[LU-6378] Quota performance issue for 2.7

• Type: Bug
• Resolution: Fixed
• Priority: Major
• Fix Version/s: Lustre 2.8.0
• Affects Version/s: Lustre 2.7.0, Lustre 2.8.0
• Severity: 3
• Assignee: nasf (Inactive)
• Reporter: Di Wang

Description

In lustre 2.7, inode I_NEW state is cleared in the later phase, which cause the following owner/group changing (in osd_attr_init0()) will be performed in I_NEW state, which means the dquot_initialize() won't be skipped for these operations,

2.7.0

/**
 * Helper function for osd_object_create()
 *
 * \retval 0, on success
 */
static int __osd_object_create(struct osd_thread_info *info,
                               struct osd_object *obj, struct lu_attr *attr,
                               struct dt_allocation_hint *hint,
                               struct dt_object_format *dof,
                               struct thandle *th)
{
        int     result;
        __u32   umask;

        /* we drop umask so that permissions we pass are not affected */
        umask = current->fs->umask;
        current->fs->umask = 0;

        result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
                                                  th);
        if (result == 0) {
                osd_attr_init(info, obj, attr, dof);
                osd_object_init0(obj);
        }

        if (obj->oo_inode != NULL) {
                LASSERT(obj->oo_inode->i_state & I_NEW);

                unlock_new_inode(obj->oo_inode);  --->It will unlock new inode here, so it will cause the above osd_attr_init to call quota_initialization (with I_NEW).  But in 2.6.0, we used to unlock the new inode in ldiskfs_new_inode.
        }

        /* restore previous umask value */
        current->fs->umask = umask;

        return result;
}

It is caused by this patch
http://review.whamcloud.com/13187

Issue Links

• is related to: LU-6381 replace global dq_state_lock/dq_list_lock with per-sb spinlocks and per-sb hash table.