nvme-fc: decouple ns references from lldd references

In the lldd api, a lldd may unregister a remoteport (loss of connectivity
or driver unload) or localport (driver unload). The lldd must wait for the
remoteport_delete or localport_delete before completing its actions post
the unregister.  The xxx_deletes currently occur only when the xxxport
structure is fully freed after all references are removed. Thus the lldd
may be held hostage until an app or in-kernel entity that has a namespace
open finally closes so the namespace can be removed, the controller
removed, thus the transport objects, thus the lldd.

This patch decouples the transport and os-facing objects from the lldd
and the remoteport and localport. There is a point in all deletions
where the transport will no longer interact with the lldd on behalf of
a controller. That point centers around the association established
with the target/subsystem. It will access the lldd whenever it attempts
to create an association and while the association is active. New
associations may only be created if the remoteport is live (thus the
localport is live). It will not access the lldd after deleting the
association.

Therefore, the patch tracks the count of active controllers - those with
associations being created or that are active - on a remoteport. It also
tracks the number of remoteports that have active controllers, on a
a localport. When a remoteport is unregistered, as soon as there are no
active controllers, the lldd's remoteport_delete may be called and the
lldd may continue. Similarly, when a localport is unregistered, as soon
as there are no remoteports with active controllers, the localport_delete
callback may be made. This significantly speeds up unregistration with
the lldd.

The transport objects continue in suspended status with reconnect timers
running, and upon expiration, normal ref-counting will occur and the
objects will be freed. The transport object may still be held hostage
by the application/kernel module, but that is acceptable.

With this change, the lldd may be fully unloaded and reloaded, and
if registrations occur prior to the timeouts, the nvme controller and
namespaces will resume normally as if a link bounce.
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

drivers/nvme/host/fc.c

@@ -126,6 +126,7 @@ struct nvme_fc_lport {
 	struct device			*dev;	/* physical device for dma */
 	struct nvme_fc_port_template	*ops;
 	struct kref			ref;
+	atomic_t                        act_rport_cnt;
 } __aligned(sizeof(u64));	/* alignment for other things alloc'd with */
 
 struct nvme_fc_rport {
@@ -138,6 +139,7 @@ struct nvme_fc_rport {
 	struct nvme_fc_lport		*lport;
 	spinlock_t			lock;
 	struct kref			ref;
+	atomic_t                        act_ctrl_cnt;
 	unsigned long			dev_loss_end;
 } __aligned(sizeof(u64));	/* alignment for other things alloc'd with */
 
@@ -153,6 +155,7 @@ struct nvme_fc_ctrl {
 	struct nvme_fc_rport	*rport;
 	u32			cnum;
 
+	bool			assoc_active;
 	u64			association_id;
 
 	struct list_head	ctrl_list;	/* rport->ctrl_list */
@@ -243,9 +246,6 @@ nvme_fc_free_lport(struct kref *ref)
 	list_del(&lport->port_list);
 	spin_unlock_irqrestore(&nvme_fc_lock, flags);
 
-	/* let the LLDD know we've finished tearing it down */
-	lport->ops->localport_delete(&lport->localport);
-
 	ida_simple_remove(&nvme_fc_local_port_cnt, lport->localport.port_num);
 	ida_destroy(&lport->endp_cnt);
 
@@ -400,6 +400,7 @@ nvme_fc_register_localport(struct nvme_fc_port_info *pinfo,
 	INIT_LIST_HEAD(&newrec->port_list);
 	INIT_LIST_HEAD(&newrec->endp_list);
 	kref_init(&newrec->ref);
+	atomic_set(&newrec->act_rport_cnt, 0);
 	newrec->ops = template;
 	newrec->dev = dev;
 	ida_init(&newrec->endp_cnt);
@@ -462,6 +463,9 @@ nvme_fc_unregister_localport(struct nvme_fc_local_port *portptr)
 
 	spin_unlock_irqrestore(&nvme_fc_lock, flags);
 
+	if (atomic_read(&lport->act_rport_cnt) == 0)
+		lport->ops->localport_delete(&lport->localport);
+
 	nvme_fc_lport_put(lport);
 
 	return 0;
@@ -515,9 +519,6 @@ nvme_fc_free_rport(struct kref *ref)
 	list_del(&rport->endp_list);
 	spin_unlock_irqrestore(&nvme_fc_lock, flags);
 
-	/* let the LLDD know we've finished tearing it down */
-	lport->ops->remoteport_delete(&rport->remoteport);
-
 	ida_simple_remove(&lport->endp_cnt, rport->remoteport.port_num);
 
 	kfree(rport);
@@ -704,6 +705,7 @@ nvme_fc_register_remoteport(struct nvme_fc_local_port *localport,
 	INIT_LIST_HEAD(&newrec->ctrl_list);
 	INIT_LIST_HEAD(&newrec->ls_req_list);
 	kref_init(&newrec->ref);
+	atomic_set(&newrec->act_ctrl_cnt, 0);
 	spin_lock_init(&newrec->lock);
 	newrec->remoteport.localport = &lport->localport;
 	newrec->dev = lport->dev;
@@ -859,6 +861,9 @@ nvme_fc_unregister_remoteport(struct nvme_fc_remote_port *portptr)
 
 	nvme_fc_abort_lsops(rport);
 
+	if (atomic_read(&rport->act_ctrl_cnt) == 0)
+		rport->lport->ops->remoteport_delete(portptr);
+
 	/*
 	 * release the reference, which will allow, if all controllers
 	 * go away, which should only occur after dev_loss_tmo occurs,
@@ -2639,6 +2644,61 @@ nvme_fc_reinit_io_queues(struct nvme_fc_ctrl *ctrl)
 	return ret;
 }
 
+static void
+nvme_fc_rport_active_on_lport(struct nvme_fc_rport *rport)
+{
+	struct nvme_fc_lport *lport = rport->lport;
+
+	atomic_inc(&lport->act_rport_cnt);
+}
+
+static void
+nvme_fc_rport_inactive_on_lport(struct nvme_fc_rport *rport)
+{
+	struct nvme_fc_lport *lport = rport->lport;
+	u32 cnt;
+
+	cnt = atomic_dec_return(&lport->act_rport_cnt);
+	if (cnt == 0 && lport->localport.port_state == FC_OBJSTATE_DELETED)
+		lport->ops->localport_delete(&lport->localport);
+}
+
+static int
+nvme_fc_ctlr_active_on_rport(struct nvme_fc_ctrl *ctrl)
+{
+	struct nvme_fc_rport *rport = ctrl->rport;
+	u32 cnt;
+
+	if (ctrl->assoc_active)
+		return 1;
+
+	ctrl->assoc_active = true;
+	cnt = atomic_inc_return(&rport->act_ctrl_cnt);
+	if (cnt == 1)
+		nvme_fc_rport_active_on_lport(rport);
+
+	return 0;
+}
+
+static int
+nvme_fc_ctlr_inactive_on_rport(struct nvme_fc_ctrl *ctrl)
+{
+	struct nvme_fc_rport *rport = ctrl->rport;
+	struct nvme_fc_lport *lport = rport->lport;
+	u32 cnt;
+
+	/* ctrl->assoc_active=false will be set independently */
+
+	cnt = atomic_dec_return(&rport->act_ctrl_cnt);
+	if (cnt == 0) {
+		if (rport->remoteport.port_state == FC_OBJSTATE_DELETED)
+			lport->ops->remoteport_delete(&rport->remoteport);
+		nvme_fc_rport_inactive_on_lport(rport);
+	}
+
+	return 0;
+}
+
 /*
  * This routine restarts the controller on the host side, and
  * on the link side, recreates the controller association.
@@ -2655,6 +2715,9 @@ nvme_fc_create_association(struct nvme_fc_ctrl *ctrl)
 	if (ctrl->rport->remoteport.port_state != FC_OBJSTATE_ONLINE)
 		return -ENODEV;
 
+	if (nvme_fc_ctlr_active_on_rport(ctrl))
+		return -ENOTUNIQ;
+
 	/*
 	 * Create the admin queue
 	 */
@@ -2762,6 +2825,8 @@ nvme_fc_create_association(struct nvme_fc_ctrl *ctrl)
 	__nvme_fc_delete_hw_queue(ctrl, &ctrl->queues[0], 0);
 out_free_queue:
 	nvme_fc_free_queue(&ctrl->queues[0]);
+	ctrl->assoc_active = false;
+	nvme_fc_ctlr_inactive_on_rport(ctrl);
 
 	return ret;
 }
@@ -2777,6 +2842,10 @@ nvme_fc_delete_association(struct nvme_fc_ctrl *ctrl)
 {
 	unsigned long flags;
 
+	if (!ctrl->assoc_active)
+		return;
+	ctrl->assoc_active = false;
+
 	spin_lock_irqsave(&ctrl->lock, flags);
 	ctrl->flags |= FCCTRL_TERMIO;
 	ctrl->iocnt = 0;
@@ -2849,6 +2918,8 @@ nvme_fc_delete_association(struct nvme_fc_ctrl *ctrl)
 
 	__nvme_fc_delete_hw_queue(ctrl, &ctrl->queues[0], 0);
 	nvme_fc_free_queue(&ctrl->queues[0]);
+
+	nvme_fc_ctlr_inactive_on_rport(ctrl);
 }
 
 static void
@@ -3048,6 +3119,7 @@ nvme_fc_init_ctrl(struct device *dev, struct nvmf_ctrl_options *opts,
 	ctrl->rport = rport;
 	ctrl->dev = lport->dev;
 	ctrl->cnum = idx;
+	ctrl->assoc_active = false;
 
 	get_device(ctrl->dev);
 	kref_init(&ctrl->ref);