scsi: qla2xxx: Fix firmware resource tracking

The storage was not draining I/Os and the work load was not spread out
across different CPUs evenly. This led to firmware resource counters
getting overrun on the busy CPU. This overrun prevented error recovery from
happening in a timely manner.

By switching the counter to atomic, it allows the count to be little more
accurate to prevent the overrun.

Cc: stable@vger.kernel.org
Fixes: da7c21b ("scsi: qla2xxx: Fix command flush during TMF")
Signed-off-by: Quinn Tran <qutran@marvell.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Link: https://lore.kernel.org/r/20230821130045.34850-4-njavali@marvell.com
Reviewed-by: Himanshu Madhani <himanshu.madhani@oracle.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>

Author: Quinn Tran

drivers/scsi/qla2xxx/qla_def.h

@@ -3770,6 +3770,16 @@ struct qla_fw_resources {
 	u16 pad;
 };
 
+struct qla_fw_res {
+	u16      iocb_total;
+	u16      iocb_limit;
+	atomic_t iocb_used;
+
+	u16      exch_total;
+	u16      exch_limit;
+	atomic_t exch_used;
+};
+
 #define QLA_IOCB_PCT_LIMIT 95
 
 struct  qla_buf_pool {
@@ -4830,6 +4840,7 @@ struct qla_hw_data {
 	u8 edif_post_stop_cnt_down;
 	struct qla_vp_map *vp_map;
 	struct qla_nvme_fc_rjt lsrjt;
+	struct qla_fw_res fwres ____cacheline_aligned;
 };
 
 #define RX_ELS_SIZE (roundup(sizeof(struct enode) + ELS_MAX_PAYLOAD, SMP_CACHE_BYTES))

drivers/scsi/qla2xxx/qla_dfs.c

@@ -276,6 +276,16 @@ qla_dfs_fw_resource_cnt_show(struct seq_file *s, void *unused)
 
 		seq_printf(s, "estimate exchange used[%d] high water limit [%d] n",
 			   exch_used, ha->base_qpair->fwres.exch_limit);
+
+		if (ql2xenforce_iocb_limit == 2) {
+			iocbs_used = atomic_read(&ha->fwres.iocb_used);
+			exch_used  = atomic_read(&ha->fwres.exch_used);
+			seq_printf(s, "        estimate iocb2 used [%d] high water limit [%d]\n",
+					iocbs_used, ha->fwres.iocb_limit);
+
+			seq_printf(s, "        estimate exchange2 used[%d] high water limit [%d] \n",
+					exch_used, ha->fwres.exch_limit);
+		}
 	}
 
 	return 0;

drivers/scsi/qla2xxx/qla_init.c

@@ -4214,6 +4214,14 @@ void qla_init_iocb_limit(scsi_qla_host_t *vha)
 			ha->queue_pair_map[i]->fwres.exch_used = 0;
 		}
 	}
+
+	ha->fwres.iocb_total = ha->orig_fw_iocb_count;
+	ha->fwres.iocb_limit = (ha->orig_fw_iocb_count * QLA_IOCB_PCT_LIMIT) / 100;
+	ha->fwres.exch_total = ha->orig_fw_xcb_count;
+	ha->fwres.exch_limit = (ha->orig_fw_xcb_count * QLA_IOCB_PCT_LIMIT) / 100;
+
+	atomic_set(&ha->fwres.iocb_used, 0);
+	atomic_set(&ha->fwres.exch_used, 0);
 }
 
 void qla_adjust_iocb_limit(scsi_qla_host_t *vha)

drivers/scsi/qla2xxx/qla_inline.h

@@ -386,14 +386,15 @@ enum {
 	RESOURCE_IOCB = BIT_0,
 	RESOURCE_EXCH = BIT_1,  /* exchange */
 	RESOURCE_FORCE = BIT_2,
+	RESOURCE_HA = BIT_3,
 };
 
 static inline int
 qla_get_fw_resources(struct qla_qpair *qp, struct iocb_resource *iores)
 {
 	u16 iocbs_used, i;
 	u16 exch_used;
-	struct qla_hw_data *ha = qp->vha->hw;
+	struct qla_hw_data *ha = qp->hw;
 
 	if (!ql2xenforce_iocb_limit) {
 		iores->res_type = RESOURCE_NONE;
@@ -428,15 +429,69 @@ qla_get_fw_resources(struct qla_qpair *qp, struct iocb_resource *iores)
 			return -ENOSPC;
 		}
 	}
+
+	if (ql2xenforce_iocb_limit == 2) {
+		if ((iores->iocb_cnt + atomic_read(&ha->fwres.iocb_used)) >=
+		    ha->fwres.iocb_limit) {
+			iores->res_type = RESOURCE_NONE;
+			return -ENOSPC;
+		}
+
+		if (iores->res_type & RESOURCE_EXCH) {
+			if ((iores->exch_cnt + atomic_read(&ha->fwres.exch_used)) >=
+			    ha->fwres.exch_limit) {
+				iores->res_type = RESOURCE_NONE;
+				return -ENOSPC;
+			}
+		}
+	}
+
 force:
 	qp->fwres.iocbs_used += iores->iocb_cnt;
 	qp->fwres.exch_used += iores->exch_cnt;
+	if (ql2xenforce_iocb_limit == 2) {
+		atomic_add(iores->iocb_cnt, &ha->fwres.iocb_used);
+		atomic_add(iores->exch_cnt, &ha->fwres.exch_used);
+		iores->res_type |= RESOURCE_HA;
+	}
 	return 0;
 }
 
+/*
+ * decrement to zero.  This routine will not decrement below zero
+ * @v:  pointer of type atomic_t
+ * @amount: amount to decrement from v
+ */
+static void qla_atomic_dtz(atomic_t *v, int amount)
+{
+	int c, old, dec;
+
+	c = atomic_read(v);
+	for (;;) {
+		dec = c - amount;
+		if (unlikely(dec < 0))
+			dec = 0;
+
+		old = atomic_cmpxchg((v), c, dec);
+		if (likely(old == c))
+			break;
+		c = old;
+	}
+}
+
 static inline void
 qla_put_fw_resources(struct qla_qpair *qp, struct iocb_resource *iores)
 {
+	struct qla_hw_data *ha = qp->hw;
+
+	if (iores->res_type & RESOURCE_HA) {
+		if (iores->res_type & RESOURCE_IOCB)
+			qla_atomic_dtz(&ha->fwres.iocb_used, iores->iocb_cnt);
+
+		if (iores->res_type & RESOURCE_EXCH)
+			qla_atomic_dtz(&ha->fwres.exch_used, iores->exch_cnt);
+	}
+
 	if (iores->res_type & RESOURCE_IOCB) {
 		if (qp->fwres.iocbs_used >= iores->iocb_cnt) {
 			qp->fwres.iocbs_used -= iores->iocb_cnt;

drivers/scsi/qla2xxx/qla_os.c

@@ -44,10 +44,11 @@ module_param(ql2xfulldump_on_mpifail, int, S_IRUGO | S_IWUSR);
 MODULE_PARM_DESC(ql2xfulldump_on_mpifail,
 		 "Set this to take full dump on MPI hang.");
 
-int ql2xenforce_iocb_limit = 1;
+int ql2xenforce_iocb_limit = 2;
 module_param(ql2xenforce_iocb_limit, int, S_IRUGO | S_IWUSR);
 MODULE_PARM_DESC(ql2xenforce_iocb_limit,
-		 "Enforce IOCB throttling, to avoid FW congestion. (default: 1)");
+		 "Enforce IOCB throttling, to avoid FW congestion. (default: 2) "
+		 "1: track usage per queue, 2: track usage per adapter");
 
 /*
  * CT6 CTX allocation cache