qla2xxx: Wait for all conflicts before ack'ing PLOGI

Until now ack'ing of a new PLOGI has only been delayed if there
was an existing session for the same WWN. Ack was released when
the session deletion completed.

If there was another WWN session with the same fc_id/loop_id pair
(aka "conflicting session"), PLOGI was still ack'ed immediately.
This potentially caused a problem when old session deletion logged
fc_id/loop_id out of FW after new session has been established.

Two work-arounds were attempted before:
1. Dropping PLOGIs until conflicting session goes away.
2. Detecting initiator being logged out of FW and issuing LOGO
to force re-login.

This patch introduces proper solution to the problem where PLOGI
is held until either existing session with same WWN or any
conflicting session goes away. Mechanism supports one session holding
two PLOGI acks as well as one PLOGI ack being held by many sessions.

Signed-off-by: Alexei Potashnik <alexei@purestorage.com>
Acked-by: Quinn Tran <quinn.tran@qlogic.com>
Signed-off-by: Himanshu Madhani <himanshu.madhani@qlogic.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>

1 files changed, 161 insertions, 44 deletions

diff --git a/drivers/scsi/qla2xxx/qla_target.c b/drivers/scsi/qla2xxx/qla_target.c
index 5ef9d4c21e38..2b218b6e9268 100644
--- a/drivers/scsi/qla2xxx/qla_target.c
+++ b/drivers/scsi/qla2xxx/qla_target.c
@@ -118,10 +118,13 @@ static void qlt_send_notify_ack(struct scsi_qla_host *vha,
 	struct imm_ntfy_from_isp *ntfy,
 	uint32_t add_flags, uint16_t resp_code, int resp_code_valid,
 	uint16_t srr_flags, uint16_t srr_reject_code, uint8_t srr_explan);
+static void qlt_send_term_imm_notif(struct scsi_qla_host *vha,
+	struct imm_ntfy_from_isp *imm, int ha_locked);
 /*
  * Global Variables
  */
 static struct kmem_cache *qla_tgt_mgmt_cmd_cachep;
+static struct kmem_cache *qla_tgt_plogi_cachep;
 static mempool_t *qla_tgt_mgmt_cmd_mempool;
 static struct workqueue_struct *qla_tgt_wq;
 static DEFINE_MUTEX(qla_tgt_mutex);
@@ -389,6 +392,85 @@ void qlt_response_pkt_all_vps(struct scsi_qla_host *vha, response_t *pkt)
 
 }
 
+/*
+ * All qlt_plogi_ack_t operations are protected by hardware_lock
+ */
+
+/*
+ * This is a zero-base ref-counting solution, since hardware_lock
+ * guarantees that ref_count is not modified concurrently.
+ * Upon successful return content of iocb is undefined
+ */
+static qlt_plogi_ack_t *
+qlt_plogi_ack_find_add(struct scsi_qla_host *vha, port_id_t *id,
+		       struct imm_ntfy_from_isp *iocb)
+{
+	qlt_plogi_ack_t *pla;
+
+	list_for_each_entry(pla, &vha->plogi_ack_list, list) {
+		if (pla->id.b24 == id->b24) {
+			qlt_send_term_imm_notif(vha, &pla->iocb, 1);
+			pla->iocb = *iocb;
+			return pla;
+		}
+	}
+
+	pla = kmem_cache_zalloc(qla_tgt_plogi_cachep, GFP_ATOMIC);
+	if (!pla) {
+		ql_dbg(ql_dbg_async, vha, 0x5088,
+		       "qla_target(%d): Allocation of plogi_ack failed\n",
+		       vha->vp_idx);
+		return NULL;
+	}
+
+	pla->iocb = *iocb;
+	pla->id = *id;
+	list_add_tail(&pla->list, &vha->plogi_ack_list);
+
+	return pla;
+}
+
+static void qlt_plogi_ack_unref(struct scsi_qla_host *vha, qlt_plogi_ack_t *pla)
+{
+	BUG_ON(!pla->ref_count);
+	pla->ref_count--;
+
+	if (pla->ref_count)
+		return;
+
+	ql_dbg(ql_dbg_async, vha, 0x5089,
+	    "Sending PLOGI ACK to wwn %8phC s_id %02x:%02x:%02x loop_id %#04x"
+	    " exch %#x ox_id %#x\n", pla->iocb.u.isp24.port_name,
+	    pla->iocb.u.isp24.port_id[2], pla->iocb.u.isp24.port_id[1],
+	    pla->iocb.u.isp24.port_id[0],
+	    le16_to_cpu(pla->iocb.u.isp24.nport_handle),
+	    pla->iocb.u.isp24.exchange_address, pla->iocb.ox_id);
+	qlt_send_notify_ack(vha, &pla->iocb, 0, 0, 0, 0, 0, 0);
+
+	list_del(&pla->list);
+	kmem_cache_free(qla_tgt_plogi_cachep, pla);
+}
+
+static void
+qlt_plogi_ack_link(struct scsi_qla_host *vha, qlt_plogi_ack_t *pla,
+    struct qla_tgt_sess *sess, qlt_plogi_link_t link)
+{
+	/* Inc ref_count first because link might already be pointing at pla */
+	pla->ref_count++;
+
+	if (sess->plogi_link[link])
+		qlt_plogi_ack_unref(vha, sess->plogi_link[link]);
+
+	ql_dbg(ql_dbg_tgt_mgt, vha, 0xf097,
+	    "Linking sess %p [%d] wwn %8phC with PLOGI ACK to wwn %8phC"
+	    " s_id %02x:%02x:%02x, ref=%d\n", sess, link, sess->port_name,
+	    pla->iocb.u.isp24.port_name, pla->iocb.u.isp24.port_id[2],
+	    pla->iocb.u.isp24.port_id[1], pla->iocb.u.isp24.port_id[0],
+	    pla->ref_count);
+
+	sess->plogi_link[link] = pla;
+}
+
 typedef struct {
 	/* These fields must be initialized by the caller */
 	port_id_t id;
@@ -429,10 +511,10 @@ qlt_send_first_logo(struct scsi_qla_host *vha, qlt_port_logo_t *logo)
 	list_del(&logo->list);
 	mutex_unlock(&vha->vha_tgt.tgt_mutex);
 
-	dev_info(&vha->hw->pdev->dev,
-		 "Finished LOGO to %02x:%02x:%02x, dropped %d cmds, res = %#x\n",
-		 logo->id.b.domain, logo->id.b.area, logo->id.b.al_pa,
-		 logo->cmd_count, res);
+	ql_dbg(ql_dbg_tgt_mgt, vha, 0xf098,
+	    "Finished LOGO to %02x:%02x:%02x, dropped %d cmds, res = %#x\n",
+	    logo->id.b.domain, logo->id.b.area, logo->id.b.al_pa,
+	    logo->cmd_count, res);
 }
 
 static void qlt_free_session_done(struct work_struct *work)
@@ -448,11 +530,11 @@ static void qlt_free_session_done(struct work_struct *work)
 
 	ql_dbg(ql_dbg_tgt_mgt, vha, 0xf084,
 		"%s: se_sess %p / sess %p from port %8phC loop_id %#04x"
-		" s_id %02x:%02x:%02x logout %d keep %d plogi %d els_logo %d\n",
+		" s_id %02x:%02x:%02x logout %d keep %d els_logo %d\n",
 		__func__, sess->se_sess, sess, sess->port_name, sess->loop_id,
 		sess->s_id.b.domain, sess->s_id.b.area, sess->s_id.b.al_pa,
 		sess->logout_on_delete, sess->keep_nport_handle,
-		sess->plogi_ack_needed, sess->send_els_logo);
+		sess->send_els_logo);
 
 	BUG_ON(!tgt);
 
@@ -508,9 +590,34 @@ static void qlt_free_session_done(struct work_struct *work)
 
 	spin_lock_irqsave(&ha->hardware_lock, flags);
 
-	if (sess->plogi_ack_needed)
-		qlt_send_notify_ack(vha, &sess->tm_iocb,
-				    0, 0, 0, 0, 0, 0);
+	{
+		qlt_plogi_ack_t *own =
+		    sess->plogi_link[QLT_PLOGI_LINK_SAME_WWN];
+		qlt_plogi_ack_t *con =
+		    sess->plogi_link[QLT_PLOGI_LINK_CONFLICT];
+
+		if (con) {
+			ql_dbg(ql_dbg_tgt_mgt, vha, 0xf099,
+			    "se_sess %p / sess %p port %8phC is gone,"
+			    " %s (ref=%d), releasing PLOGI for %8phC (ref=%d)\n",
+			    sess->se_sess, sess, sess->port_name,
+			    own ? "releasing own PLOGI" :
+			    "no own PLOGI pending",
+			    own ? own->ref_count : -1,
+			    con->iocb.u.isp24.port_name, con->ref_count);
+			qlt_plogi_ack_unref(vha, con);
+		} else {
+			ql_dbg(ql_dbg_tgt_mgt, vha, 0xf09a,
+			    "se_sess %p / sess %p port %8phC is gone, %s (ref=%d)\n",
+			    sess->se_sess, sess, sess->port_name,
+			    own ? "releasing own PLOGI" :
+			    "no own PLOGI pending",
+			    own ? own->ref_count : -1);
+		}
+
+		if (own)
+			qlt_plogi_ack_unref(vha, own);
+	}
 
 	list_del(&sess->sess_list_entry);
 
@@ -4092,15 +4199,6 @@ void qlt_logo_completion_handler(fc_port_t *fcport, int rc)
 	}
 }
 
-static void qlt_swap_imm_ntfy_iocb(struct imm_ntfy_from_isp *a,
-    struct imm_ntfy_from_isp *b)
-{
-	struct imm_ntfy_from_isp tmp;
-	memcpy(&tmp, a, sizeof(struct imm_ntfy_from_isp));
-	memcpy(a, b, sizeof(struct imm_ntfy_from_isp));
-	memcpy(b, &tmp, sizeof(struct imm_ntfy_from_isp));
-}
-
 /*
 * ha->hardware_lock supposed to be held on entry (to protect tgt->sess_list)
 *
@@ -4110,11 +4208,13 @@ static void qlt_swap_imm_ntfy_iocb(struct imm_ntfy_from_isp *a,
 */
 static struct qla_tgt_sess *
 qlt_find_sess_invalidate_other(struct qla_tgt *tgt, uint64_t wwn,
-    port_id_t port_id, uint16_t loop_id)
+    port_id_t port_id, uint16_t loop_id, struct qla_tgt_sess **conflict_sess)
 {
 	struct qla_tgt_sess *sess = NULL, *other_sess;
 	uint64_t other_wwn;
 
+	*conflict_sess = NULL;
+
 	list_for_each_entry(other_sess, &tgt->sess_list, sess_list_entry) {
 
 		other_wwn = wwn_to_u64(other_sess->port_name);
@@ -4142,9 +4242,10 @@ qlt_find_sess_invalidate_other(struct qla_tgt *tgt, uint64_t wwn,
 			} else {
 				/*
 				 * Another wwn used to have our s_id/loop_id
-				 * combo - kill the session, but don't log out
+				 * kill the session, but don't free the loop_id
 				 */
-				sess->logout_on_delete = 0;
+				other_sess->keep_nport_handle = 1;
+				*conflict_sess = other_sess;
 				qlt_schedule_sess_for_deletion(other_sess,
 				    true);
 			}
@@ -4206,12 +4307,13 @@ static int qlt_24xx_handle_els(struct scsi_qla_host *vha,
 {
 	struct qla_tgt *tgt = vha->vha_tgt.qla_tgt;
 	struct qla_hw_data *ha = vha->hw;
-	struct qla_tgt_sess *sess = NULL;
+	struct qla_tgt_sess *sess = NULL, *conflict_sess = NULL;
 	uint64_t wwn;
 	port_id_t port_id;
 	uint16_t loop_id;
 	uint16_t wd3_lo;
 	int res = 0;
+	qlt_plogi_ack_t *pla;
 
 	wwn = wwn_to_u64(iocb->u.isp24.port_name);
 
@@ -4237,25 +4339,15 @@ static int qlt_24xx_handle_els(struct scsi_qla_host *vha,
 
 		if (wwn)
 			sess = qlt_find_sess_invalidate_other(tgt, wwn,
-			    port_id, loop_id);
+			    port_id, loop_id, &conflict_sess);
 
-		if (!sess || IS_SW_RESV_ADDR(sess->s_id)) {
+		if (IS_SW_RESV_ADDR(port_id) || (!sess && !conflict_sess)) {
 			res = 1;
 			break;
 		}
 
-		if (sess->plogi_ack_needed) {
-			/*
-			 * Initiator sent another PLOGI before last PLOGI could
-			 * finish. Swap plogi iocbs and terminate old one
-			 * without acking, new one will get acked when session
-			 * deletion completes.
-			 */
-			ql_log(ql_log_warn, sess->vha, 0xf094,
-			    "sess %p received double plogi.\n", sess);
-
-			qlt_swap_imm_ntfy_iocb(iocb, &sess->tm_iocb);
-
+		pla = qlt_plogi_ack_find_add(vha, &port_id, iocb);
+		if (!pla) {
 			qlt_send_term_imm_notif(vha, iocb, 1);
 
 			res = 0;
@@ -4264,13 +4356,14 @@ static int qlt_24xx_handle_els(struct scsi_qla_host *vha,
 
 		res = 0;
 
-		/*
-		 * Save immediate Notif IOCB for Ack when sess is done
-		 * and being deleted.
-		 */
-		memcpy(&sess->tm_iocb, iocb, sizeof(sess->tm_iocb));
-		sess->plogi_ack_needed  = 1;
+		if (conflict_sess)
+			qlt_plogi_ack_link(vha, pla, conflict_sess,
+			    QLT_PLOGI_LINK_CONFLICT);
 
+		if (!sess)
+			break;
+
+		qlt_plogi_ack_link(vha, pla, sess, QLT_PLOGI_LINK_SAME_WWN);
 		 /*
 		  * Under normal circumstances we want to release nport handle
 		  * during LOGO process to avoid nport handle leaks inside FW.
@@ -4299,7 +4392,16 @@ static int qlt_24xx_handle_els(struct scsi_qla_host *vha,
 
 		if (wwn)
 			sess = qlt_find_sess_invalidate_other(tgt, wwn, port_id,
-			    loop_id);
+			    loop_id, &conflict_sess);
+
+		if (conflict_sess) {
+			ql_dbg(ql_dbg_tgt_mgt, vha, 0xf09b,
+			    "PRLI with conflicting sess %p port %8phC\n",
+			    conflict_sess, conflict_sess->port_name);
+			qlt_send_term_imm_notif(vha, iocb, 1);
+			res = 0;
+			break;
+		}
 
 		if (sess != NULL) {
 			if (sess->deleted) {
@@ -6615,13 +6717,25 @@ int __init qlt_init(void)
 		return -ENOMEM;
 	}
 
+	qla_tgt_plogi_cachep = kmem_cache_create("qla_tgt_plogi_cachep",
+						 sizeof(qlt_plogi_ack_t),
+						 __alignof__(qlt_plogi_ack_t),
+						 0, NULL);
+
+	if (!qla_tgt_plogi_cachep) {
+		ql_log(ql_log_fatal, NULL, 0xe06d,
+		    "kmem_cache_create for qla_tgt_plogi_cachep failed\n");
+		ret = -ENOMEM;
+		goto out_mgmt_cmd_cachep;
+	}
+
 	qla_tgt_mgmt_cmd_mempool = mempool_create(25, mempool_alloc_slab,
 	    mempool_free_slab, qla_tgt_mgmt_cmd_cachep);
 	if (!qla_tgt_mgmt_cmd_mempool) {
 		ql_log(ql_log_fatal, NULL, 0xe06e,
 		    "mempool_create for qla_tgt_mgmt_cmd_mempool failed\n");
 		ret = -ENOMEM;
-		goto out_mgmt_cmd_cachep;
+		goto out_plogi_cachep;
 	}
 
 	qla_tgt_wq = alloc_workqueue("qla_tgt_wq", 0, 0);
@@ -6638,6 +6752,8 @@ int __init qlt_init(void)
 
 out_cmd_mempool:
 	mempool_destroy(qla_tgt_mgmt_cmd_mempool);
+out_plogi_cachep:
+	kmem_cache_destroy(qla_tgt_plogi_cachep);
 out_mgmt_cmd_cachep:
 	kmem_cache_destroy(qla_tgt_mgmt_cmd_cachep);
 	return ret;
@@ -6650,5 +6766,6 @@ void qlt_exit(void)
 
 	destroy_workqueue(qla_tgt_wq);
 	mempool_destroy(qla_tgt_mgmt_cmd_mempool);
+	kmem_cache_destroy(qla_tgt_plogi_cachep);
 	kmem_cache_destroy(qla_tgt_mgmt_cmd_cachep);
 }