vattr.va_uid = crgetuid(kcred);
vattr.va_gid = crgetgid(kcred);
- sharezp = kmem_cache_alloc(znode_cache, KM_SLEEP);
+ sharezp = kmem_cache_alloc(znode_cache, KM_PUSHPAGE);
sharezp->z_moved = 0;
sharezp->z_unlinked = 0;
sharezp->z_atime_dirty = 0;
{
znode_t *zp;
- zp = kmem_cache_alloc(znode_cache, KM_SLEEP);
+ zp = kmem_cache_alloc(znode_cache, KM_PUSHPAGE);
*ip = ZTOI(zp);
return (0);
*/
static znode_t *
zfs_znode_alloc(zfs_sb_t *zsb, dmu_buf_t *db, int blksz,
- dmu_object_type_t obj_type, uint64_t obj, sa_handle_t *hdl)
+ dmu_object_type_t obj_type, uint64_t obj, sa_handle_t *hdl,
+ struct dentry *dentry)
{
znode_t *zp;
struct inode *ip;
}
ip->i_ino = obj;
- ip->i_mode = zp->z_mode;
- ip->i_mtime = ip->i_atime = ip->i_ctime = CURRENT_TIME_SEC;
+ zfs_inode_update(zp);
zfs_inode_set_ops(zsb, ip);
if (insert_inode_locked(ip))
goto error;
+ if (dentry)
+ d_instantiate(dentry, ip);
+
mutex_enter(&zsb->z_znodes_lock);
list_insert_tail(&zsb->z_all_znodes, zp);
membar_producer();
VERIFY(sa_replace_all_by_template(sa_hdl, sa_attrs, cnt, tx) == 0);
if (!(flag & IS_ROOT_NODE)) {
- *zpp = zfs_znode_alloc(zsb, db, 0, obj_type, obj, sa_hdl);
+ *zpp = zfs_znode_alloc(zsb, db, 0, obj_type, obj, sa_hdl,
+ vap->va_dentry);
ASSERT(*zpp != NULL);
ASSERT(dzp != NULL);
err = zpl_xattr_security_init(ZTOI(*zpp), ZTOI(dzp));
znode_t *zp;
int err;
sa_handle_t *hdl;
+ struct inode *ip;
*zpp = NULL;
+again:
+ ip = ilookup(zsb->z_sb, obj_num);
+
ZFS_OBJ_HOLD_ENTER(zsb, obj_num);
err = sa_buf_hold(zsb->z_os, obj_num, NULL, &db);
if (err) {
ZFS_OBJ_HOLD_EXIT(zsb, obj_num);
+ iput(ip);
return (err);
}
doi.doi_bonus_size < sizeof (znode_phys_t)))) {
sa_buf_rele(db, NULL);
ZFS_OBJ_HOLD_EXIT(zsb, obj_num);
+ iput(ip);
return (EINVAL);
}
hdl = dmu_buf_get_user(db);
if (hdl != NULL) {
- zp = sa_get_userdata(hdl);
+ if (ip == NULL) {
+ /*
+ * ilookup returned NULL, which means
+ * the znode is dying - but the SA handle isn't
+ * quite dead yet, we need to drop any locks
+ * we're holding, re-schedule the task and try again.
+ */
+ sa_buf_rele(db, NULL);
+ ZFS_OBJ_HOLD_EXIT(zsb, obj_num);
+
+ schedule();
+ goto again;
+ }
+ zp = sa_get_userdata(hdl);
/*
* Since "SA" does immediate eviction we
sa_buf_rele(db, NULL);
mutex_exit(&zp->z_lock);
ZFS_OBJ_HOLD_EXIT(zsb, obj_num);
+ iput(ip);
return (err);
}
+ ASSERT3P(ip, ==, NULL);
+
/*
* Not found create new znode/vnode but only if file exists.
*
* bonus buffer.
*/
zp = zfs_znode_alloc(zsb, db, doi.doi_data_block_size,
- doi.doi_bonus_type, obj_num, NULL);
+ doi.doi_bonus_type, obj_num, NULL, NULL);
if (zp == NULL) {
err = ENOENT;
} else {
{
zfs_sb_t *zsb = ZTOZSB(zp);
uint64_t z_id = zp->z_id;
+ boolean_t drop_mutex = 0;
ASSERT(zp->z_sa_hdl);
/*
- * Don't allow a zfs_zget() while were trying to release this znode
+ * Don't allow a zfs_zget() while were trying to release this znode.
+ *
+ * Linux allows direct memory reclaim which means that any KM_SLEEP
+ * allocation may trigger inode eviction. This can lead to a deadlock
+ * through the ->shrink_icache_memory()->evict()->zfs_inactive()->
+ * zfs_zinactive() call path. To avoid this deadlock the process
+ * must not reacquire the mutex when it is already holding it.
*/
- ZFS_OBJ_HOLD_ENTER(zsb, z_id);
+ if (!ZFS_OBJ_HOLD_OWNED(zsb, z_id)) {
+ ZFS_OBJ_HOLD_ENTER(zsb, z_id);
+ drop_mutex = 1;
+ }
+
mutex_enter(&zp->z_lock);
/*
*/
if (zp->z_unlinked) {
mutex_exit(&zp->z_lock);
- ZFS_OBJ_HOLD_EXIT(zsb, z_id);
+
+ if (drop_mutex)
+ ZFS_OBJ_HOLD_EXIT(zsb, z_id);
+
zfs_rmnode(zp);
return;
}
mutex_exit(&zp->z_lock);
zfs_znode_dmu_fini(zp);
- ZFS_OBJ_HOLD_EXIT(zsb, z_id);
+
+ if (drop_mutex)
+ ZFS_OBJ_HOLD_EXIT(zsb, z_id);
}
void