Fix stack dbuf_hold_impl()

This commit preserves the recursive function dbuf_hold_impl() but moves
the local variables and function arguments to the heap to minimize
the stack frame size.  Enough space is initially allocated on the
stack for 20 levels of recursion.  This technique was based on commit
34229a2f2ac07363f64ddd63e014964fff2f0671 which reduced stack usage of
traverse_visitbp().

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

1 files changed, 140 insertions, 64 deletions

diff --git a/module/zfs/dbuf.c b/module/zfs/dbuf.c
index d08359156..ccab0bce4 100644
--- a/module/zfs/dbuf.c
+++ b/module/zfs/dbuf.c
@@ -36,6 +36,29 @@
 #include <sys/sa.h>
 #include <sys/sa_impl.h>
 
+struct dbuf_hold_impl_data {
+	/* Function arguments */
+	dnode_t *dh_dn;
+	uint8_t dh_level;
+	uint64_t dh_blkid;
+	int dh_fail_sparse;
+	void *dh_tag;
+	dmu_buf_impl_t **dh_dbp;
+	/* Local variables */
+	dmu_buf_impl_t *dh_db;
+	dmu_buf_impl_t *dh_parent;
+	blkptr_t *dh_bp;
+	int dh_err;
+	dbuf_dirty_record_t *dh_dr;
+	arc_buf_contents_t dh_type;
+	int dh_depth;
+};
+
+static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
+    dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
+    void *tag, dmu_buf_impl_t **dbp, int depth);
+static int __dbuf_hold_impl(struct dbuf_hold_impl_data *dh);
+
 static void dbuf_destroy(dmu_buf_impl_t *db);
 static int dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
@@ -1586,7 +1609,7 @@ dbuf_clear(dmu_buf_impl_t *db)
 
 static int
 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
-    dmu_buf_impl_t **parentp, blkptr_t **bpp)
+    dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh)
 {
 	int nlevels, epbs;
 
@@ -1623,8 +1646,17 @@ dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
 		return (ENOENT);
 	} else if (level < nlevels-1) {
 		/* this block is referenced from an indirect block */
-		int err = dbuf_hold_impl(dn, level+1,
-		    blkid >> epbs, fail_sparse, NULL, parentp);
+		int err;
+		if (dh == NULL) {
+			err = dbuf_hold_impl(dn, level+1, blkid >> epbs,
+					fail_sparse, NULL, parentp);
+		}
+		else {
+			__dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1,
+					blkid >> epbs, fail_sparse, NULL,
+					parentp, dh->dh_depth + 1);
+			err = __dbuf_hold_impl(dh + 1);
+		}
 		if (err)
 			return (err);
 		err = dbuf_read(*parentp, NULL,
@@ -1823,7 +1855,7 @@ dbuf_prefetch(dnode_t *dn, uint64_t blkid)
 		return;
 	}
 
-	if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
+	if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp, NULL) == 0) {
 		if (bp && !BP_IS_HOLE(bp)) {
 			int priority = dn->dn_type == DMU_OT_DDT_ZAP ?
 			    ZIO_PRIORITY_DDT_PREFETCH : ZIO_PRIORITY_ASYNC_READ;
@@ -1850,98 +1882,142 @@ dbuf_prefetch(dnode_t *dn, uint64_t blkid)
 	}
 }
 
+#define DBUF_HOLD_IMPL_MAX_DEPTH	20
+
 /*
  * Returns with db_holds incremented, and db_mtx not held.
  * Note: dn_struct_rwlock must be held.
  */
-int
-dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
-    void *tag, dmu_buf_impl_t **dbp)
+static int
+__dbuf_hold_impl(struct dbuf_hold_impl_data *dh)
 {
-	dmu_buf_impl_t *db, *parent = NULL;
+	ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH);
+	dh->dh_parent = NULL;
 
-	ASSERT(blkid != DMU_BONUS_BLKID);
-	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
-	ASSERT3U(dn->dn_nlevels, >, level);
+	ASSERT(dh->dh_blkid != DMU_BONUS_BLKID);
+	ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock));
+	ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level);
 
-	*dbp = NULL;
+	*(dh->dh_dbp) = NULL;
 top:
 	/* dbuf_find() returns with db_mtx held */
-	db = dbuf_find(dn, level, blkid);
-
-	if (db == NULL) {
-		blkptr_t *bp = NULL;
-		int err;
-
-		ASSERT3P(parent, ==, NULL);
-		err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
-		if (fail_sparse) {
-			if (err == 0 && bp && BP_IS_HOLE(bp))
-				err = ENOENT;
-			if (err) {
-				if (parent)
-					dbuf_rele(parent, NULL);
-				return (err);
+	dh->dh_db = dbuf_find(dh->dh_dn, dh->dh_level, dh->dh_blkid);
+
+	if (dh->dh_db == NULL) {
+		dh->dh_bp = NULL;
+
+		ASSERT3P(dh->dh_parent, ==, NULL);
+		dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid,
+					dh->dh_fail_sparse, &dh->dh_parent,
+					&dh->dh_bp, dh);
+		if (dh->dh_fail_sparse) {
+			if (dh->dh_err == 0 && dh->dh_bp && BP_IS_HOLE(dh->dh_bp))
+				dh->dh_err = ENOENT;
+			if (dh->dh_err) {
+				if (dh->dh_parent)
+					dbuf_rele(dh->dh_parent, NULL);
+				return (dh->dh_err);
 			}
 		}
-		if (err && err != ENOENT)
-			return (err);
-		db = dbuf_create(dn, level, blkid, parent, bp);
-	}
-
-	if (db->db_buf && refcount_is_zero(&db->db_holds)) {
-		arc_buf_add_ref(db->db_buf, db);
-		if (db->db_buf->b_data == NULL) {
-			dbuf_clear(db);
-			if (parent) {
-				dbuf_rele(parent, NULL);
-				parent = NULL;
+		if (dh->dh_err && dh->dh_err != ENOENT)
+			return (dh->dh_err);
+		dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid,
+					dh->dh_parent, dh->dh_bp);
+	}
+
+	if (dh->dh_db->db_buf && refcount_is_zero(&dh->dh_db->db_holds)) {
+		arc_buf_add_ref(dh->dh_db->db_buf, dh->dh_db);
+		if (dh->dh_db->db_buf->b_data == NULL) {
+			dbuf_clear(dh->dh_db);
+			if (dh->dh_parent) {
+				dbuf_rele(dh->dh_parent, NULL);
+				dh->dh_parent = NULL;
 			}
 			goto top;
 		}
-		ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
+		ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data);
 	}
 
-	ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
+	ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf));
 
 	/*
 	 * If this buffer is currently syncing out, and we are are
 	 * still referencing it from db_data, we need to make a copy
 	 * of it in case we decide we want to dirty it again in this txg.
 	 */
-	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
-	    dn->dn_object != DMU_META_DNODE_OBJECT &&
-	    db->db_state == DB_CACHED && db->db_data_pending) {
-		dbuf_dirty_record_t *dr = db->db_data_pending;
-
-		if (dr->dt.dl.dr_data == db->db_buf) {
-			arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
-
-			dbuf_set_data(db,
-			    arc_buf_alloc(dn->dn_objset->os_spa,
-			    db->db.db_size, db, type));
-			bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
-			    db->db.db_size);
+	if (dh->dh_db->db_level == 0 &&
+	    dh->dh_db->db_blkid != DMU_BONUS_BLKID &&
+	    dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT &&
+	    dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) {
+		dh->dh_dr = dh->dh_db->db_data_pending;
+
+		if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf) {
+			dh->dh_type = DBUF_GET_BUFC_TYPE(dh->dh_db);
+
+			dbuf_set_data(dh->dh_db,
+			    arc_buf_alloc(dh->dh_dn->dn_objset->os_spa,
+			    dh->dh_db->db.db_size, dh->dh_db, dh->dh_type));
+			bcopy(dh->dh_dr->dt.dl.dr_data->b_data,
+			    dh->dh_db->db.db_data, dh->dh_db->db.db_size);
 		}
 	}
 
-	(void) refcount_add(&db->db_holds, tag);
-	dbuf_update_data(db);
-	DBUF_VERIFY(db);
-	mutex_exit(&db->db_mtx);
+	(void) refcount_add(&dh->dh_db->db_holds, dh->dh_tag);
+	dbuf_update_data(dh->dh_db);
+	DBUF_VERIFY(dh->dh_db);
+	mutex_exit(&dh->dh_db->db_mtx);
 
 	/* NOTE: we can't rele the parent until after we drop the db_mtx */
-	if (parent)
-		dbuf_rele(parent, NULL);
+	if (dh->dh_parent)
+		dbuf_rele(dh->dh_parent, NULL);
 
-	ASSERT3P(DB_DNODE(db), ==, dn);
-	ASSERT3U(db->db_blkid, ==, blkid);
-	ASSERT3U(db->db_level, ==, level);
-	*dbp = db;
+	ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn);
+	ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid);
+	ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level);
+	*(dh->dh_dbp) = dh->dh_db;
 
 	return (0);
 }
 
+/*
+ * The following code preserves the recursive function dbuf_hold_impl()
+ * but moves the local variables AND function arguments to the heap to
+ * minimize the stack frame size.  Enough space is initially allocated
+ * on the stack for 20 levels of recursion.
+ */
+int
+dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
+    void *tag, dmu_buf_impl_t **dbp)
+{
+	struct dbuf_hold_impl_data *dh;
+	int error;
+
+	dh = kmem_zalloc(sizeof(struct dbuf_hold_impl_data) *
+	    DBUF_HOLD_IMPL_MAX_DEPTH, KM_SLEEP);
+	__dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse, tag, dbp, 0);
+
+	error = __dbuf_hold_impl(dh);
+
+	kmem_free(dh, sizeof(struct dbuf_hold_impl_data) *
+	    DBUF_HOLD_IMPL_MAX_DEPTH);
+
+	return (error);
+}
+
+static void
+__dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
+    dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
+    void *tag, dmu_buf_impl_t **dbp, int depth)
+{
+	dh->dh_dn = dn;
+	dh->dh_level = level;
+	dh->dh_blkid = blkid;
+	dh->dh_fail_sparse = fail_sparse;
+	dh->dh_tag = tag;
+	dh->dh_dbp = dbp;
+	dh->dh_depth = depth;
+}
+
 dmu_buf_impl_t *
 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
 {