Fix zfsctl_expire_snapshot() deadlock

[zfs.git] / module / zfs / zfs_ctldir.c

diff --git a/module/zfs/zfs_ctldir.c b/module/zfs/zfs_ctldir.c
index d139b86..168f853 100644 (file)
--- a/module/zfs/zfs_ctldir.c
+++ b/module/zfs/zfs_ctldir.c
@@ -90,6 +90,11 @@
  */
 int zfs_expire_snapshot = ZFSCTL_EXPIRE_SNAPSHOT;
 
+/*
+ * Dedicated task queue for unmounting snapshots.
+ */
+static taskq_t *zfs_expire_taskq;
+
 static zfs_snapentry_t *
 zfsctl_sep_alloc(void)
 {
@@ -112,16 +117,15 @@ zfsctl_sep_free(zfs_snapentry_t *sep)
 static void
 zfsctl_expire_snapshot(void *data)
 {
-       zfs_snapentry_t *sep;
-       zfs_sb_t *zsb;
+       zfs_snapentry_t *sep = (zfs_snapentry_t *)data;
+       zfs_sb_t *zsb = ITOZSB(sep->se_inode);
        int error;
 
-       sep = spl_get_work_data(data, zfs_snapentry_t, se_work.work);
-       zsb = ITOZSB(sep->se_inode);
-
        error = zfsctl_unmount_snapshot(zsb, sep->se_name, MNT_EXPIRE);
        if (error == EBUSY)
-               schedule_delayed_work(&sep->se_work, zfs_expire_snapshot * HZ);
+               sep->se_taskqid = taskq_dispatch_delay(zfs_expire_taskq,
+                   zfsctl_expire_snapshot, sep, TQ_SLEEP,
+                   ddi_get_lbolt() + zfs_expire_snapshot * HZ);
 }
 
 int
@@ -193,6 +197,7 @@ zfsctl_inode_alloc(zfs_sb_t *zsb, uint64_t id,
        zp->z_is_mapped = B_FALSE;
        zp->z_is_ctldir = B_TRUE;
        zp->z_is_sa = B_FALSE;
+       zp->z_is_stale = B_FALSE;
        ip->i_ino = id;
        ip->i_mode = (S_IFDIR | S_IRUGO | S_IXUGO);
        ip->i_uid = 0;
@@ -438,7 +443,7 @@ zfsctl_snapdir_lookup(struct inode *dip, char *name, struct inode **ipp,
 
        ZFS_ENTER(zsb);
 
-       error = dmu_snapshot_id(zsb->z_os, name, &id);
+       error = dmu_snapshot_lookup(zsb->z_os, name, &id);
        if (error) {
                ZFS_EXIT(zsb);
                return (error);
@@ -661,7 +666,7 @@ zfsctl_snapdir_inactive(struct inode *ip)
 
                if (sep->se_inode == ip) {
                        avl_remove(&zsb->z_ctldir_snaps, sep);
-                       cancel_delayed_work_sync(&sep->se_work);
+                       taskq_cancel_id(zfs_expire_taskq, sep->se_taskqid);
                        zfsctl_sep_free(sep);
                        break;
                }
@@ -681,7 +686,7 @@ zfsctl_snapdir_inactive(struct inode *ip)
        "exec 0</dev/null " \
        "     1>/dev/null " \
        "     2>/dev/null; " \
-       "umount -t zfs -n '%s%s'"
+       "umount -t zfs -n %s'%s'"
 
 static int
 __zfsctl_unmount_snapshot(zfs_snapentry_t *sep, int flags)
@@ -692,7 +697,7 @@ __zfsctl_unmount_snapshot(zfs_snapentry_t *sep, int flags)
 
        argv[2] = kmem_asprintf(SET_UNMOUNT_CMD,
            flags & MNT_FORCE ? "-f " : "", sep->se_path);
-       error = call_usermodehelper(argv[0], argv, envp, 1);
+       error = call_usermodehelper(argv[0], argv, envp, UMH_WAIT_PROC);
        strfree(argv[2]);
 
        /*
@@ -708,7 +713,8 @@ __zfsctl_unmount_snapshot(zfs_snapentry_t *sep, int flags)
         * to prevent zfsctl_expire_snapshot() from attempting a unmount.
         */
        if ((error == 0) && !(flags & MNT_EXPIRE))
-               cancel_delayed_work(&sep->se_work);
+               taskq_cancel_id(zfs_expire_taskq, sep->se_taskqid);
+
 
        return (error);
 }
@@ -726,7 +732,11 @@ zfsctl_unmount_snapshot(zfs_sb_t *zsb, char *name, int flags)
        sep = avl_find(&zsb->z_ctldir_snaps, &search, NULL);
        if (sep) {
                avl_remove(&zsb->z_ctldir_snaps, sep);
+               mutex_exit(&zsb->z_ctldir_lock);
+
                error = __zfsctl_unmount_snapshot(sep, flags);
+
+               mutex_enter(&zsb->z_ctldir_lock);
                if (error == EBUSY)
                        avl_add(&zsb->z_ctldir_snaps, sep);
                else
@@ -761,7 +771,11 @@ zfsctl_unmount_snapshots(zfs_sb_t *zsb, int flags, int *count)
        while (sep != NULL) {
                next = AVL_NEXT(&zsb->z_ctldir_snaps, sep);
                avl_remove(&zsb->z_ctldir_snaps, sep);
+               mutex_exit(&zsb->z_ctldir_lock);
+
                error = __zfsctl_unmount_snapshot(sep, flags);
+
+               mutex_enter(&zsb->z_ctldir_lock);
                if (error == EBUSY) {
                        avl_add(&zsb->z_ctldir_snaps, sep);
                        (*count)++;
@@ -817,7 +831,7 @@ zfsctl_mount_snapshot(struct path *path, int flags)
         * to safely abort the automount.  This should be very rare.
         */
        argv[2] = kmem_asprintf(SET_MOUNT_CMD, full_name, full_path);
-       error = call_usermodehelper(argv[0], argv, envp, 1);
+       error = call_usermodehelper(argv[0], argv, envp, UMH_WAIT_PROC);
        strfree(argv[2]);
        if (error) {
                printk("ZFS: Unable to automount %s at %s: %d\n",
@@ -837,7 +851,7 @@ zfsctl_mount_snapshot(struct path *path, int flags)
        sep = avl_find(&zsb->z_ctldir_snaps, &search, NULL);
        if (sep) {
                avl_remove(&zsb->z_ctldir_snaps, sep);
-               cancel_delayed_work_sync(&sep->se_work);
+               taskq_cancel_id(zfs_expire_taskq, sep->se_taskqid);
                zfsctl_sep_free(sep);
        }
 
@@ -847,8 +861,9 @@ zfsctl_mount_snapshot(struct path *path, int flags)
        sep->se_inode = ip;
        avl_add(&zsb->z_ctldir_snaps, sep);
 
-        spl_init_delayed_work(&sep->se_work, zfsctl_expire_snapshot, sep);
-       schedule_delayed_work(&sep->se_work, zfs_expire_snapshot * HZ);
+       sep->se_taskqid = taskq_dispatch_delay(zfs_expire_taskq,
+           zfsctl_expire_snapshot, sep, TQ_SLEEP,
+           ddi_get_lbolt() + zfs_expire_snapshot * HZ);
 
        mutex_exit(&zsb->z_ctldir_lock);
 error:
@@ -902,7 +917,7 @@ zfsctl_lookup_objset(struct super_block *sb, uint64_t objsetid, zfs_sb_t **zsbp)
         */
        sep = avl_first(&zsb->z_ctldir_snaps);
        while (sep != NULL) {
-               error = dmu_snapshot_id(zsb->z_os, sep->se_name, &id);
+               error = dmu_snapshot_lookup(zsb->z_os, sep->se_name, &id);
                if (error)
                        goto out;
 
@@ -977,6 +992,8 @@ zfsctl_shares_lookup(struct inode *dip, char *name, struct inode **ipp,
 void
 zfsctl_init(void)
 {
+       zfs_expire_taskq = taskq_create("z_unmount", 1, maxclsyspri,
+           1, 8, TASKQ_PREPOPULATE);
 }
 
 /*
@@ -986,6 +1003,7 @@ zfsctl_init(void)
 void
 zfsctl_fini(void)
 {
+       taskq_destroy(zfs_expire_taskq);
 }
 
 module_param(zfs_expire_snapshot, int, 0644);