On Thu, 22 Feb 2018, 12:49am, Hannes Reinecke wrote:
> Commit d8630bb95f46 ('Serialize session deletion by using work_lock') tries
> to fixup a deadlock when deleting sessions, but fails to take
> into account the locking rules. This patch resolves the situation
> by introducing a separate lock for processing the GNLIST response, and
> ensures that sess_lock is released before calling
> qlt_schedule_sess_delete().
>
> Cc: Himanshu Madhani <himanshu.madh...@cavium.com>
> Cc: Quinn Tran <quinn.t...@cavium.com>
> Fixes: d8630bb95f46 ("scsi: qla2xxx: Serialize session deletion by using
> work_lock")
> Signed-off-by: Hannes Reinecke <h...@suse.com>
> ---
> drivers/scsi/qla2xxx/qla_def.h | 4 ++--
> drivers/scsi/qla2xxx/qla_init.c | 24 +++++++++++++++---------
> drivers/scsi/qla2xxx/qla_os.c | 7 ++++++-
> drivers/scsi/qla2xxx/qla_target.c | 17 ++++++-----------
> 4 files changed, 29 insertions(+), 23 deletions(-)
>
> diff --git a/drivers/scsi/qla2xxx/qla_def.h b/drivers/scsi/qla2xxx/qla_def.h
> index be7d6824581a..3ca4b6a5eddd 100644
> --- a/drivers/scsi/qla2xxx/qla_def.h
> +++ b/drivers/scsi/qla2xxx/qla_def.h
> @@ -261,9 +261,9 @@
> struct name_list_extended {
> struct get_name_list_extended *l;
> dma_addr_t ldma;
> - struct list_head fcports; /* protect by sess_list */
> + struct list_head fcports;
> + spinlock_t fcports_lock;
> u32 size;
> - u8 sent;
> };
> /*
> * Timeout timer counts in seconds
> diff --git a/drivers/scsi/qla2xxx/qla_init.c b/drivers/scsi/qla2xxx/qla_init.c
> index 590aa904fdef..4dd897c2c2b1 100644
> --- a/drivers/scsi/qla2xxx/qla_init.c
> +++ b/drivers/scsi/qla2xxx/qla_init.c
> @@ -660,8 +660,7 @@ qla24xx_async_gnl_sp_done(void *s, int res)
> (loop_id & 0x7fff));
> }
>
> - spin_lock_irqsave(&vha->hw->tgt.sess_lock, flags);
> - vha->gnl.sent = 0;
> + spin_lock_irqsave(&vha->gnl.fcports_lock, flags);
>
> INIT_LIST_HEAD(&h);
> fcport = tf = NULL;
> @@ -670,12 +669,16 @@ qla24xx_async_gnl_sp_done(void *s, int res)
>
> list_for_each_entry_safe(fcport, tf, &h, gnl_entry) {
> list_del_init(&fcport->gnl_entry);
> + spin_lock(&vha->hw->tgt.sess_lock);
> fcport->flags &= ~(FCF_ASYNC_SENT | FCF_ASYNC_ACTIVE);
> + spin_unlock(&vha->hw->tgt.sess_lock);
> ea.fcport = fcport;
>
> qla2x00_fcport_event_handler(vha, &ea);
> }
> + spin_unlock_irqrestore(&vha->gnl.fcports_lock, flags);
>
> + spin_lock_irqsave(&vha->hw->tgt.sess_lock, flags);
> /* create new fcport if fw has knowledge of new sessions */
> for (i = 0; i < n; i++) {
> port_id_t id;
> @@ -727,18 +730,21 @@ int qla24xx_async_gnl(struct scsi_qla_host *vha,
> fc_port_t *fcport)
> ql_dbg(ql_dbg_disc, vha, 0x20d9,
> "Async-gnlist WWPN %8phC \n", fcport->port_name);
>
> - spin_lock_irqsave(&vha->hw->tgt.sess_lock, flags);
> + spin_lock_irqsave(&vha->gnl.fcports_lock, flags);
> + if (!list_empty(&fcport->gnl_entry)) {
> + spin_unlock_irqrestore(&vha->gnl.fcports_lock, flags);
> + rval = QLA_SUCCESS;
> + goto done;
> + }
> +
> + spin_lock(&vha->hw->tgt.sess_lock);
> fcport->disc_state = DSC_GNL;
> fcport->last_rscn_gen = fcport->rscn_gen;
> fcport->last_login_gen = fcport->login_gen;
> + spin_unlock(&vha->hw->tgt.sess_lock);
>
> list_add_tail(&fcport->gnl_entry, &vha->gnl.fcports);
> - if (vha->gnl.sent) {
> - spin_unlock_irqrestore(&vha->hw->tgt.sess_lock, flags);
> - return QLA_SUCCESS;
> - }
> - vha->gnl.sent = 1;
> - spin_unlock_irqrestore(&vha->hw->tgt.sess_lock, flags);
> + spin_unlock_irqrestore(&vha->gnl.fcports_lock, flags);
>
> sp = qla2x00_get_sp(vha, fcport, GFP_KERNEL);
> if (!sp)
> diff --git a/drivers/scsi/qla2xxx/qla_os.c b/drivers/scsi/qla2xxx/qla_os.c
> index 12ee6e02d146..dc2b188fdce1 100644
> --- a/drivers/scsi/qla2xxx/qla_os.c
> +++ b/drivers/scsi/qla2xxx/qla_os.c
> @@ -4575,6 +4575,7 @@ struct scsi_qla_host *qla2x00_create_host(struct
> scsi_host_template *sht,
>
> spin_lock_init(&vha->work_lock);
> spin_lock_init(&vha->cmd_list_lock);
> + spin_lock_init(&vha->gnl.fcports_lock);
> init_waitqueue_head(&vha->fcport_waitQ);
> init_waitqueue_head(&vha->vref_waitq);
>
> @@ -4875,6 +4876,8 @@ void qla24xx_create_new_sess(struct scsi_qla_host *vha,
> struct qla_work_evt *e)
> }
> qlt_plogi_ack_unref(vha, pla);
> } else {
> + fc_port_t *dfcp = NULL;
> +
> spin_lock_irqsave(&vha->hw->tgt.sess_lock, flags);
> tfcp = qla2x00_find_fcport_by_nportid(vha,
> &e->u.new_sess.id, 1);
> @@ -4897,11 +4900,13 @@ void qla24xx_create_new_sess(struct scsi_qla_host
> *vha, struct qla_work_evt *e)
> default:
> fcport->login_pause = 1;
> tfcp->conflict = fcport;
> - qlt_schedule_sess_for_deletion(tfcp);
> + dfcp = tfcp;
> break;
> }
> }
> spin_unlock_irqrestore(&vha->hw->tgt.sess_lock, flags);
> + if (dfcp)
> + qlt_schedule_sess_for_deletion(tfcp);
>
> wwn = wwn_to_u64(fcport->node_name);
>
> diff --git a/drivers/scsi/qla2xxx/qla_target.c
> b/drivers/scsi/qla2xxx/qla_target.c
> index 3735ebd83012..990ef1183f3a 100644
> --- a/drivers/scsi/qla2xxx/qla_target.c
> +++ b/drivers/scsi/qla2xxx/qla_target.c
> @@ -1224,10 +1224,10 @@ static void qla24xx_chk_fcp_state(struct fc_port
> *sess)
> }
> }
>
> -/* ha->tgt.sess_lock supposed to be held on entry */
> void qlt_schedule_sess_for_deletion(struct fc_port *sess)
> {
> struct qla_tgt *tgt = sess->tgt;
> + struct qla_hw_data *ha = sess->vha->hw;
> unsigned long flags;
>
> if (sess->disc_state == DSC_DELETE_PEND)
> @@ -1244,16 +1244,16 @@ void qlt_schedule_sess_for_deletion(struct fc_port
> *sess)
> return;
> }
>
> + spin_lock_irqsave(&ha->tgt.sess_lock, flags);
> if (sess->deleted == QLA_SESS_DELETED)
> sess->logout_on_delete = 0;
>
> - spin_lock_irqsave(&sess->vha->work_lock, flags);
> if (sess->deleted == QLA_SESS_DELETION_IN_PROGRESS) {
> - spin_unlock_irqrestore(&sess->vha->work_lock, flags);
> + spin_unlock_irqrestore(&ha->tgt.sess_lock, flags);
> return;
> }
> sess->deleted = QLA_SESS_DELETION_IN_PROGRESS;
> - spin_unlock_irqrestore(&sess->vha->work_lock, flags);
> + spin_unlock_irqrestore(&ha->tgt.sess_lock, flags);
>
> sess->disc_state = DSC_DELETE_PEND;
>
> @@ -1262,13 +1262,10 @@ void qlt_schedule_sess_for_deletion(struct fc_port
> *sess)
> ql_dbg(ql_dbg_tgt, sess->vha, 0xe001,
> "Scheduling sess %p for deletion\n", sess);
>
> - /* use cancel to push work element through before re-queue */
> - cancel_work_sync(&sess->del_work);
> INIT_WORK(&sess->del_work, qla24xx_delete_sess_fn);
> - queue_work(sess->vha->hw->wq, &sess->del_work);
> + WARN_ON(!queue_work(sess->vha->hw->wq, &sess->del_work));
> }
>
> -/* ha->tgt.sess_lock supposed to be held on entry */
> static void qlt_clear_tgt_db(struct qla_tgt *tgt)
> {
> struct fc_port *sess;
> @@ -1451,8 +1448,8 @@ qlt_fc_port_deleted(struct scsi_qla_host *vha,
> fc_port_t *fcport, int max_gen)
> ql_dbg(ql_dbg_tgt_mgt, vha, 0xf008, "qla_tgt_fc_port_deleted %p", sess);
>
> sess->local = 1;
> - qlt_schedule_sess_for_deletion(sess);
> spin_unlock_irqrestore(&vha->hw->tgt.sess_lock, flags);
> + qlt_schedule_sess_for_deletion(sess);
> }
>
> static inline int test_tgt_sess_count(struct qla_tgt *tgt)
> @@ -1512,10 +1509,8 @@ int qlt_stop_phase1(struct qla_tgt *tgt)
> * Lock is needed, because we still can get an incoming packet.
> */
> mutex_lock(&vha->vha_tgt.tgt_mutex);
> - spin_lock_irqsave(&ha->tgt.sess_lock, flags);
> tgt->tgt_stop = 1;
> qlt_clear_tgt_db(tgt);
> - spin_unlock_irqrestore(&ha->tgt.sess_lock, flags);
> mutex_unlock(&vha->vha_tgt.tgt_mutex);
> mutex_unlock(&qla_tgt_mutex);
>
>
Thanks for the patch. Looks good.
Acked-by: Himanshu Madhani <himanshu.madh...@cavium.com>
