Prefetch on deadlists merge

During snapshot deletion ZFS may issue several reads for each deadlist
to merge them into next snapshot's or pool's bpobj.  Number of the dead
lists increases with number of snapshots.  On HDD pools it may take
significant time during which sync thread is blocked.

This patch introduces prescient prefetch of required blocks for up to
128 deadlists ahead.  Tests show reduction of time required to delete
dataset with 720 snapshots with randomly overwritten file on wide HDD
pool from 75-85 to 22-28 seconds.

Reviewed-by: Allan Jude <[email protected]>
Reviewed-by: Brian Behlendorf <[email protected]>
Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
Issue openzfs#14276 
Closes openzfs#14402

include/sys/bpobj.h

@@ -87,6 +87,7 @@ int livelist_bpobj_iterate_from_nofree(bpobj_t *bpo, bpobj_itor_t func,
     void *arg, int64_t start);
 
 void bpobj_enqueue_subobj(bpobj_t *bpo, uint64_t subobj, dmu_tx_t *tx);
+void bpobj_prefetch_subobj(bpobj_t *bpo, uint64_t subobj);
 void bpobj_enqueue(bpobj_t *bpo, const blkptr_t *bp, boolean_t bp_freed,
     dmu_tx_t *tx);
 

module/zfs/bpobj.c

@@ -663,14 +663,13 @@ bpobj_enqueue_subobj(bpobj_t *bpo, uint64_t subobj, dmu_tx_t *tx)
 	}
 
 	VERIFY3U(0, ==, bpobj_open(&subbpo, bpo->bpo_os, subobj));
-	VERIFY3U(0, ==, bpobj_space(&subbpo, &used, &comp, &uncomp));
-
 	if (bpobj_is_empty(&subbpo)) {
 		/* No point in having an empty subobj. */
 		bpobj_close(&subbpo);
 		bpobj_free(bpo->bpo_os, subobj, tx);
 		return;
 	}
+	VERIFY3U(0, ==, bpobj_space(&subbpo, &used, &comp, &uncomp));
 
 	mutex_enter(&bpo->bpo_lock);
 	dmu_buf_will_dirty(bpo->bpo_dbuf, tx);
@@ -780,6 +779,68 @@ bpobj_enqueue_subobj(bpobj_t *bpo, uint64_t subobj, dmu_tx_t *tx)
 
 }
 
+/*
+ * Prefetch metadata required for bpobj_enqueue_subobj().
+ */
+void
+bpobj_prefetch_subobj(bpobj_t *bpo, uint64_t subobj)
+{
+	dmu_object_info_t doi;
+	bpobj_t subbpo;
+	uint64_t subsubobjs;
+	boolean_t copy_subsub = B_TRUE;
+	boolean_t copy_bps = B_TRUE;
+
+	ASSERT(bpobj_is_open(bpo));
+	ASSERT(subobj != 0);
+
+	if (subobj == dmu_objset_pool(bpo->bpo_os)->dp_empty_bpobj)
+		return;
+
+	if (bpobj_open(&subbpo, bpo->bpo_os, subobj) != 0)
+		return;
+	if (bpobj_is_empty(&subbpo)) {
+		bpobj_close(&subbpo);
+		return;
+	}
+	subsubobjs = subbpo.bpo_phys->bpo_subobjs;
+	bpobj_close(&subbpo);
+
+	if (subsubobjs != 0) {
+		if (dmu_object_info(bpo->bpo_os, subsubobjs, &doi) != 0)
+			return;
+		if (doi.doi_max_offset > doi.doi_data_block_size)
+			copy_subsub = B_FALSE;
+	}
+
+	if (dmu_object_info(bpo->bpo_os, subobj, &doi) != 0)
+		return;
+	if (doi.doi_max_offset > doi.doi_data_block_size || !copy_subsub)
+		copy_bps = B_FALSE;
+
+	if (copy_subsub && subsubobjs != 0) {
+		if (bpo->bpo_phys->bpo_subobjs) {
+			dmu_prefetch(bpo->bpo_os, bpo->bpo_phys->bpo_subobjs, 0,
+			    bpo->bpo_phys->bpo_num_subobjs * sizeof (subobj), 1,
+			    ZIO_PRIORITY_ASYNC_READ);
+		}
+		dmu_prefetch(bpo->bpo_os, subsubobjs, 0, 0, 1,
+		    ZIO_PRIORITY_ASYNC_READ);
+	}
+
+	if (copy_bps) {
+		dmu_prefetch(bpo->bpo_os, bpo->bpo_object, 0,
+		    bpo->bpo_phys->bpo_num_blkptrs * sizeof (blkptr_t), 1,
+		    ZIO_PRIORITY_ASYNC_READ);
+		dmu_prefetch(bpo->bpo_os, subobj, 0, 0, 1,
+		    ZIO_PRIORITY_ASYNC_READ);
+	} else if (bpo->bpo_phys->bpo_subobjs) {
+		dmu_prefetch(bpo->bpo_os, bpo->bpo_phys->bpo_subobjs, 0,
+		    bpo->bpo_phys->bpo_num_subobjs * sizeof (subobj), 1,
+		    ZIO_PRIORITY_ASYNC_READ);
+	}
+}
+
 void
 bpobj_enqueue(bpobj_t *bpo, const blkptr_t *bp, boolean_t bp_freed,
     dmu_tx_t *tx)

module/zfs/dsl_deadlist.c

@@ -438,6 +438,18 @@ dle_enqueue_subobj(dsl_deadlist_t *dl, dsl_deadlist_entry_t *dle,
 	}
 }
 
+/*
+ * Prefetch metadata required for dle_enqueue_subobj().
+ */
+static void
+dle_prefetch_subobj(dsl_deadlist_t *dl, dsl_deadlist_entry_t *dle,
+    uint64_t obj)
+{
+	if (dle->dle_bpobj.bpo_object !=
+	    dmu_objset_pool(dl->dl_os)->dp_empty_bpobj)
+		bpobj_prefetch_subobj(&dle->dle_bpobj, obj);
+}
+
 void
 dsl_deadlist_insert(dsl_deadlist_t *dl, const blkptr_t *bp, boolean_t bp_freed,
     dmu_tx_t *tx)
@@ -810,6 +822,27 @@ dsl_deadlist_insert_bpobj(dsl_deadlist_t *dl, uint64_t obj, uint64_t birth,
 	dle_enqueue_subobj(dl, dle, obj, tx);
 }
 
+/*
+ * Prefetch metadata required for dsl_deadlist_insert_bpobj().
+ */
+static void
+dsl_deadlist_prefetch_bpobj(dsl_deadlist_t *dl, uint64_t obj, uint64_t birth)
+{
+	dsl_deadlist_entry_t dle_tofind;
+	dsl_deadlist_entry_t *dle;
+	avl_index_t where;
+
+	ASSERT(MUTEX_HELD(&dl->dl_lock));
+
+	dsl_deadlist_load_tree(dl);
+
+	dle_tofind.dle_mintxg = birth;
+	dle = avl_find(&dl->dl_tree, &dle_tofind, &where);
+	if (dle == NULL)
+		dle = avl_nearest(&dl->dl_tree, where, AVL_BEFORE);
+	dle_prefetch_subobj(dl, dle, obj);
+}
+
 static int
 dsl_deadlist_insert_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
     dmu_tx_t *tx)
@@ -826,12 +859,12 @@ dsl_deadlist_insert_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
 void
 dsl_deadlist_merge(dsl_deadlist_t *dl, uint64_t obj, dmu_tx_t *tx)
 {
-	zap_cursor_t zc;
-	zap_attribute_t za;
+	zap_cursor_t zc, pzc;
+	zap_attribute_t za, pza;
 	dmu_buf_t *bonus;
 	dsl_deadlist_phys_t *dlp;
 	dmu_object_info_t doi;
-	int error;
+	int error, perror, i;
 
 	VERIFY0(dmu_object_info(dl->dl_os, obj, &doi));
 	if (doi.doi_type == DMU_OT_BPOBJ) {
@@ -843,15 +876,32 @@ dsl_deadlist_merge(dsl_deadlist_t *dl, uint64_t obj, dmu_tx_t *tx)
 	}
 
 	mutex_enter(&dl->dl_lock);
+	/*
+	 * Prefetch up to 128 deadlists first and then more as we progress.
+	 * The limit is a balance between ARC use and diminishing returns.
+	 */
+	for (zap_cursor_init(&pzc, dl->dl_os, obj), i = 0;
+	    (perror = zap_cursor_retrieve(&pzc, &pza)) == 0 && i < 128;
+	    zap_cursor_advance(&pzc), i++) {
+		dsl_deadlist_prefetch_bpobj(dl, pza.za_first_integer,
+		    zfs_strtonum(pza.za_name, NULL));
+	}
 	for (zap_cursor_init(&zc, dl->dl_os, obj);
 	    (error = zap_cursor_retrieve(&zc, &za)) == 0;
 	    zap_cursor_advance(&zc)) {
 		uint64_t mintxg = zfs_strtonum(za.za_name, NULL);
 		dsl_deadlist_insert_bpobj(dl, za.za_first_integer, mintxg, tx);
 		VERIFY0(zap_remove_int(dl->dl_os, obj, mintxg, tx));
+		if (perror == 0) {
+			dsl_deadlist_prefetch_bpobj(dl, pza.za_first_integer,
+			    zfs_strtonum(pza.za_name, NULL));
+			zap_cursor_advance(&pzc);
+			perror = zap_cursor_retrieve(&pzc, &pza);
+		}
 	}
 	VERIFY3U(error, ==, ENOENT);
 	zap_cursor_fini(&zc);
+	zap_cursor_fini(&pzc);
 
 	VERIFY0(dmu_bonus_hold(dl->dl_os, obj, FTAG, &bonus));
 	dlp = bonus->db_data;
@@ -869,8 +919,9 @@ dsl_deadlist_move_bpobj(dsl_deadlist_t *dl, bpobj_t *bpo, uint64_t mintxg,
     dmu_tx_t *tx)
 {
 	dsl_deadlist_entry_t dle_tofind;
-	dsl_deadlist_entry_t *dle;
+	dsl_deadlist_entry_t *dle, *pdle;
 	avl_index_t where;
+	int i;
 
 	ASSERT(!dl->dl_oldfmt);
 
@@ -882,11 +933,23 @@ dsl_deadlist_move_bpobj(dsl_deadlist_t *dl, bpobj_t *bpo, uint64_t mintxg,
 	dle = avl_find(&dl->dl_tree, &dle_tofind, &where);
 	if (dle == NULL)
 		dle = avl_nearest(&dl->dl_tree, where, AVL_AFTER);
+	/*
+	 * Prefetch up to 128 deadlists first and then more as we progress.
+	 * The limit is a balance between ARC use and diminishing returns.
+	 */
+	for (pdle = dle, i = 0; pdle && i < 128; ) {
+		bpobj_prefetch_subobj(bpo, pdle->dle_bpobj.bpo_object);
+		pdle = AVL_NEXT(&dl->dl_tree, pdle);
+	}
 	while (dle) {
 		uint64_t used, comp, uncomp;
 		dsl_deadlist_entry_t *dle_next;
 
 		bpobj_enqueue_subobj(bpo, dle->dle_bpobj.bpo_object, tx);
+		if (pdle) {
+			bpobj_prefetch_subobj(bpo, pdle->dle_bpobj.bpo_object);
+			pdle = AVL_NEXT(&dl->dl_tree, pdle);
+		}
 
 		VERIFY0(bpobj_space(&dle->dle_bpobj,
 		    &used, &comp, &uncomp));