bnxt_en: Workaround occasional TX timeout on 57500 A0.

Hardware can sometimes not generate NQ MSIX with a single pending
CP ring entry.  This seems to always happen at the last entry of
the CP ring before it wraps.  Add logic to check all the CP rings for
pending entries without the CP ring consumer index advancing.  Calling
HWRM_DBG_RING_INFO_GET to read the context of the CP ring will flush
out the NQ entry and MSIX.
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/broadcom/bnxt/bnxt.c

@@ -8714,6 +8714,26 @@ static int bnxt_set_features(struct net_device *dev, netdev_features_t features)
 	return rc;
 }
 
+static int bnxt_dbg_hwrm_ring_info_get(struct bnxt *bp, u8 ring_type,
+				       u32 ring_id, u32 *prod, u32 *cons)
+{
+	struct hwrm_dbg_ring_info_get_output *resp = bp->hwrm_cmd_resp_addr;
+	struct hwrm_dbg_ring_info_get_input req = {0};
+	int rc;
+
+	bnxt_hwrm_cmd_hdr_init(bp, &req, HWRM_DBG_RING_INFO_GET, -1, -1);
+	req.ring_type = ring_type;
+	req.fw_ring_id = cpu_to_le32(ring_id);
+	mutex_lock(&bp->hwrm_cmd_lock);
+	rc = _hwrm_send_message(bp, &req, sizeof(req), HWRM_CMD_TIMEOUT);
+	if (!rc) {
+		*prod = le32_to_cpu(resp->producer_index);
+		*cons = le32_to_cpu(resp->consumer_index);
+	}
+	mutex_unlock(&bp->hwrm_cmd_lock);
+	return rc;
+}
+
 static void bnxt_dump_tx_sw_state(struct bnxt_napi *bnapi)
 {
 	struct bnxt_tx_ring_info *txr = bnapi->tx_ring;
@@ -8821,6 +8841,11 @@ static void bnxt_timer(struct timer_list *t)
 			bnxt_queue_sp_work(bp);
 		}
 	}
+
+	if ((bp->flags & BNXT_FLAG_CHIP_P5) && netif_carrier_ok(dev)) {
+		set_bit(BNXT_RING_COAL_NOW_SP_EVENT, &bp->sp_event);
+		bnxt_queue_sp_work(bp);
+	}
 bnxt_restart_timer:
 	mod_timer(&bp->timer, jiffies + bp->current_interval);
 }
@@ -8851,6 +8876,43 @@ static void bnxt_reset(struct bnxt *bp, bool silent)
 	bnxt_rtnl_unlock_sp(bp);
 }
 
+static void bnxt_chk_missed_irq(struct bnxt *bp)
+{
+	int i;
+
+	if (!(bp->flags & BNXT_FLAG_CHIP_P5))
+		return;
+
+	for (i = 0; i < bp->cp_nr_rings; i++) {
+		struct bnxt_napi *bnapi = bp->bnapi[i];
+		struct bnxt_cp_ring_info *cpr;
+		u32 fw_ring_id;
+		int j;
+
+		if (!bnapi)
+			continue;
+
+		cpr = &bnapi->cp_ring;
+		for (j = 0; j < 2; j++) {
+			struct bnxt_cp_ring_info *cpr2 = cpr->cp_ring_arr[j];
+			u32 val[2];
+
+			if (!cpr2 || cpr2->has_more_work ||
+			    !bnxt_has_work(bp, cpr2))
+				continue;
+
+			if (cpr2->cp_raw_cons != cpr2->last_cp_raw_cons) {
+				cpr2->last_cp_raw_cons = cpr2->cp_raw_cons;
+				continue;
+			}
+			fw_ring_id = cpr2->cp_ring_struct.fw_ring_id;
+			bnxt_dbg_hwrm_ring_info_get(bp,
+				DBG_RING_INFO_GET_REQ_RING_TYPE_L2_CMPL,
+				fw_ring_id, &val[0], &val[1]);
+		}
+	}
+}
+
 static void bnxt_cfg_ntp_filters(struct bnxt *);
 
 static void bnxt_sp_task(struct work_struct *work)
@@ -8930,6 +8992,9 @@ static void bnxt_sp_task(struct work_struct *work)
 	if (test_and_clear_bit(BNXT_FLOW_STATS_SP_EVENT, &bp->sp_event))
 		bnxt_tc_flow_stats_work(bp);
 
+	if (test_and_clear_bit(BNXT_RING_COAL_NOW_SP_EVENT, &bp->sp_event))
+		bnxt_chk_missed_irq(bp);
+
 	/* These functions below will clear BNXT_STATE_IN_SP_TASK.  They
 	 * must be the last functions to be called before exiting.
 	 */

drivers/net/ethernet/broadcom/bnxt/bnxt.h

@@ -798,6 +798,8 @@ struct bnxt_cp_ring_info {
 	u8			had_work_done:1;
 	u8			has_more_work:1;
 
+	u32			last_cp_raw_cons;
+
 	struct bnxt_coal	rx_ring_coal;
 	u64			rx_packets;
 	u64			rx_bytes;
@@ -1527,6 +1529,7 @@ struct bnxt {
 #define BNXT_LINK_SPEED_CHNG_SP_EVENT	14
 #define BNXT_FLOW_STATS_SP_EVENT	15
 #define BNXT_UPDATE_PHY_SP_EVENT	16
+#define BNXT_RING_COAL_NOW_SP_EVENT	17
 
 	struct bnxt_hw_resc	hw_resc;
 	struct bnxt_pf_info	pf;