				z_walk->zst_direction = diff < 0 ?
-				    ZFETCH_BACKWARD : ZFETCH_FORWARD;
-				z_walk->zst_stride =
-				    diff * z_walk->zst_direction;
-				z_walk->zst_ph_offset =
-				    zh->zst_offset + z_walk->zst_stride;
-				dmu_zfetch_stream_remove(zf, z_comp);
-				mutex_destroy(&z_comp->zst_lock);
-				kmem_free(z_comp, sizeof (zstream_t));
-
-				dmu_zfetch_dofetch(zf, z_walk);
-
-				rw_exit(&zf->zf_rwlock);
-				return (1);
-			}
-		}
-	}
-
-	rw_exit(&zf->zf_rwlock);
-	return (0);
-}
-
-/*
- * Given a zstream_t, determine the bounds of the prefetch.  Then call the
- * routine that actually prefetches the individual blocks.
- */
-static void
-dmu_zfetch_dofetch(zfetch_t *zf, zstream_t *zs)
-{
-	uint64_t	prefetch_tail;
-	uint64_t	prefetch_limit;
-	uint64_t	prefetch_ofst;
-	uint64_t	prefetch_len;
-	uint64_t	blocks_fetched;
-
-	zs->zst_stride = MAX((int64_t)zs->zst_stride, zs->zst_len);
-	zs->zst_cap = MIN(zfetch_block_cap, 2 * zs->zst_cap);
-
-	prefetch_tail = MAX((int64_t)zs->zst_ph_offset,
-	    (int64_t)(zs->zst_offset + zs->zst_stride));
-	/*
-	 * XXX: use a faster division method?
-	 */
-	prefetch_limit = zs->zst_offset + zs->zst_len +
-	    (zs->zst_cap * zs->zst_stride) / zs->zst_len;
-
-	while (prefetch_tail < prefetch_limit) {
-		prefetch_ofst = zs->zst_offset + zs->zst_direction *
-		    (prefetch_tail - zs->zst_offset);
-
-		prefetch_len = zs->zst_len;
-
-		/*
-		 * Don't prefetch beyond the end of the file, if working
-		 * backwards.
-		 */
-		if ((zs->zst_direction == ZFETCH_BACKWARD) &&
-		    (prefetch_ofst > prefetch_tail)) {
-			prefetch_len += prefetch_ofst;
-			prefetch_ofst = 0;
-		}
-
-		/* don't prefetch more than we're supposed to */
-		if (prefetch_len > zs->zst_len)
-			break;
-
-		blocks_fetched = dmu_zfetch_fetch(zf->zf_dnode,
-		    prefetch_ofst, zs->zst_len);
-
-		prefetch_tail += zs->zst_stride;
-		/* stop if we've run out of stuff to prefetch */
-		if (blocks_fetched < zs->zst_len)
-			break;
-	}
-	zs->zst_ph_offset = prefetch_tail;
-	zs->zst_last = ddi_get_lbolt();
-}
-
 void
 zfetch_init(void)
 {
-
 	zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc",
 	    KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t),
 	    KSTAT_FLAG_VIRTUAL);
@@ -266,273 +100,41 @@ zfetch_fini(void)
 void
 dmu_zfetch_init(zfetch_t *zf, dnode_t *dno)
 {
-	if (zf == NULL) {
+	if (zf == NULL)
 		return;
-	}
 
 	zf->zf_dnode = dno;
-	zf->zf_stream_cnt = 0;
-	zf->zf_alloc_fail = 0;
 
 	list_create(&zf->zf_stream, sizeof (zstream_t),
-	    offsetof(zstream_t, zst_node));
+	    offsetof(zstream_t, zs_node));
 
 	rw_init(&zf->zf_rwlock, NULL, RW_DEFAULT, NULL);
 }
 
-/*
- * This function computes the actual size, in blocks, that can be prefetched,
- * and fetches it.
- */
-static uint64_t
-dmu_zfetch_fetch(dnode_t *dn, uint64_t blkid, uint64_t nblks)
-{
-	uint64_t	fetchsz;
-	uint64_t	i;
-
-	fetchsz = dmu_zfetch_fetchsz(dn, blkid, nblks);
-
-	for (i = 0; i < fetchsz; i++) {
-		dbuf_prefetch(dn, 0, blkid + i, ZIO_PRIORITY_ASYNC_READ,
-		    ARC_FLAG_PREFETCH);
-	}
-
-	return (fetchsz);
-}
-
-/*
- * this function returns the number of blocks that would be prefetched, based
- * upon the supplied dnode, blockid, and nblks.  This is used so that we can
- * update streams in place, and then prefetch with their old value after the
- * fact.  This way, we can delay the prefetch, but subsequent accesses to the
- * stream won't result in the same data being prefetched multiple times.
- */
-static uint64_t
-dmu_zfetch_fetchsz(dnode_t *dn, uint64_t blkid, uint64_t nblks)
-{
-	uint64_t	fetchsz;
-
-	if (blkid > dn->dn_maxblkid) {
-		return (0);
-	}
-
-	/* compute fetch size */
-	if (blkid + nblks + 1 > dn->dn_maxblkid) {
-		fetchsz = (dn->dn_maxblkid - blkid) + 1;
-		ASSERT(blkid + fetchsz - 1 <= dn->dn_maxblkid);
-	} else {
-		fetchsz = nblks;
-	}
-
-
-	return (fetchsz);
-}
-
-/*
- * given a zfetch and a zstream structure, see if there is an associated zstream
- * for this block read.  If so, it starts a prefetch for the stream it
- * located and returns true, otherwise it returns false
- */
-static boolean_t
-dmu_zfetch_find(zfetch_t *zf, zstream_t *zh, int prefetched)
+static void
+dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
 {
-	zstream_t	*zs;
-	int64_t		diff;
-	int		reset = !prefetched;
-	int		rc = 0;
-
-	if (zh == NULL)
-		return (0);
-
-	/*
-	 * XXX: This locking strategy is a bit coarse; however, it's impact has
-	 * yet to be tested.  If this turns out to be an issue, it can be
-	 * modified in a number of different ways.
-	 */
-
-	rw_enter(&zf->zf_rwlock, RW_READER);
-top:
-
-	for (zs = list_head(&zf->zf_stream); zs;
-	    zs = list_next(&zf->zf_stream, zs)) {
-
-		/*
-		 * XXX - should this be an assert?
-		 */
-		if (zs->zst_len == 0) {
-			/* bogus stream */
-			ZFETCHSTAT_BUMP(zfetchstat_bogus_streams);
-			continue;
-		}
-
-		/*
-		 * We hit this case when we are in a strided prefetch stream:
-		 * we will read "len" blocks before "striding".
-		 */
-		if (zh->zst_offset >= zs->zst_offset &&
-		    zh->zst_offset < zs->zst_offset + zs->zst_len) {
-			if (prefetched) {
-				/* already fetched */
-				ZFETCHSTAT_BUMP(zfetchstat_stride_hits);
-				rc = 1;
-				goto out;
-			} else {
-				ZFETCHSTAT_BUMP(zfetchstat_stride_misses);
-			}
-		}
-
-		/*
-		 * This is the forward sequential read case: we increment
-		 * len by one each time we hit here, so we will enter this
-		 * case on every read.
-		 */
-		if (zh->zst_offset == zs->zst_offset + zs->zst_len) {
-
-			reset = !prefetched && zs->zst_len > 1;
-
-			mutex_enter(&zs->zst_lock);
-
-			if (zh->zst_offset != zs->zst_offset + zs->zst_len) {
-				mutex_exit(&zs->zst_lock);
-				goto top;
-			}
-			zs->zst_len += zh->zst_len;
-			diff = zs->zst_len - zfetch_block_cap;
-			if (diff > 0) {
-				zs->zst_offset += diff;
-				zs->zst_len = zs->zst_len > diff ?
-				    zs->zst_len - diff : 0;
-			}
-			zs->zst_direction = ZFETCH_FORWARD;
-
-			break;
-
-		/*
-		 * Same as above, but reading backwards through the file.
-		 */
-		} else if (zh->zst_offset == zs->zst_offset - zh->zst_len) {
-			/* backwards sequential access */
-
-			reset = !prefetched && zs->zst_len > 1;
-
-			mutex_enter(&zs->zst_lock);
-
-			if (zh->zst_offset != zs->zst_offset - zh->zst_len) {
-				mutex_exit(&zs->zst_lock);
-				goto top;
-			}
-
-			zs->zst_offset = zs->zst_offset > zh->zst_len ?
-			    zs->zst_offset - zh->zst_len : 0;
-			zs->zst_ph_offset = zs->zst_ph_offset > zh->zst_len ?
-			    zs->zst_ph_offset - zh->zst_len : 0;
-			zs->zst_len += zh->zst_len;
-
-			diff = zs->zst_len - zfetch_block_cap;
-			if (diff > 0) {
-				zs->zst_ph_offset = zs->zst_ph_offset > diff ?
-				    zs->zst_ph_offset - diff : 0;
-				zs->zst_len = zs->zst_len > diff ?
-				    zs->zst_len - diff : zs->zst_len;
-			}
-			zs->zst_direction = ZFETCH_BACKWARD;
-
-			break;
-
-		} else if ((zh->zst_offset - zs->zst_offset - zs->zst_stride <
-		    zs->zst_len) && (zs->zst_len != zs->zst_stride)) {
-			/* strided forward access */
-
-			mutex_enter(&zs->zst_lock);
-
-			if ((zh->zst_offset - zs->zst_offset - zs->zst_stride >=
-			    zs->zst_len) || (zs->zst_len == zs->zst_stride)) {
-				mutex_exit(&zs->zst_lock);
-				goto top;
-			}
-
-			zs->zst_offset += zs->zst_stride;
-			zs->zst_direction = ZFETCH_FORWARD;
-
-			break;
-
-		} else if ((zh->zst_offset - zs->zst_offset + zs->zst_stride <
-		    zs->zst_len) && (zs->zst_len != zs->zst_stride)) {
-			/* strided reverse access */
-
-			mutex_enter(&zs->zst_lock);
-
-			if ((zh->zst_offset - zs->zst_offset + zs->zst_stride >=
-			    zs->zst_len) || (zs->zst_len == zs->zst_stride)) {
-				mutex_exit(&zs->zst_lock);
-				goto top;
-			}
-
-			zs->zst_offset = zs->zst_offset > zs->zst_stride ?
-			    zs->zst_offset - zs->zst_stride : 0;
-			zs->zst_ph_offset = (zs->zst_ph_offset >
-			    (2 * zs->zst_stride)) ?
-			    (zs->zst_ph_offset - (2 * zs->zst_stride)) : 0;
-			zs->zst_direction = ZFETCH_BACKWARD;
-
-			break;
-		}
-	}
-
-	if (zs) {
-		if (reset) {
-			zstream_t *remove = zs;
-
-			ZFETCHSTAT_BUMP(zfetchstat_stream_resets);
-			rc = 0;
-			mutex_exit(&zs->zst_lock);
-			rw_exit(&zf->zf_rwlock);
-			rw_enter(&zf->zf_rwlock, RW_WRITER);
-			/*
-			 * Relocate the stream, in case someone removes
-			 * it while we were acquiring the WRITER lock.
-			 */
-			for (zs = list_head(&zf->zf_stream); zs;
-			    zs = list_next(&zf->zf_stream, zs)) {
-				if (zs == remove) {
-					dmu_zfetch_stream_remove(zf, zs);
-					mutex_destroy(&zs->zst_lock);
-					kmem_free(zs, sizeof (zstream_t));
-					break;
-				}
-			}
-		} else {
-			ZFETCHSTAT_BUMP(zfetchstat_stream_noresets);
-			rc = 1;
-			dmu_zfetch_dofetch(zf, zs);
-			mutex_exit(&zs->zst_lock);
-		}
-	}
-out:
-	rw_exit(&zf->zf_rwlock);
-	return (rc);
+	ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
+	list_remove(&zf->zf_stream, zs);
+	mutex_destroy(&zs->zs_lock);
+	kmem_free(zs, sizeof (*zs));
 }
 
 /*
- * Clean-up state associated with a zfetch structure.  This frees allocated
- * structure members, empties the zf_stream tree, and generally makes things
- * nice.  This doesn't free the zfetch_t itself, that's left to the caller.
+ * Clean-up state associated with a zfetch structure (e.g. destroy the
+ * streams).  This doesn't free the zfetch_t itself, that's left to the caller.
  */
 void
-dmu_zfetch_rele(zfetch_t *zf)
+dmu_zfetch_fini(zfetch_t *zf)
 {
-	zstream_t	*zs;
-	zstream_t	*zs_next;
+	zstream_t *zs;
 
 	ASSERT(!RW_LOCK_HELD(&zf->zf_rwlock));
 
-	for (zs = list_head(&zf->zf_stream); zs; zs = zs_next) {
-		zs_next = list_next(&zf->zf_stream, zs);
-
-		list_remove(&zf->zf_stream, zs);
-		mutex_destroy(&zs->zst_lock);
-		kmem_free(zs, sizeof (zstream_t));
-	}
+	rw_enter(&zf->zf_rwlock, RW_WRITER);
+	while ((zs = list_head(&zf->zf_stream)) != NULL)
+		dmu_zfetch_stream_remove(zf, zs);
+	rw_exit(&zf->zf_rwlock);
 	list_destroy(&zf->zf_stream);
 	rw_destroy(&zf->zf_rwlock);
 
@@ -540,101 +142,57 @@ dmu_zfetch_rele(zfetch_t *zf)
 }
 
 /*
- * Given a zfetch and zstream structure, insert the zstream structure into the
- * AVL tree contained within the zfetch structure.  Peform the appropriate
- * book-keeping.  It is possible that another thread has inserted a stream which
- * matches one that we are about to insert, so we must be sure to check for this
- * case.  If one is found, return failure, and let the caller cleanup the
- * duplicates.
+ * If there aren't too many streams already, create a new stream.
+ * The "blkid" argument is the next block that we expect this stream to access.
+ * While we're here, clean up old streams (which haven't been
+ * accessed for at least zfetch_min_sec_reap seconds).
  */
-static int
-dmu_zfetch_stream_insert(zfetch_t *zf, zstream_t *zs)
+static void
+dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid)
 {
-	zstream_t	*zs_walk;
-	zstream_t	*zs_next;
+	zstream_t *zs;
+	zstream_t *zs_next;
+	int numstreams = 0;
+	uint32_t max_streams;
 
 	ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
 
-	for (zs_walk = list_head(&zf->zf_stream); zs_walk; zs_walk = zs_next) {
-		zs_next = list_next(&zf->zf_stream, zs_walk);
-
-		if (dmu_zfetch_streams_equal(zs_walk, zs)) {
-			return (0);
-		}
-	}
-
-	list_insert_head(&zf->zf_stream, zs);
-	zf->zf_stream_cnt++;
-	return (1);
-}
-
-
-/*
- * Walk the list of zstreams in the given zfetch, find an old one (by time), and
- * reclaim it for use by the caller.
- */
-static zstream_t *
-dmu_zfetch_stream_reclaim(zfetch_t *zf)
-{
-	zstream_t	*zs;
-
-	if (! rw_tryenter(&zf->zf_rwlock, RW_WRITER))
-		return (0);
-
-	for (zs = list_head(&zf->zf_stream); zs;
-	    zs = list_next(&zf->zf_stream, zs)) {
-
-		if (((ddi_get_lbolt() - zs->zst_last)/hz) > zfetch_min_sec_reap)
-			break;
+	/*
+	 * Clean up old streams.
+	 */
+	for (zs = list_head(&zf->zf_stream);
+	    zs != NULL; zs = zs_next) {
+		zs_next = list_next(&zf->zf_stream, zs);
+		if (((gethrtime() - zs->zs_atime) / NANOSEC) >
+		    zfetch_min_sec_reap)
+			dmu_zfetch_stream_remove(zf, zs);
+		else
+			numstreams++;
 	}
 
-	if (zs) {
-		dmu_zfetch_stream_remove(zf, zs);
-		mutex_destroy(&zs->zst_lock);
-		bzero(zs, sizeof (zstream_t));
-	} else {
-		zf->zf_alloc_fail++;
+	/*
+	 * The maximum number of streams is normally zfetch_max_streams,
+	 * but for small files we lower it such that it's at least possible
+	 * for all the streams to be non-overlapping.
+	 *
+	 * If we are already at the maximum number of streams for this file,
+	 * even after removing old streams, then don't create this stream.
+	 */
+	max_streams = MAX(1, MIN(zfetch_max_streams,
+	    zf->zf_dnode->dn_maxblkid * zf->zf_dnode->dn_datablksz /
+	    zfetch_max_distance));
+	if (numstreams >= max_streams) {
+		ZFETCHSTAT_BUMP(zfetchstat_max_streams);
+		return;
 	}
-	rw_exit(&zf->zf_rwlock);
-
-	return (zs);
-}
 
-/*
- * Given a zfetch and zstream structure, remove the zstream structure from its
- * container in the zfetch structure.  Perform the appropriate book-keeping.
- */
-static void
-dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
-{
-	ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
+	zs = kmem_zalloc(sizeof (*zs), KM_SLEEP);
+	zs->zs_blkid = blkid;
+	zs->zs_pf_blkid = blkid;
+	zs->zs_atime = gethrtime();
+	mutex_init(&zs->zs_lock, NULL, MUTEX_DEFAULT, NULL);
 
-	list_remove(&zf->zf_stream, zs);
-	zf->zf_stream_cnt--;
-}
-
-static int
-dmu_zfetch_streams_equal(zstream_t *zs1, zstream_t *zs2)
-{
-	if (zs1->zst_offset != zs2->zst_offset)
-		return (0);
-
-	if (zs1->zst_len != zs2->zst_len)
-		return (0);
-
-	if (zs1->zst_stride != zs2->zst_stride)
-		return (0);
-
-	if (zs1->zst_ph_offset != zs2->zst_ph_offset)
-		return (0);
-
-	if (zs1->zst_cap != zs2->zst_cap)
-		return (0);
-
-	if (zs1->zst_direction != zs2->zst_direction)
-		return (0);
-
-	return (1);
+	list_insert_head(&zf->zf_stream, zs);
 }
 
 /*
@@ -642,93 +200,91 @@ dmu_zfetch_streams_equal(zstream_t *zs1, zstream_t *zs2)
  * routines to create, delete, find, or operate upon prefetch streams.
  */
 void
-dmu_zfetch(zfetch_t *zf, uint64_t offset, uint64_t size, int prefetched)
+dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks)
 {
-	zstream_t	zst;
-	zstream_t	*newstream;
-	boolean_t	fetched;
-	int		inserted;
-	unsigned int	blkshft;
-	uint64_t	blksz;
+	zstream_t *zs;
+	int64_t pf_start;
+	int pf_nblks;
+	int i;
 
 	if (zfs_prefetch_disable)
 		return;
 
-	/* files that aren't ln2 blocksz are only one block -- nothing to do */
-	if (!zf->zf_dnode->dn_datablkshift)
+	/*
+	 * As a fast path for small (single-block) files, ignore access
+	 * to the first block.
+	 */
+	if (blkid == 0)
 		return;
 
-	/* convert offset and size, into blockid and nblocks */
-	blkshft = zf->zf_dnode->dn_datablkshift;
-	blksz = (1 << blkshft);
-
-	bzero(&zst, sizeof (zstream_t));
-	zst.zst_offset = offset >> blkshft;
-	zst.zst_len = (P2ROUNDUP(offset + size, blksz) -
-	    P2ALIGN(offset, blksz)) >> blkshft;
+	rw_enter(&zf->zf_rwlock, RW_READER);
 
-	fetched = dmu_zfetch_find(zf, &zst, prefetched);
-	if (fetched) {
-		ZFETCHSTAT_BUMP(zfetchstat_hits);
-	} else {
-		ZFETCHSTAT_BUMP(zfetchstat_misses);
-		if ((fetched = dmu_zfetch_colinear(zf, &zst))) {
-			ZFETCHSTAT_BUMP(zfetchstat_colinear_hits);
-		} else {
-			ZFETCHSTAT_BUMP(zfetchstat_colinear_misses);
+	for (zs = list_head(&zf->zf_stream); zs != NULL;
+	    zs = list_next(&zf->zf_stream, zs)) {
+		if (blkid == zs->zs_blkid) {
+			mutex_enter(&zs->zs_lock);
+			/*
+			 * zs_blkid could have changed before we
+			 * acquired zs_lock; re-check them here.
+			 */
+			if (blkid != zs->zs_blkid) {
+				mutex_exit(&zs->zs_lock);
+				continue;
+			}
+			break;
 		}
 	}
 
-	if (!fetched) {
-		newstream = dmu_zfetch_stream_reclaim(zf);
-
+	if (zs == NULL) {
 		/*
-		 * we still couldn't find a stream, drop the lock, and allocate
-		 * one if possible.  Otherwise, give up and go home.
+		 * This access is not part of any existing stream.  Create
+		 * a new stream for it.
 		 */
-		if (newstream) {
-			ZFETCHSTAT_BUMP(zfetchstat_reclaim_successes);
-		} else {
-			uint64_t	maxblocks;
-			uint32_t	max_streams;
-			uint32_t	cur_streams;
-
-			ZFETCHSTAT_BUMP(zfetchstat_reclaim_failures);
-			cur_streams = zf->zf_stream_cnt;
-			maxblocks = zf->zf_dnode->dn_maxblkid;
-
-			max_streams = MIN(zfetch_max_streams,
-			    (maxblocks / zfetch_block_cap));
-			if (max_streams == 0) {
-				max_streams++;
-			}
-
-			if (cur_streams >= max_streams) {
-				return;
-			}
-			newstream =
-			    kmem_zalloc(sizeof (zstream_t), KM_SLEEP);
-		}
+		ZFETCHSTAT_BUMP(zfetchstat_misses);
+		if (rw_tryupgrade(&zf->zf_rwlock))
+			dmu_zfetch_stream_create(zf, blkid + nblks);
+		rw_exit(&zf->zf_rwlock);
+		return;
+	}
 
-		newstream->zst_offset = zst.zst_offset;
-		newstream->zst_len = zst.zst_len;
-		newstream->zst_stride = zst.zst_len;
-		newstream->zst_ph_offset = zst.zst_len + zst.zst_offset;
-		newstream->zst_cap = zst.zst_len;
-		newstream->zst_direction = ZFETCH_FORWARD;
-		newstream->zst_last = ddi_get_lbolt();
+	/*
+	 * This access was to a block that we issued a prefetch for on
+	 * behalf of this stream. Issue further prefetches for this stream.
+	 *
+	 * 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,
+	 * start just after the block we just accessed.
+	 */
+	pf_start = MAX(zs->zs_pf_blkid, blkid + nblks);
 
-		mutex_init(&newstream->zst_lock, NULL, MUTEX_DEFAULT, NULL);
+	/*
+	 * Double our amount of prefetched data, but don't let the
+	 * prefetch get further ahead than zfetch_max_distance.
+	 */
+	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);
 
-		rw_enter(&zf->zf_rwlock, RW_WRITER);
-		inserted = dmu_zfetch_stream_insert(zf, newstream);
-		rw_exit(&zf->zf_rwlock);
+	zs->zs_pf_blkid = pf_start + pf_nblks;
+	zs->zs_atime = gethrtime();
+	zs->zs_blkid = blkid + nblks;
 
-		if (!inserted) {
-			mutex_destroy(&newstream->zst_lock);
-			kmem_free(newstream, sizeof (zstream_t));
-		}
+	/*
+	 * 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.
+	 */
+	mutex_exit(&zs->zs_lock);
+	rw_exit(&zf->zf_rwlock);
+	for (i = 0; i < pf_nblks; i++) {
+		dbuf_prefetch(zf->zf_dnode, 0, pf_start + i,
+		    ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
 	}
+	ZFETCHSTAT_BUMP(zfetchstat_hits);
 }
 
 #if defined(_KERNEL) && defined(HAVE_SPL)
@@ -741,8 +297,9 @@ MODULE_PARM_DESC(zfetch_max_streams, "Max number of streams per zfetch");
 module_param(zfetch_min_sec_reap, uint, 0644);
 MODULE_PARM_DESC(zfetch_min_sec_reap, "Min time before stream reclaim");
 
-module_param(zfetch_block_cap, uint, 0644);
-MODULE_PARM_DESC(zfetch_block_cap, "Max number of blocks to fetch at a time");
+module_param(zfetch_max_distance, uint, 0644);
+MODULE_PARM_DESC(zfetch_max_distance,
+	"Max bytes to prefetch per stream (default 8MB)");
 
 module_param(zfetch_array_rd_sz, ulong, 0644);
 MODULE_PARM_DESC(zfetch_array_rd_sz, "Number of bytes in a array_read");
diff --git a/module/zfs/dnode.c b/module/zfs/dnode.c
index 190e66b74..38bcecd46 100644
--- a/module/zfs/dnode.c
+++ b/module/zfs/dnode.c
@@ -524,7 +524,7 @@ dnode_destroy(dnode_t *dn)
 	dn->dn_id_flags = 0;
 	dn->dn_unlisted_l0_blkid = 0;
 
-	dmu_zfetch_rele(&dn->dn_zfetch);
+	dmu_zfetch_fini(&dn->dn_zfetch);
 	kmem_cache_free(dnode_cache, dn);
 	arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
 
@@ -773,8 +773,6 @@ dnode_move_impl(dnode_t *odn, dnode_t *ndn)
 	dmu_zfetch_init(&ndn->dn_zfetch, NULL);
 	list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
 	ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
-	ndn->dn_zfetch.zf_stream_cnt = odn->dn_zfetch.zf_stream_cnt;
-	ndn->dn_zfetch.zf_alloc_fail = odn->dn_zfetch.zf_alloc_fail;
 
 	/*
 	 * Update back pointers. Updating the handle fixes the back pointer of