| Message ID | 20201012091100.55305-1-dwagner@suse.de (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | qla2xxx: Return EBUSY on fcport deletion | expand |
On Mon, 12 Oct 2020, 2:11am, Daniel Wagner wrote: > When the fcport is about to be deleted we should return EBUSY instead > of ENODEV. Only for EBUSY the request will be requeued in a multipath > setup. > > Also in case we have a valid qpair but the firmware has not yet > started return EBUSY to avoid dropping the request. > > Signed-off-by: Daniel Wagner <dwagner@suse.de> > --- > Hi, > > During port bounce and fail tests we observed that requests get > dropped on a failing path because the driver returned ENODEV and thus > the multipath code didn't requeue the request. > > The tests were done with only the 'fcport && fcport->deleted' condition > but Hannes suggested we might as well do the same for 'qpair && > !qpair->fw_started'. > > Thanks, > Daniel > > drivers/scsi/qla2xxx/qla_nvme.c | 5 ++++- > 1 file changed, 4 insertions(+), 1 deletion(-) > > diff --git a/drivers/scsi/qla2xxx/qla_nvme.c b/drivers/scsi/qla2xxx/qla_nvme.c > index 5cc1bbb1ed74..db8b802b147c 100644 > --- a/drivers/scsi/qla2xxx/qla_nvme.c > +++ b/drivers/scsi/qla2xxx/qla_nvme.c > @@ -555,8 +555,11 @@ static int qla_nvme_post_cmd(struct nvme_fc_local_port *lport, > > fcport = qla_rport->fcport; > > - if (!qpair || !fcport || (qpair && !qpair->fw_started) || > + if ((qpair && !qpair->fw_started) || > (fcport && fcport->deleted)) > + return -EBUSY; > + > + if (!qpair || !fcport) > return rval; > > vha = fcport->vha; > This does not appear to be cut against the latest for-next/staging; "rval" is not used there for the initial set of returns. Anyway, returning EBUSY is the right way to go. Regards, -Arun
On Mon, Oct 12, 2020 at 09:07:15AM -0700, Arun Easi wrote: > This does not appear to be cut against the latest for-next/staging; "rval" > is not used there for the initial set of returns. Indeed, forgot to use staging. It's against queue. Let me update it. > Anyway, returning EBUSY is the right way to go. Thanks for the confirmation. Daniel
diff --git a/drivers/scsi/qla2xxx/qla_nvme.c b/drivers/scsi/qla2xxx/qla_nvme.c index 5cc1bbb1ed74..db8b802b147c 100644 --- a/drivers/scsi/qla2xxx/qla_nvme.c +++ b/drivers/scsi/qla2xxx/qla_nvme.c @@ -555,8 +555,11 @@ static int qla_nvme_post_cmd(struct nvme_fc_local_port *lport, fcport = qla_rport->fcport; - if (!qpair || !fcport || (qpair && !qpair->fw_started) || + if ((qpair && !qpair->fw_started) || (fcport && fcport->deleted)) + return -EBUSY; + + if (!qpair || !fcport) return rval; vha = fcport->vha;
When the fcport is about to be deleted we should return EBUSY instead of ENODEV. Only for EBUSY the request will be requeued in a multipath setup. Also in case we have a valid qpair but the firmware has not yet started return EBUSY to avoid dropping the request. Signed-off-by: Daniel Wagner <dwagner@suse.de> --- Hi, During port bounce and fail tests we observed that requests get dropped on a failing path because the driver returned ENODEV and thus the multipath code didn't requeue the request. The tests were done with only the 'fcport && fcport->deleted' condition but Hannes suggested we might as well do the same for 'qpair && !qpair->fw_started'. Thanks, Daniel drivers/scsi/qla2xxx/qla_nvme.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)