Wait for all znodes to be released before tearing down the superblock

By the time we're tearing down our superblock the VFS has started releasing
all our inodes/znodes. Some of this work may have been handed off to our
iput taskq so we need to wait for that work to complete. However the iput
from the taskq can itself result in additional work being added to the
taskq:

dsl_pool_iput_taskq
 iput
  iput_final
   evict
    destroy_inode
     zpl_inode_destroy
      zfs_inode_destroy
       zfs_iput_async(ZTOI(zp->z_xattr_parent))
        taskq_dispatch(dsl_pool_iput_taskq..., iput, ...)

Let's wait until all our znodes have been released.

Signed-off-by: Chris Dunlop <chris@onthe.net.au>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3281

diff --git a/module/zfs/zfs_vfsops.c b/module/zfs/zfs_vfsops.c
index 41a1c4d8849d281900a60652da98a147f384325c..59f73776cc1d37bb2aab9c28efbe707aa4c3f079 100644 (file)
--- a/module/zfs/zfs_vfsops.c
+++ b/module/zfs/zfs_vfsops.c
@@ -1137,8 +1137,28 @@ zfs_sb_teardown(zfs_sb_t *zsb, boolean_t unmounting)
         * drain the iput_taskq to ensure all active references to the
         * zfs_sb_t have been handled only then can it be safely destroyed.
         */
-       if (zsb->z_os)
-               taskq_wait(dsl_pool_iput_taskq(dmu_objset_pool(zsb->z_os)));
+       if (zsb->z_os) {
+               /*
+                * If we're unmounting we have to wait for the list to
+                * drain completely.
+                *
+                * If we're not unmounting there's no guarantee the list
+                * will drain completely, but iputs run from the taskq
+                * may add the parents of dir-based xattrs to the taskq
+                * so we want to wait for these.
+                *
+                * We can safely read z_nr_znodes without locking because the
+                * VFS has already blocked operations which add to the
+                * z_all_znodes list and thus increment z_nr_znodes.
+                */
+               int round = 0;
+               while (zsb->z_nr_znodes > 0) {
+                       taskq_wait(dsl_pool_iput_taskq(dmu_objset_pool(
+                           zsb->z_os)));
+                       if (++round > 1 && !unmounting)
+                               break;
+               }
+       }
 
        rrw_enter(&zsb->z_teardown_lock, RW_WRITER, FTAG);
 
@@ -1182,13 +1202,15 @@ zfs_sb_teardown(zfs_sb_t *zsb, boolean_t unmounting)
         *
         * Release all holds on dbufs.
         */
-       mutex_enter(&zsb->z_znodes_lock);
-       for (zp = list_head(&zsb->z_all_znodes); zp != NULL;
-           zp = list_next(&zsb->z_all_znodes, zp)) {
-               if (zp->z_sa_hdl)
-                       zfs_znode_dmu_fini(zp);
+       if (!unmounting) {
+               mutex_enter(&zsb->z_znodes_lock);
+               for (zp = list_head(&zsb->z_all_znodes); zp != NULL;
+               zp = list_next(&zsb->z_all_znodes, zp)) {
+                       if (zp->z_sa_hdl)
+                               zfs_znode_dmu_fini(zp);
+               }
+               mutex_exit(&zsb->z_znodes_lock);
        }
-       mutex_exit(&zsb->z_znodes_lock);
 
        /*
         * If we are unmounting, set the unmounted flag and let new VFS ops