Illumos 5095 - panic when adding a duplicate dbuf to dn_dbufs

5095 panic when adding a duplicate dbuf to dn_dbufs
Author: Alex Reece <alex@delphix.com>
Reviewed by: Adam Leventhal <adam.leventhal@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Mattew Ahrens <mahrens@delphix.com>
Reviewed by: Dan Kimmel <dan.kimmel@delphix.com>
Reviewed by: Dan McDonald <danmcd@omniti.com>
Reviewed by: Josef Sipek <jeffpc@josefsipek.net>
Approved by: Robert Mustacchi <rm@joyent.com>

References:
  https://www.illumos.org/issues/5095
  https://github.com/illumos/illumos-gate/commit/86bb58a

Ported-by: Chris Dunlop <chris@onthe.net.au>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>

include/sys/dbuf.h

@@ -66,8 +66,13 @@ extern "C" {
  *		|			 |
  *		|			 |
  *		+--------> NOFILL -------+
+ *
+ * DB_SEARCH is an invalid state for a dbuf. It is used by dbuf_free_range
+ * to find all dbufs in a range of a dnode and must be less than any other
+ * dbuf_states_t (see comment on dn_dbufs in dnode.h).
  */
 typedef enum dbuf_states {
+	DB_SEARCH = -1,
 	DB_UNCACHED,
 	DB_FILL,
 	DB_NOFILL,
@@ -213,9 +218,6 @@ typedef struct dmu_buf_impl {
 	/* pointer to most recent dirty record for this buffer */
 	dbuf_dirty_record_t *db_last_dirty;
 
-	/* Creation time of dbuf (see comment in dbuf_compare). */
-	hrtime_t db_creation;
-
 	/*
 	 * Our link on the owner dnodes's dn_dbufs list.
 	 * Protected by its dn_dbufs_mtx.

include/sys/dnode.h

@@ -233,7 +233,18 @@ typedef struct dnode {
 	refcount_t dn_holds;
 
 	kmutex_t dn_dbufs_mtx;
-	avl_tree_t dn_dbufs;		/* descendent dbufs */
+	/*
+	 * Descendent dbufs, ordered by dbuf_compare. Note that dn_dbufs
+	 * can contain multiple dbufs of the same (level, blkid) when a
+	 * dbuf is marked DB_EVICTING without being removed from
+	 * dn_dbufs. To maintain the avl invariant that there cannot be
+	 * duplicate entries, we order the dbufs by an arbitrary value -
+	 * their address in memory. This means that dn_dbufs cannot be used to
+	 * directly look up a dbuf. Instead, callers must use avl_walk, have
+	 * a reference to the dbuf, or look up a non-existant node with
+	 * db_state = DB_SEARCH (see dbuf_free_range for an example).
+	 */
+	avl_tree_t dn_dbufs;
 
 	/* protected by dn_struct_rwlock */
 	struct dmu_buf_impl *dn_bonus;	/* bonus buffer dbuf */

module/zfs/dbuf.c

@@ -94,8 +94,6 @@ dbuf_cons(void *vdb, void *unused, int kmflag)
 	cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
 	refcount_create(&db->db_holds);
 
-	db->db_creation = gethrtime();
-
 	return (0);
 }
 
@@ -884,7 +882,7 @@ dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
 	db_seach = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
 	db_search->db_level = 0;
 	db_search->db_blkid = start_blkid;
-	db_search->db_creation = 0;
+	db_search->db_state = DB_SEARCH;
 
 	mutex_enter(&dn->dn_dbufs_mtx);
 	if (start_blkid >= dn->dn_unlisted_l0_blkid && !freespill) {

module/zfs/dnode.c

@@ -70,33 +70,35 @@ dbuf_compare(const void *x1, const void *x2)
 
 	if (d1->db_level < d2->db_level) {
 		return (-1);
-	} else if (d1->db_level > d2->db_level) {
+	}
+	if (d1->db_level > d2->db_level) {
 		return (1);
 	}
 
 	if (d1->db_blkid < d2->db_blkid) {
 		return (-1);
-	} else if (d1->db_blkid > d2->db_blkid) {
+	}
+	if (d1->db_blkid > d2->db_blkid) {
 		return (1);
 	}
 
-	/*
-	 * If a dbuf is being evicted while dn_dbufs_mutex is not held, we set
-	 * the db_state to DB_EVICTING but do not remove it from dn_dbufs. If
-	 * another thread creates a dbuf of the same blkid before the dbuf is
-	 * removed from dn_dbufs, we can reach a state where there are two
-	 * dbufs of the same blkid and level in db_dbufs. To maintain the avl
-	 * invariant that there cannot be duplicate items, we distinguish
-	 * between these two dbufs based on the time they were created.
-	 */
-	if (d1->db_creation < d2->db_creation) {
+	if (d1->db_state < d2->db_state) {
 		return (-1);
-	} else if (d1->db_creation > d2->db_creation) {
-		return (1);
-	} else {
-		ASSERT3P(d1, ==, d2);
-		return (0);
 	}
+	if (d1->db_state > d2->db_state) {
+		return (1);
+	}
+
+	ASSERT3S(d1->db_state, !=, DB_SEARCH);
+	ASSERT3S(d2->db_state, !=, DB_SEARCH);
+
+	if ((uintptr_t)d1 < (uintptr_t)d2) {
+		return (-1);
+	}
+	if ((uintptr_t)d1 > (uintptr_t)d2) {
+		return (1);
+	}
+	return (0);
 }
 
 /* ARGSUSED */