-
Bug
-
Resolution: Unresolved
-
Medium
-
None
-
None
-
None
-
3
-
9223372036854775807
Quota slave could be offline at the moment when lqe_edquot is set to some quota ID at the QMT, i.e. it wouldn't get -EDQUOT notification. When this target comes back and new data is going to be written to this quota ID, qsd enqueues intent per-ID quota lock(IT_QUOTA_DQACQ) to grant quotas. At the QMT side it tries to grant more quotas(qmt_intent_policy) and fill lvbo(qmt_lvbo_fill). Qunit and LQUOTA_FL_EDQUOT are set at this step in qmt_lvbo_fill. However, they both are ignored at he qsd side because ret == -EDQUOT:
void qsd_req_completion(const struct lu_env *env, struct qsd_qtype_info *qqi,
struct quota_body *reqbody, struct quota_body *repbody,
struct lustre_handle *lockh, struct lquota_lvb *lvb,
void *arg, int ret)
{
...
/* extract information from lvb */
if (ret == 0 && lvb != NULL) {
if (lvb->lvb_id_qunit != 0)
qsd_set_qunit(lqe, lvb->lvb_id_qunit);
qsd_set_edquot(lqe, !!(lvb->lvb_flags & LQUOTA_FL_EDQUOT));
} else if (repbody != NULL && repbody->qb_qunit != 0) {
qsd_set_qunit(lqe, repbody->qb_qunit);
}
Usually, it is not critical because EDQUOT flag would be sent through qmt_glimpse. However, I faced a case when qmt_glimpse wasn't called and as a result qsd didn't set edquot. That said, it was possible to write at this OST despite of the -EDQUOT at the MDT.
Even if qmt_glimpse would be sent, there is a delay between handling the 1st qsd_req_completion and qsd_id_glimpse_ast. During this time new data could be written to the OST, despite of the -EDQUOT at the QMT.
It is easy to fix. I'll send a fix soon.