Mon Jun 17 08:09:57 2019 UTC ()

Add native vfs_suspend()/vfs_resume() before and after
zfs_suspend_fs()/zfs_resume_fs() and get rid of dead "z_sa_hdl == NULL"
znodes before vfs_resume() to keep the vnode cache consistent.

Live rollback should work now.

PR port-xen/54273 ("zpool create pool xbd2" panics DOMU kernel)


(hannken)

diff -r1.23 -r1.24 src/external/cddl/osnet/dist/uts/common/fs/zfs/zfs_vfsops.c

--- src/external/cddl/osnet/dist/uts/common/fs/zfs/zfs_vfsops.c 2019/05/22 08:45:32	1.23
+++ src/external/cddl/osnet/dist/uts/common/fs/zfs/zfs_vfsops.c 2019/06/17 08:09:57	1.24

 @@ -111,26 +111,27 @@ struct vfsops zfs_vfsops = {
 	.vfs_statfs =		zfs_statfs,
 	.vfs_vget =		zfs_vget,
 	.vfs_sync =		zfs_sync,
 	.vfs_checkexp =		zfs_checkexp,
 	.vfs_fhtovp =		zfs_fhtovp,
 };
 VFS_SET(zfs_vfsops, zfs, VFCF_JAIL | VFCF_DELEGADMIN);
 #endif /* __FreeBSD_kernel__ */
 #ifdef __NetBSD__
 #include <sys/fstrans.h>
 #include <sys/mkdev.h>
 #include <miscfs/genfs/genfs.h>
 int zfs_debug_level;
 kmutex_t zfs_debug_mtx;
 #define	DROP_GIANT()	/* nothing */
 #define PICKUP_GIANT()	/* nothing */
 #define vfs_stdsync(a, b) 0
 static int zfs_mount(vfs_t *vfsp, const char *path, void *data, size_t *data_len);
 static int zfs_umount(vfs_t *vfsp, int fflag);
 static int zfs_root(vfs_t *vfsp, int flags, vnode_t **vpp);
 @@ -2659,39 +2660,58 @@ zfs_fhtovp(vfs_t *vfsp, fid_t *fidp, int
 /*
  * Block out VOPs and close zfsvfs_t::z_os
+ *
  * Note, if successful, then we return with the 'z_teardown_lock' and
  * 'z_teardown_inactive_lock' write held.  We leave ownership of the underlying
  * dataset and objset intact so that they can be atomically handed off during
  * a subsequent rollback or recv operation and the resume thereafter.
  */
 int
 zfs_suspend_fs(zfsvfs_t *zfsvfs)
+{
 	int error;
 #ifdef __NetBSD__
 	if ((error = vfs_suspend(zfsvfs->z_vfs, 0)) != 0)
 		return error;
 	if ((error = zfsvfs_teardown(zfsvfs, B_FALSE)) != 0) {
 		vfs_resume(zfsvfs->z_vfs);
 		return (error);
+	}
 #else
 	if ((error = zfsvfs_teardown(zfsvfs, B_FALSE)) != 0)
 		return (error);
 #endif
 	return (0);
+}
 /*
  * Rebuild SA and release VOPs.  Note that ownership of the underlying dataset
  * is an invariant across any of the operations that can be performed while the
  * filesystem was suspended.  Whether it succeeded or failed, the preconditions
  * are the same: the relevant objset and associated dataset are owned by
  * zfsvfs, held, and long held on entry.
  */
 #ifdef __NetBSD__
 static bool
 zfs_resume_selector(void *cl, struct vnode *vp)
+{
 	if (zfsctl_is_node(vp))
 		return false;
 	return (VTOZ(vp)->z_sa_hdl == NULL);
+}
 #endif
 int
 zfs_resume_fs(zfsvfs_t *zfsvfs, dsl_dataset_t *ds)
+{
 	int err;
 	znode_t *zp;
 	ASSERT(RRM_WRITE_HELD(&zfsvfs->z_teardown_lock));
 	ASSERT(RW_WRITE_HELD(&zfsvfs->z_teardown_inactive_lock));
 	/*
 	 * We already own this, so just update the objset_t, as the one we
 	 * had before may have been evicted.
 	 */
 @@ -2715,26 +2735,38 @@ zfs_resume_fs(zfsvfs_t *zfsvfs, dsl_data
 	 * when they try to use their znode.
 	 */
 	mutex_enter(&zfsvfs->z_znodes_lock);
 	for (zp = list_head(&zfsvfs->z_all_znodes); zp;
 	    zp = list_next(&zfsvfs->z_all_znodes, zp)) {
 		(void) zfs_rezget(zp);
+	}
 	mutex_exit(&zfsvfs->z_znodes_lock);
 bail:
 	/* release the VOPs */
 	rw_exit(&zfsvfs->z_teardown_inactive_lock);
 	rrm_exit(&zfsvfs->z_teardown_lock, FTAG);
 #ifdef __NetBSD__
 	struct vnode_iterator *marker;
 	vnode_t *vp;
 	vfs_vnode_iterator_init(zfsvfs->z_vfs, &marker);
 	while ((vp = vfs_vnode_iterator_next(marker,
 	    zfs_resume_selector, NULL))) {
 		vgone(vp);
+	}
 	vfs_vnode_iterator_destroy(marker);
 	vfs_resume(zfsvfs->z_vfs);
 #endif
 	if (err) {
 		/*
 		 * Since we couldn't setup the sa framework, try to force
 		 * unmount this file system.
 		 */
 		if (vn_vfswlock(zfsvfs->z_vfs->vfs_vnodecovered) == 0) {
 			vfs_ref(zfsvfs->z_vfs);
 			(void) dounmount(zfsvfs->z_vfs, MS_FORCE, curthread);
+		}
+	}
 	return (err);
+}