Thanks for the review, Ewan. My response inline..
On Mon, 17 Jun 2019, 11:38am, Ewan D. Milne wrote:
> See below.
>
> On Fri, 2019-06-14 at 15:10 -0700, Himanshu Madhani wrote:
> > From: Arun Easi <ae...@marvell.com>
> >
> > BUG: unable to handle kernel NULL pointer dereference at (null)
> > IP: [<ffffffffc050d10c>] qla_nvme_unregister_remote_port+0x6c/0xf0 [qla2xxx]
> > PGD 800000084cf41067 PUD 84d288067 PMD 0
> > Oops: 0000 [#1] SMP
> > Call Trace:
> > [<ffffffff98abcfdf>] process_one_work+0x17f/0x440
> > [<ffffffff98abdca6>] worker_thread+0x126/0x3c0
> > [<ffffffff98abdb80>] ? manage_workers.isra.26+0x2a0/0x2a0
> > [<ffffffff98ac4f81>] kthread+0xd1/0xe0
> > [<ffffffff98ac4eb0>] ? insert_kthread_work+0x40/0x40
> > [<ffffffff9918ad37>] ret_from_fork_nospec_begin+0x21/0x21
> > [<ffffffff98ac4eb0>] ? insert_kthread_work+0x40/0x40
> > RIP [<ffffffffc050d10c>] qla_nvme_unregister_remote_port+0x6c/0xf0
> > [qla2xxx]
> >
> > The crash is due to a bad entry in the nvme_rport_list. This list is not
> > protected, and when a remoteport_delete callback is called, driver
> > traverses the list and crashes.
> >
> > Actually, the list could be removed and driver could traverse the main
> > fcport list instead. Fix does exactly that.
> >
> > Signed-off-by: Arun Easi <ae...@marvell.com>
> > Signed-off-by: Himanshu Madhani <hmadh...@marvell.com>
> > ---
> > drivers/scsi/qla2xxx/qla_def.h | 1 -
> > drivers/scsi/qla2xxx/qla_nvme.c | 52
> > ++++++++++++++++++++---------------------
> > drivers/scsi/qla2xxx/qla_nvme.h | 1 -
> > drivers/scsi/qla2xxx/qla_os.c | 1 -
> > 4 files changed, 25 insertions(+), 30 deletions(-)
> >
> > diff --git a/drivers/scsi/qla2xxx/qla_def.h b/drivers/scsi/qla2xxx/qla_def.h
> > index 1a4095c56eee..602ed24bb806 100644
> > --- a/drivers/scsi/qla2xxx/qla_def.h
> > +++ b/drivers/scsi/qla2xxx/qla_def.h
> > @@ -4376,7 +4376,6 @@ typedef struct scsi_qla_host {
> >
> > struct nvme_fc_local_port *nvme_local_port;
> > struct completion nvme_del_done;
> > - struct list_head nvme_rport_list;
> >
> > uint16_t fcoe_vlan_id;
> > uint16_t fcoe_fcf_idx;
> > diff --git a/drivers/scsi/qla2xxx/qla_nvme.c
> > b/drivers/scsi/qla2xxx/qla_nvme.c
> > index 22e3fba28e51..99220a3cf734 100644
> > --- a/drivers/scsi/qla2xxx/qla_nvme.c
> > +++ b/drivers/scsi/qla2xxx/qla_nvme.c
> > @@ -14,6 +14,18 @@ static struct nvme_fc_port_template
> > qla_nvme_fc_transport;
> >
> > static void qla_nvme_unregister_remote_port(struct work_struct *);
> >
> > +static inline
> > +int qla_is_active_nvme_fcport(struct fc_port *fcport)
> > +{
> > + return fcport->nvme_flag & NVME_FLAG_REGISTERED;
> > +}
> > +
>
> Nitpick: "qla_is_active_nvme_fcport", the qualifier "active"
> does not match any of the flag definitions and this is only
> checking the REGISTERED flag. Maybe "qla_is_registered_nvme_fcport"
> instead? Would help to understand vs. DELETING and RESETTING case.
Sure, this sounds better, will do in v2.
>
> > +#define qla_list_for_each_nvme_fcport(_fcport, _vha) \
> > +{ \
> > + list_for_each_entry(_fcport, &_vha->vp_fcports, list) \
> > + if (qla_is_active_nvme_fcport(_fcport)) \
> > +}
> > +
>
> Like here, "if (qla_is_registered_nvme_fcport(" would be clearer.
>
> > int qla_nvme_register_remote(struct scsi_qla_host *vha, struct fc_port
> > *fcport)
> > {
> > struct qla_nvme_rport *rport;
> > @@ -74,7 +86,6 @@ int qla_nvme_register_remote(struct scsi_qla_host *vha,
> > struct fc_port *fcport)
> >
> > rport = fcport->nvme_remote_port->private;
> > rport->fcport = fcport;
> > - list_add_tail(&rport->list, &vha->nvme_rport_list);
> >
> > fcport->nvme_flag |= NVME_FLAG_REGISTERED;
> > return 0;
> > @@ -542,19 +553,12 @@ static void qla_nvme_localport_delete(struct
> > nvme_fc_local_port *lport)
> > static void qla_nvme_remoteport_delete(struct nvme_fc_remote_port *rport)
> > {
> > fc_port_t *fcport;
> > - struct qla_nvme_rport *qla_rport = rport->private, *trport;
> > + struct qla_nvme_rport *qla_rport = rport->private;
> >
> > fcport = qla_rport->fcport;
> > fcport->nvme_remote_port = NULL;
> > fcport->nvme_flag &= ~NVME_FLAG_REGISTERED;
> >
> > - list_for_each_entry_safe(qla_rport, trport,
> > - &fcport->vha->nvme_rport_list, list) {
> > - if (qla_rport->fcport == fcport) {
> > - list_del(&qla_rport->list);
> > - break;
> > - }
> > - }
> > complete(&fcport->nvme_del_done);
> >
> > if (!test_bit(UNLOADING, &fcport->vha->dpc_flags)) {
> > @@ -590,31 +594,25 @@ static void qla_nvme_unregister_remote_port(struct
> > work_struct *work)
> > {
> > struct fc_port *fcport = container_of(work, struct fc_port,
> > nvme_del_work);
> > - struct qla_nvme_rport *qla_rport, *trport;
> > + int ret;
> >
> > if (!IS_ENABLED(CONFIG_NVME_FC))
> > return;
> >
> > + if (!qla_is_active_nvme_fcport(fcport))
> > + return;
> > +
>
> This looks like it is prone to a race because the REGISTERED flag
> is cleared in the callback from the NVMe transport?
The remoteport_delete callback gets invoked only after a
nvme_fc_unregister_remoteport() which has not yet happened, so I am
not sure where the race is.
> The flag was already checked in qlt_unreg_sess(), then the work item
> runs here, then later qla_nvme_remoteport_delete() runs. What is
> happening here that the check is preventing?
This was merely a defensive check added, that I guess got stuck in mind
somehow. Anycase, you are right, that is not needed and can be removed.
-- arun
>
> -Ewan
>
> > ql_log(ql_log_warn, NULL, 0x2112,
> > "%s: unregister remoteport on %p\n",__func__, fcport);
> >
> > - list_for_each_entry_safe(qla_rport, trport,
> > - &fcport->vha->nvme_rport_list, list) {
> > - if (qla_rport->fcport == fcport) {
> > - ql_log(ql_log_info, fcport->vha, 0x2113,
> > - "%s: fcport=%p\n", __func__, fcport);
> > - nvme_fc_set_remoteport_devloss
> > - (fcport->nvme_remote_port, 0);
> > - init_completion(&fcport->nvme_del_done);
> > - if (nvme_fc_unregister_remoteport
> > - (fcport->nvme_remote_port))
> > - ql_log(ql_log_info, fcport->vha, 0x2114,
> > - "%s: Failed to unregister
> > nvme_remote_port\n",
> > - __func__);
> > - wait_for_completion(&fcport->nvme_del_done);
> > - break;
> > - }
> > - }
> > + nvme_fc_set_remoteport_devloss(fcport->nvme_remote_port, 0);
> > + init_completion(&fcport->nvme_del_done);
> > + ret = nvme_fc_unregister_remoteport(fcport->nvme_remote_port);
> > + if (ret)
> > + ql_log(ql_log_info, fcport->vha, 0x2114,
> > + "%s: Failed to unregister nvme_remote_port (%d)\n",
> > + __func__, ret);
> > + wait_for_completion(&fcport->nvme_del_done);
> > }
> >
> > void qla_nvme_delete(struct scsi_qla_host *vha)
> > diff --git a/drivers/scsi/qla2xxx/qla_nvme.h
> > b/drivers/scsi/qla2xxx/qla_nvme.h
> > index d3b8a6440113..2d088add7011 100644
> > --- a/drivers/scsi/qla2xxx/qla_nvme.h
> > +++ b/drivers/scsi/qla2xxx/qla_nvme.h
> > @@ -37,7 +37,6 @@ struct nvme_private {
> > };
> >
> > struct qla_nvme_rport {
> > - struct list_head list;
> > struct fc_port *fcport;
> > };
> >
> > diff --git a/drivers/scsi/qla2xxx/qla_os.c b/drivers/scsi/qla2xxx/qla_os.c
> > index 00fee5bf4de1..ae93ae2b6090 100644
> > --- a/drivers/scsi/qla2xxx/qla_os.c
> > +++ b/drivers/scsi/qla2xxx/qla_os.c
> > @@ -4789,7 +4789,6 @@ struct scsi_qla_host *qla2x00_create_host(struct
> > scsi_host_template *sht,
> > INIT_LIST_HEAD(&vha->plogi_ack_list);
> > INIT_LIST_HEAD(&vha->qp_list);
> > INIT_LIST_HEAD(&vha->gnl.fcports);
> > - INIT_LIST_HEAD(&vha->nvme_rport_list);
> > INIT_LIST_HEAD(&vha->gpnid_list);
> > INIT_WORK(&vha->iocb_work, qla2x00_iocb_work_fn);
> >
>
