target: Add percpu refcounting for se_lun access

This patch adds percpu refcounting for se_lun access that allows the
association of an se_lun + se_cmd in transport_lookup_cmd_lun() to
occur without an extra list_head for tracking outstanding I/O during
se_lun shutdown.

This effectively changes se_lun shutdown logic to wait for outstanding
I/O percpu references to complete in transport_lun_remove_cmd() using
se_lun->lun_ref_comp, instead of explicitly draining the per se_lun
command list and waiting for individual se_cmd descriptor processing
to complete.

Cc: Kent Overstreet <[e-mail]>
Signed-off-by: Nicholas Bellinger <[e-mail]>

drivers/target/target_core_device.c

···
 		se_cmd->pr_res_key = deve->pr_res_key;
 		se_cmd->orig_fe_lun = unpacked_lun;
 		se_cmd->se_cmd_flags |= SCF_SE_LUN_CMD;
+
+		percpu_ref_get(&se_lun->lun_ref);
+		se_cmd->lun_ref_active = true;
 	}
 	spin_unlock_irqrestore(&se_sess->se_node_acl->device_list_lock, flags);
 
···
 		se_cmd->se_lun = &se_sess->se_tpg->tpg_virt_lun0;
 		se_cmd->orig_fe_lun = 0;
 		se_cmd->se_cmd_flags |= SCF_SE_LUN_CMD;
+
+		percpu_ref_get(&se_lun->lun_ref);
+		se_cmd->lun_ref_active = true;
 	}
 
 	/* Directly associate cmd with se_dev */
···
 	else if (se_cmd->data_direction == DMA_FROM_DEVICE)
 		dev->read_bytes += se_cmd->data_length;
 	spin_unlock_irqrestore(&dev->stats_lock, flags);
-
-	spin_lock_irqsave(&se_lun->lun_cmd_lock, flags);
-	list_add_tail(&se_cmd->se_lun_node, &se_lun->lun_cmd_list);
-	spin_unlock_irqrestore(&se_lun->lun_cmd_lock, flags);
 
 	return 0;
 }

drivers/target/target_core_internal.h

···
 int	transport_dump_vpd_ident_type(struct t10_vpd *, unsigned char *, int);
 int	transport_dump_vpd_ident(struct t10_vpd *, unsigned char *, int);
 bool	target_stop_cmd(struct se_cmd *cmd, unsigned long *flags);
-int	transport_clear_lun_from_sessions(struct se_lun *);
+int	transport_clear_lun_ref(struct se_lun *);
 void	transport_send_task_abort(struct se_cmd *);
 sense_reason_t	target_cmd_size_check(struct se_cmd *cmd, unsigned int size);
 void	target_qf_do_work(struct work_struct *work);

drivers/target/target_core_tpg.c

···
 }
 EXPORT_SYMBOL(core_tpg_set_initiator_node_tag);
 
+static void core_tpg_lun_ref_release(struct percpu_ref *ref)
+{
+	struct se_lun *lun = container_of(ref, struct se_lun, lun_ref);
+
+	complete(&lun->lun_ref_comp);
+}
+
 static int core_tpg_setup_virtual_lun0(struct se_portal_group *se_tpg)
 {
 	/* Set in core_dev_setup_virtual_lun0() */
···
 	spin_lock_init(&lun->lun_acl_lock);
 	spin_lock_init(&lun->lun_cmd_lock);
 	spin_lock_init(&lun->lun_sep_lock);
+	init_completion(&lun->lun_ref_comp);
 
-	ret = core_tpg_post_addlun(se_tpg, lun, lun_access, dev);
+	ret = percpu_ref_init(&lun->lun_ref, core_tpg_lun_ref_release);
 	if (ret < 0)
 		return ret;
+
+	ret = core_tpg_post_addlun(se_tpg, lun, lun_access, dev);
+	if (ret < 0) {
+		percpu_ref_cancel_init(&lun->lun_ref);
+		return ret;
+	}
 
 	return 0;
 }
···
 		spin_lock_init(&lun->lun_acl_lock);
 		spin_lock_init(&lun->lun_cmd_lock);
 		spin_lock_init(&lun->lun_sep_lock);
+		init_completion(&lun->lun_ref_comp);
 	}
 
 	se_tpg->se_tpg_type = se_tpg_type;
···
 {
 	int ret;
 
-	ret = core_dev_export(lun_ptr, tpg, lun);
+	ret = percpu_ref_init(&lun->lun_ref, core_tpg_lun_ref_release);
 	if (ret < 0)
 		return ret;
+
+	ret = core_dev_export(lun_ptr, tpg, lun);
+	if (ret < 0) {
+		percpu_ref_cancel_init(&lun->lun_ref);
+		return ret;
+	}
 
 	spin_lock(&tpg->tpg_lun_lock);
 	lun->lun_access = lun_access;
···
 	struct se_lun *lun)
 {
 	core_clear_lun_from_tpg(lun, tpg);
-	transport_clear_lun_from_sessions(lun);
+	transport_clear_lun_ref(lun);
 }
 
 struct se_lun *core_tpg_pre_dellun(

drivers/target/target_core_transport.c

···
 static void transport_lun_remove_cmd(struct se_cmd *cmd)
 {
 	struct se_lun *lun = cmd->se_lun;
-	unsigned long flags;
 
-	if (!lun)
+	if (!lun || !cmd->lun_ref_active)
 		return;
 
-	spin_lock_irqsave(&lun->lun_cmd_lock, flags);
-	if (!list_empty(&cmd->se_lun_node))
-		list_del_init(&cmd->se_lun_node);
-	spin_unlock_irqrestore(&lun->lun_cmd_lock, flags);
+	percpu_ref_put(&lun->lun_ref);
 }
 
 void transport_cmd_finish_abort(struct se_cmd *cmd, int remove)
···
 	spin_unlock_irqrestore(&lun->lun_cmd_lock, lun_flags);
 }
 
-static int transport_clear_lun_thread(void *p)
+static int transport_clear_lun_ref_thread(void *p)
 {
 	struct se_lun *lun = p;
 
-	__transport_clear_lun_from_sessions(lun);
+	percpu_ref_kill(&lun->lun_ref);
+
+	wait_for_completion(&lun->lun_ref_comp);
 	complete(&lun->lun_shutdown_comp);
 
 	return 0;
 }
 
-int transport_clear_lun_from_sessions(struct se_lun *lun)
+int transport_clear_lun_ref(struct se_lun *lun)
 {
 	struct task_struct *kt;
 
-	kt = kthread_run(transport_clear_lun_thread, lun,
+	kt = kthread_run(transport_clear_lun_ref_thread, lun,
 			"tcm_cl_%u", lun->unpacked_lun);
 	if (IS_ERR(kt)) {
 		pr_err("Unable to start clear_lun thread\n");

drivers/target/target_core_xcopy.c

···
 	spin_lock_init(&pt_cmd->se_lun->lun_acl_lock);
 	spin_lock_init(&pt_cmd->se_lun->lun_cmd_lock);
 	spin_lock_init(&pt_cmd->se_lun->lun_sep_lock);
+	init_completion(&pt_cmd->se_lun->lun_ref_comp);
 
 	pt_cmd->se_dev = se_dev;
 

include/target/target_core_base.h

···
 
 	/* backend private data */
 	void			*priv;
+
+	/* Used for lun->lun_ref counting */
+	bool			lun_ref_active;
 };
 
 struct se_ua {
···
 	struct se_port		*lun_sep;
 	struct config_group	lun_group;
 	struct se_port_stat_grps port_stat_grps;
+	struct completion	lun_ref_comp;
+	struct percpu_ref	lun_ref;
 };
 
 struct scsi_port_stats {