Fix use-after-free in case of L2ARC prefetch failure

In case L2ARC read failed, l2arc_read_done() creates _different_ ZIO
to read data from the original storage device.  Unfortunately pointer
to the failed ZIO remains in hdr->b_l1hdr.b_acb->acb_zio_head, and if
some other read try to bump the ZIO priority, it will crash.

The problem is reproducible by corrupting L2ARC content and reading
some data with prefetch if l2arc_noprefetch tunable is changed to 0.
With the default setting the issue is probably not reproducible now.

Reviewed-by: Tom Caputi <[email protected]>
Reviewed-by: Brian Behlendorf <[email protected]>
Signed-off-by: Alexander Motin <[email protected]>
Sponsored-By: iXsystems, Inc.
Closes openzfs#9648

module/zfs/arc.c

@@ -8561,7 +8561,6 @@ l2arc_read_done(zio_t *zio)
 		zio->io_private = hdr;
 		arc_read_done(zio);
 	} else {
-		mutex_exit(hash_lock);
 		/*
 		 * Buffer didn't survive caching.  Increment stats and
 		 * reissue to the original storage device.
@@ -8586,10 +8585,24 @@ l2arc_read_done(zio_t *zio)
 
 			ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
 
-			zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
+			zio = zio_read(pio, zio->io_spa, zio->io_bp,
 			    abd, zio->io_size, arc_read_done,
 			    hdr, zio->io_priority, cb->l2rcb_flags,
-			    &cb->l2rcb_zb));
+			    &cb->l2rcb_zb);
+
+			/*
+			 * Original ZIO will be freed, so we need to update
+			 * ARC header with the new ZIO pointer to be used
+			 * by zio_change_priority() in arc_read().
+			 */
+			for (struct arc_callback *acb = hdr->b_l1hdr.b_acb;
+			    acb != NULL; acb = acb->acb_next)
+				acb->acb_zio_head = zio;
+
+			mutex_exit(hash_lock);
+			zio_nowait(zio);
+		} else {
+			mutex_exit(hash_lock);
 		}
 	}