Add prefetch property

ZFS prefetch is currently governed by the zfs_prefetch_disable
tunable. However, this is a module-wide settings - if a specific
dataset benefits from prefetch, while others have issue with it,
an optimal solution does not exists.

This commit introduce the "prefetch" tri-state property, which enable
granular control (at dataset/volume level) for prefetching.

This patch does not remove the zfs_prefetch_disable, which remains
a system-wide switch for enable/disable prefetch. However, to avoid
duplication, it would be preferable to deprecate and then remove
the module tunable.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Ameer Hamza <ahamza@ixsystems.com>
Signed-off-by: Gionatan Danti <g.danti@assyoma.it>
Co-authored-by: Gionatan Danti <g.danti@assyoma.it>
Closes #15237 
Closes #15436

include/sys/dmu_objset.h

@@ -132,6 +132,7 @@ struct objset {
 	zfs_logbias_op_t os_logbias;
 	zfs_cache_type_t os_primary_cache;
 	zfs_cache_type_t os_secondary_cache;
+	zfs_prefetch_type_t os_prefetch;
 	zfs_sync_type_t os_sync;
 	zfs_redundant_metadata_type_t os_redundant_metadata;
 	uint64_t os_recordsize;

include/sys/fs/zfs.h

@@ -191,6 +191,7 @@ typedef enum {
 	ZFS_PROP_REDACTED,
 	ZFS_PROP_REDACT_SNAPS,
 	ZFS_PROP_SNAPSHOTS_CHANGED,
+	ZFS_PROP_PREFETCH,
 	ZFS_NUM_PROPS
 } zfs_prop_t;
 
@@ -546,6 +547,12 @@ typedef enum zfs_key_location {
 	ZFS_KEYLOCATION_LOCATIONS
 } zfs_keylocation_t;
 
+typedef enum {
+	ZFS_PREFETCH_NONE = 0,
+	ZFS_PREFETCH_METADATA = 1,
+	ZFS_PREFETCH_ALL = 2
+} zfs_prefetch_type_t;
+
 #define	DEFAULT_PBKDF2_ITERATIONS 350000
 #define	MIN_PBKDF2_ITERATIONS 100000
 

lib/libzfs/libzfs.abi

@@ -1867,7 +1867,8 @@
       <enumerator name='ZFS_PROP_REDACTED' value='93'/>
       <enumerator name='ZFS_PROP_REDACT_SNAPS' value='94'/>
       <enumerator name='ZFS_PROP_SNAPSHOTS_CHANGED' value='95'/>
-      <enumerator name='ZFS_NUM_PROPS' value='96'/>
+      <enumerator name='ZFS_PROP_PREFETCH' value='96'/>
+      <enumerator name='ZFS_NUM_PROPS' value='97'/>
     </enum-decl>
     <typedef-decl name='zfs_prop_t' type-id='4b000d60' id='58603c44'/>
     <enum-decl name='zprop_source_t' naming-typedef-id='a2256d42' id='5903f80e'>

man/man7/zfsprops.7

@@ -1613,6 +1613,23 @@ If this property is set to
 then only metadata is cached.
 The default value is
 .Sy all .
+.It Sy prefetch Ns = Ns Sy all Ns | Ns Sy none Ns | Ns Sy metadata
+Controls what speculative prefetch does.
+If this property is set to
+.Sy all ,
+then both user data and metadata are prefetched.
+If this property is set to
+.Sy none ,
+then neither user data nor metadata are prefetched.
+If this property is set to
+.Sy metadata ,
+then only metadata are prefetched.
+The default value is
+.Sy all .
+.Pp
+Please note that the module parameter zfs_disable_prefetch=1 can
+be used to totally disable speculative prefetch, bypassing anything
+this property does.
 .It Sy setuid Ns = Ns Sy on Ns | Ns Sy off
 Controls whether the setuid bit is respected for the file system.
 The default value is

module/zcommon/zfs_prop.c

@@ -345,6 +345,13 @@ zfs_prop_init(void)
 		{ NULL }
 	};
 
+	static const zprop_index_t prefetch_table[] = {
+		{ "none",	ZFS_PREFETCH_NONE },
+		{ "metadata",	ZFS_PREFETCH_METADATA },
+		{ "all",	ZFS_PREFETCH_ALL },
+		{ NULL }
+	};
+
 	static const zprop_index_t sync_table[] = {
 		{ "standard",	ZFS_SYNC_STANDARD },
 		{ "always",	ZFS_SYNC_ALWAYS },
@@ -453,6 +460,10 @@ zfs_prop_init(void)
 	    ZFS_CACHE_ALL, PROP_INHERIT,
 	    ZFS_TYPE_FILESYSTEM | ZFS_TYPE_SNAPSHOT | ZFS_TYPE_VOLUME,
 	    "all | none | metadata", "SECONDARYCACHE", cache_table, sfeatures);
+	zprop_register_index(ZFS_PROP_PREFETCH, "prefetch",
+	    ZFS_PREFETCH_ALL, PROP_INHERIT,
+	    ZFS_TYPE_FILESYSTEM | ZFS_TYPE_SNAPSHOT | ZFS_TYPE_VOLUME,
+	    "none | metadata | all", "PREFETCH", prefetch_table, sfeatures);
 	zprop_register_index(ZFS_PROP_LOGBIAS, "logbias", ZFS_LOGBIAS_LATENCY,
 	    PROP_INHERIT, ZFS_TYPE_FILESYSTEM | ZFS_TYPE_VOLUME,
 	    "latency | throughput", "LOGBIAS", logbias_table, sfeatures);

module/zfs/dmu_objset.c

@@ -263,6 +263,19 @@ secondary_cache_changed_cb(void *arg, uint64_t newval)
 	os->os_secondary_cache = newval;
 }
 
+static void
+prefetch_changed_cb(void *arg, uint64_t newval)
+{
+	objset_t *os = arg;
+
+	/*
+	 * Inheritance should have been done by now.
+	 */
+	ASSERT(newval == ZFS_PREFETCH_ALL || newval == ZFS_PREFETCH_NONE ||
+	    newval == ZFS_PREFETCH_METADATA);
+	os->os_prefetch = newval;
+}
+
 static void
 sync_changed_cb(void *arg, uint64_t newval)
 {
@@ -562,6 +575,11 @@ dmu_objset_open_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *bp,
 			    zfs_prop_to_name(ZFS_PROP_SECONDARYCACHE),
 			    secondary_cache_changed_cb, os);
 		}
+		if (err == 0) {
+			err = dsl_prop_register(ds,
+			    zfs_prop_to_name(ZFS_PROP_PREFETCH),
+			    prefetch_changed_cb, os);
+		}
 		if (!ds->ds_is_snapshot) {
 			if (err == 0) {
 				err = dsl_prop_register(ds,
@@ -635,6 +653,7 @@ dmu_objset_open_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *bp,
 		os->os_primary_cache = ZFS_CACHE_ALL;
 		os->os_secondary_cache = ZFS_CACHE_ALL;
 		os->os_dnodesize = DNODE_MIN_SIZE;
+		os->os_prefetch = ZFS_PREFETCH_ALL;
 	}
 
 	if (ds == NULL || !ds->ds_is_snapshot)

module/zfs/dmu_zfetch.c

@@ -471,9 +471,14 @@ dmu_zfetch_prepare(zfetch_t *zf, uint64_t blkid, uint64_t nblks,
 {
 	zstream_t *zs;
 	spa_t *spa = zf->zf_dnode->dn_objset->os_spa;
+	zfs_prefetch_type_t os_prefetch = zf->zf_dnode->dn_objset->os_prefetch;
 
-	if (zfs_prefetch_disable)
+	if (zfs_prefetch_disable || os_prefetch == ZFS_PREFETCH_NONE)
 		return (NULL);
+
+	if (os_prefetch == ZFS_PREFETCH_METADATA)
+		fetch_data = B_FALSE;
+
 	/*
 	 * If we haven't yet loaded the indirect vdevs' mappings, we
 	 * can only read from blocks that we carefully ensure are on