IB/hfi1: Convert PortXmitWait/PortVLXmitWait counters to flit times

HFI's counters SendWaitCnt and SendWaitVlCnt are in units
of TXE cycle time (at 805MHz). OPA counters PortXmitWait and
PortVLXmtWait are in units of flit times.
Convert the counter values to flit units using following
conversion formula:

PortXmitWait =
	SendWaitCnt * 2 * (4 /link_width) * (25 Gbps /link_speed)
PortVLXmitWait =
	SendWaitVLCnt * 2 * (4 /link_width) * (25 Gbps /link_speed)

At link up or downgrade events, the link width can change. To ensure
accurate counter calculations, sample the counters after the events,
during counter requests, and then aggregate the OPA counters.

Reviewed-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Kamenee Arumugam <kamenee.arumugam@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>

drivers/infiniband/hw/hfi1/chip.c

@@ -1083,6 +1083,7 @@ static int qos_rmt_entries(struct hfi1_devdata *dd, unsigned int *mp,
 static void clear_full_mgmt_pkey(struct hfi1_pportdata *ppd);
 static int wait_link_transfer_active(struct hfi1_devdata *dd, int wait_ms);
 static void clear_rsm_rule(struct hfi1_devdata *dd, u8 rule_index);
+static void update_xmit_counters(struct hfi1_pportdata *ppd, u16 link_width);
 
 /*
  * Error interrupt table entry.  This is used as input to the interrupt
@@ -6905,6 +6906,32 @@ void handle_freeze(struct work_struct *work)
 	/* no longer frozen */
 }
 
+/**
+ * update_xmit_counters - update PortXmitWait/PortVlXmitWait
+ * counters.
+ * @ppd: info of physical Hfi port
+ * @link_width: new link width after link up or downgrade
+ *
+ * Update the PortXmitWait and PortVlXmitWait counters after
+ * a link up or downgrade event to reflect a link width change.
+ */
+static void update_xmit_counters(struct hfi1_pportdata *ppd, u16 link_width)
+{
+	int i;
+	u16 tx_width;
+	u16 link_speed;
+
+	tx_width = tx_link_width(link_width);
+	link_speed = get_link_speed(ppd->link_speed_active);
+
+	/*
+	 * There are C_VL_COUNT number of PortVLXmitWait counters.
+	 * Adding 1 to C_VL_COUNT to include the PortXmitWait counter.
+	 */
+	for (i = 0; i < C_VL_COUNT + 1; i++)
+		get_xmit_wait_counters(ppd, tx_width, link_speed, i);
+}
+
 /*
  * Handle a link up interrupt from the 8051.
  *
@@ -7526,18 +7553,29 @@ void handle_verify_cap(struct work_struct *work)
 	set_link_state(ppd, HLS_GOING_UP);
 }
 
-/*
- * Apply the link width downgrade enabled policy against the current active
- * link widths.
+/**
+ * apply_link_downgrade_policy - Apply the link width downgrade enabled
+ * policy against the current active link widths.
+ * @ppd: info of physical Hfi port
+ * @refresh_widths: True indicates link downgrade event
+ * @return: True indicates a successful link downgrade. False indicates
+ *	    link downgrade event failed and the link will bounce back to
+ *	    default link width.
  *
- * Called when the enabled policy changes or the active link widths change.
+ * Called when the enabled policy changes or the active link widths
+ * change.
+ * Refresh_widths indicates that a link downgrade occurred. The
+ * link_downgraded variable is set by refresh_widths and
+ * determines the success/failure of the policy application.
  */
-void apply_link_downgrade_policy(struct hfi1_pportdata *ppd, int refresh_widths)
+bool apply_link_downgrade_policy(struct hfi1_pportdata *ppd,
+				 bool refresh_widths)
 {
 	int do_bounce = 0;
 	int tries;
 	u16 lwde;
 	u16 tx, rx;
+	bool link_downgraded = refresh_widths;
 
 	/* use the hls lock to avoid a race with actual link up */
 	tries = 0;
@@ -7571,6 +7609,7 @@ retry:
 	    ppd->link_width_downgrade_rx_active == 0) {
 		/* the 8051 reported a dead link as a downgrade */
 		dd_dev_err(ppd->dd, "Link downgrade is really a link down, ignoring\n");
+		link_downgraded = false;
 	} else if (lwde == 0) {
 		/* downgrade is disabled */
 
@@ -7587,6 +7626,7 @@ retry:
 				   ppd->link_width_downgrade_tx_active,
 				   ppd->link_width_downgrade_rx_active);
 			do_bounce = 1;
+			link_downgraded = false;
 		}
 	} else if ((lwde & ppd->link_width_downgrade_tx_active) == 0 ||
 		   (lwde & ppd->link_width_downgrade_rx_active) == 0) {
@@ -7598,6 +7638,7 @@ retry:
 			   lwde, ppd->link_width_downgrade_tx_active,
 			   ppd->link_width_downgrade_rx_active);
 		do_bounce = 1;
+		link_downgraded = false;
 	}
 
 done:
@@ -7609,6 +7650,8 @@ done:
 		set_link_state(ppd, HLS_DN_OFFLINE);
 		start_link(ppd);
 	}
+
+	return link_downgraded;
 }
 
 /*
@@ -7622,7 +7665,8 @@ void handle_link_downgrade(struct work_struct *work)
 							link_downgrade_work);
 
 	dd_dev_info(ppd->dd, "8051: Link width downgrade\n");
-	apply_link_downgrade_policy(ppd, 1);
+	if (apply_link_downgrade_policy(ppd, true))
+		update_xmit_counters(ppd, ppd->link_width_downgrade_tx_active);
 }
 
 static char *dcc_err_string(char *buf, int buf_len, u64 flags)
@@ -10597,6 +10641,14 @@ int set_link_state(struct hfi1_pportdata *ppd, u32 state)
 		add_rcvctrl(dd, RCV_CTRL_RCV_PORT_ENABLE_SMASK);
 
 		handle_linkup_change(dd, 1);
+
+		/*
+		 * After link up, a new link width will have been set.
+		 * Update the xmit counters with regards to the new
+		 * link width.
+		 */
+		update_xmit_counters(ppd, ppd->link_width_active);
+
 		ppd->host_link_state = HLS_UP_INIT;
 		update_statusp(ppd, IB_PORT_INIT);
 		break;

drivers/infiniband/hw/hfi1/chip.h

@@ -736,8 +736,8 @@ int read_8051_config(struct hfi1_devdata *, u8, u8, u32 *);
 int start_link(struct hfi1_pportdata *ppd);
 int bringup_serdes(struct hfi1_pportdata *ppd);
 void set_intr_state(struct hfi1_devdata *dd, u32 enable);
-void apply_link_downgrade_policy(struct hfi1_pportdata *ppd,
-				 int refresh_widths);
+bool apply_link_downgrade_policy(struct hfi1_pportdata *ppd,
+				 bool refresh_widths);
 void update_usrhead(struct hfi1_ctxtdata *rcd, u32 hd, u32 updegr, u32 egrhd,
 		    u32 intr_adjust, u32 npkts);
 int stop_drain_data_vls(struct hfi1_devdata *dd);

drivers/infiniband/hw/hfi1/hfi.h

@@ -858,6 +858,13 @@ struct hfi1_pportdata {
 	struct work_struct linkstate_active_work;
 	/* Does this port need to prescan for FECNs */
 	bool cc_prescan;
+	/*
+	 * Sample sendWaitCnt & sendWaitVlCnt during link transition
+	 * and counter request.
+	 */
+	u64 port_vl_xmit_wait_last[C_VL_COUNT + 1];
+	u16 prev_link_width;
+	u64 vl_xmit_flit_cnt[C_VL_COUNT + 1];
 };
 
 typedef int (*rhf_rcv_function_ptr)(struct hfi1_packet *packet);

drivers/infiniband/hw/hfi1/init.c

@@ -637,6 +637,15 @@ void hfi1_init_pportdata(struct pci_dev *pdev, struct hfi1_pportdata *ppd,
 	ppd->dd = dd;
 	ppd->hw_pidx = hw_pidx;
 	ppd->port = port; /* IB port number, not index */
+	ppd->prev_link_width = LINK_WIDTH_DEFAULT;
+	/*
+	 * There are C_VL_COUNT number of PortVLXmitWait counters.
+	 * Adding 1 to C_VL_COUNT to include the PortXmitWait counter.
+	 */
+	for (i = 0; i < C_VL_COUNT + 1; i++) {
+		ppd->port_vl_xmit_wait_last[i] = 0;
+		ppd->vl_xmit_flit_cnt[i] = 0;
+	}
 
 	default_pkey_idx = 1;
 

drivers/infiniband/hw/hfi1/mad.c

@@ -2649,6 +2649,79 @@ static void a0_portstatus(struct hfi1_pportdata *ppd,
 	}
 }
 
+/**
+ * tx_link_width - convert link width bitmask to integer
+ * value representing actual link width.
+ * @link_width: width of active link
+ * @return: return index of the bit set in link_width var
+ *
+ * The function convert and return the index of bit set
+ * that indicate the current link width.
+ */
+u16 tx_link_width(u16 link_width)
+{
+	int n = LINK_WIDTH_DEFAULT;
+	u16 tx_width = n;
+
+	while (link_width && n) {
+		if (link_width & (1 << (n - 1))) {
+			tx_width = n;
+			break;
+		}
+		n--;
+	}
+
+	return tx_width;
+}
+
+/**
+ * get_xmit_wait_counters - Convert HFI 's SendWaitCnt/SendWaitVlCnt
+ * counter in unit of TXE cycle times to flit times.
+ * @ppd: info of physical Hfi port
+ * @link_width: width of active link
+ * @link_speed: speed of active link
+ * @vl: represent VL0-VL7, VL15 for PortVLXmitWait counters request
+ * and if vl value is C_VL_COUNT, it represent SendWaitCnt
+ * counter request
+ * @return: return SendWaitCnt/SendWaitVlCnt counter value per vl.
+ *
+ * Convert SendWaitCnt/SendWaitVlCnt counter from TXE cycle times to
+ * flit times. Call this function to samples these counters. This
+ * function will calculate for previous state transition and update
+ * current state at end of function using ppd->prev_link_width and
+ * ppd->port_vl_xmit_wait_last to port_vl_xmit_wait_curr and link_width.
+ */
+u64 get_xmit_wait_counters(struct hfi1_pportdata *ppd,
+			   u16 link_width, u16 link_speed, int vl)
+{
+	u64 port_vl_xmit_wait_curr;
+	u64 delta_vl_xmit_wait;
+	u64 xmit_wait_val;
+
+	if (vl > C_VL_COUNT)
+		return  0;
+	if (vl < C_VL_COUNT)
+		port_vl_xmit_wait_curr =
+			read_port_cntr(ppd, C_TX_WAIT_VL, vl);
+	else
+		port_vl_xmit_wait_curr =
+			read_port_cntr(ppd, C_TX_WAIT, CNTR_INVALID_VL);
+
+	xmit_wait_val =
+		port_vl_xmit_wait_curr -
+		ppd->port_vl_xmit_wait_last[vl];
+	delta_vl_xmit_wait =
+		convert_xmit_counter(xmit_wait_val,
+				     ppd->prev_link_width,
+				     link_speed);
+
+	ppd->vl_xmit_flit_cnt[vl] += delta_vl_xmit_wait;
+	ppd->port_vl_xmit_wait_last[vl] = port_vl_xmit_wait_curr;
+	ppd->prev_link_width = link_width;
+
+	return ppd->vl_xmit_flit_cnt[vl];
+}
+
 static int pma_get_opa_portstatus(struct opa_pma_mad *pmp,
 				  struct ib_device *ibdev,
 				  u8 port, u32 *resp_len)
@@ -2668,6 +2741,8 @@ static int pma_get_opa_portstatus(struct opa_pma_mad *pmp,
 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 	int vfi;
 	u64 tmp, tmp2;
+	u16 link_width;
+	u16 link_speed;
 
 	response_data_size = sizeof(struct opa_port_status_rsp) +
 				num_vls * sizeof(struct _vls_pctrs);
@@ -2711,8 +2786,16 @@ static int pma_get_opa_portstatus(struct opa_pma_mad *pmp,
 	rsp->port_multicast_rcv_pkts =
 		cpu_to_be64(read_dev_cntr(dd, C_DC_MC_RCV_PKTS,
 					  CNTR_INVALID_VL));
+	/*
+	 * Convert PortXmitWait counter from TXE cycle times
+	 * to flit times.
+	 */
+	link_width =
+		tx_link_width(ppd->link_width_downgrade_tx_active);
+	link_speed = get_link_speed(ppd->link_speed_active);
 	rsp->port_xmit_wait =
-		cpu_to_be64(read_port_cntr(ppd, C_TX_WAIT, CNTR_INVALID_VL));
+		cpu_to_be64(get_xmit_wait_counters(ppd, link_width,
+						   link_speed, C_VL_COUNT));
 	rsp->port_rcv_fecn =
 		cpu_to_be64(read_dev_cntr(dd, C_DC_RCV_FCN, CNTR_INVALID_VL));
 	rsp->port_rcv_becn =
@@ -2777,10 +2860,14 @@ static int pma_get_opa_portstatus(struct opa_pma_mad *pmp,
 		rsp->vls[vfi].port_vl_xmit_pkts =
 			cpu_to_be64(read_port_cntr(ppd, C_TX_PKT_VL,
 						   idx_from_vl(vl)));
-
+		/*
+		 * Convert PortVlXmitWait counter from TXE cycle
+		 * times to flit times.
+		 */
 		rsp->vls[vfi].port_vl_xmit_wait =
-			cpu_to_be64(read_port_cntr(ppd, C_TX_WAIT_VL,
-						   idx_from_vl(vl)));
+			cpu_to_be64(get_xmit_wait_counters(ppd, link_width,
+							   link_speed,
+							   idx_from_vl(vl)));
 
 		rsp->vls[vfi].port_vl_rcv_fecn =
 			cpu_to_be64(read_dev_cntr(dd, C_DC_RCV_FCN_VL,
@@ -2910,6 +2997,8 @@ static int pma_get_opa_datacounters(struct opa_pma_mad *pmp,
 	unsigned long vl;
 	u32 vl_select_mask;
 	int vfi;
+	u16 link_width;
+	u16 link_speed;
 
 	num_ports = be32_to_cpu(pmp->mad_hdr.attr_mod) >> 24;
 	num_vls = hweight32(be32_to_cpu(req->vl_select_mask));
@@ -2959,8 +3048,16 @@ static int pma_get_opa_datacounters(struct opa_pma_mad *pmp,
 	rsp->link_quality_indicator = cpu_to_be32((u32)lq);
 	pma_get_opa_port_dctrs(ibdev, rsp);
 
+	/*
+	 * Convert PortXmitWait counter from TXE
+	 * cycle times to flit times.
+	 */
+	link_width =
+		tx_link_width(ppd->link_width_downgrade_tx_active);
+	link_speed = get_link_speed(ppd->link_speed_active);
 	rsp->port_xmit_wait =
-		cpu_to_be64(read_port_cntr(ppd, C_TX_WAIT, CNTR_INVALID_VL));
+		cpu_to_be64(get_xmit_wait_counters(ppd, link_width,
+						   link_speed, C_VL_COUNT));
 	rsp->port_rcv_fecn =
 		cpu_to_be64(read_dev_cntr(dd, C_DC_RCV_FCN, CNTR_INVALID_VL));
 	rsp->port_rcv_becn =
@@ -2996,9 +3093,14 @@ static int pma_get_opa_datacounters(struct opa_pma_mad *pmp,
 			cpu_to_be64(read_dev_cntr(dd, C_DC_RX_PKT_VL,
 						  idx_from_vl(vl)));
 
+		/*
+		 * Convert PortVlXmitWait counter from TXE
+		 * cycle times to flit times.
+		 */
 		rsp->vls[vfi].port_vl_xmit_wait =
-			cpu_to_be64(read_port_cntr(ppd, C_TX_WAIT_VL,
-						   idx_from_vl(vl)));
+			cpu_to_be64(get_xmit_wait_counters(ppd, link_width,
+							   link_speed,
+							   idx_from_vl(vl)));
 
 		rsp->vls[vfi].port_vl_rcv_fecn =
 			cpu_to_be64(read_dev_cntr(dd, C_DC_RCV_FCN_VL,
@@ -3416,9 +3518,11 @@ static int pma_set_opa_portstatus(struct opa_pma_mad *pmp,
 	if (counter_select & CS_PORT_MCAST_RCV_PKTS)
 		write_dev_cntr(dd, C_DC_MC_RCV_PKTS, CNTR_INVALID_VL, 0);
 
-	if (counter_select & CS_PORT_XMIT_WAIT)
+	if (counter_select & CS_PORT_XMIT_WAIT) {
 		write_port_cntr(ppd, C_TX_WAIT, CNTR_INVALID_VL, 0);
-
+		ppd->port_vl_xmit_wait_last[C_VL_COUNT] = 0;
+		ppd->vl_xmit_flit_cnt[C_VL_COUNT] = 0;
+	}
 	/* ignore cs_sw_portCongestion for HFIs */
 
 	if (counter_select & CS_PORT_RCV_FECN)
@@ -3491,8 +3595,11 @@ static int pma_set_opa_portstatus(struct opa_pma_mad *pmp,
 		if (counter_select & CS_PORT_RCV_PKTS)
 			write_dev_cntr(dd, C_DC_RX_PKT_VL, idx_from_vl(vl), 0);
 
-		if (counter_select & CS_PORT_XMIT_WAIT)
+		if (counter_select & CS_PORT_XMIT_WAIT) {
 			write_port_cntr(ppd, C_TX_WAIT_VL, idx_from_vl(vl), 0);
+			ppd->port_vl_xmit_wait_last[idx_from_vl(vl)] = 0;
+			ppd->vl_xmit_flit_cnt[idx_from_vl(vl)] = 0;
+		}
 
 		/* sw_port_vl_congestion is 0 for HFIs */
 		if (counter_select & CS_PORT_RCV_FECN)

drivers/infiniband/hw/hfi1/mad.h

@@ -180,6 +180,15 @@ struct opa_mad_notice_attr {
 #define OPA_VLARB_PREEMPT_MATRIX     3
 
 #define IB_PMA_PORT_COUNTERS_CONG       cpu_to_be16(0xFF00)
+#define LINK_SPEED_25G		1
+#define LINK_SPEED_12_5G	2
+#define LINK_WIDTH_DEFAULT	4
+#define DECIMAL_FACTORING	1000
+/*
+ * The default link width is multiplied by 1000
+ * to get accurate value after division.
+ */
+#define FACTOR_LINK_WIDTH	(LINK_WIDTH_DEFAULT * DECIMAL_FACTORING)
 
 struct ib_pma_portcounters_cong {
 	u8 reserved;
@@ -429,5 +438,41 @@ struct sc2vlnt {
 
 void hfi1_event_pkey_change(struct hfi1_devdata *dd, u8 port);
 void hfi1_handle_trap_timer(struct timer_list *t);
-
+u16 tx_link_width(u16 link_width);
+u64 get_xmit_wait_counters(struct hfi1_pportdata *ppd, u16 link_width,
+			   u16 link_speed, int vl);
+/**
+ * get_link_speed - determine whether 12.5G or 25G speed
+ * @link_speed: the speed of active link
+ * @return: Return 2 if link speed identified as 12.5G
+ * or return 1 if link speed is 25G.
+ *
+ * The function indirectly calculate required link speed
+ * value for convert_xmit_counter function. If the link
+ * speed is 25G, the function return as 1 as it is required
+ * by xmit counter conversion formula :-( 25G / link_speed).
+ * This conversion will provide value 1 if current
+ * link speed is 25G or 2 if 12.5G.This is done to avoid
+ * 12.5 float number conversion.
+ */
+static inline u16 get_link_speed(u16 link_speed)
+{
+	return (link_speed == 1) ?
+		 LINK_SPEED_12_5G : LINK_SPEED_25G;
+}
+
+/**
+ * convert_xmit_counter - calculate flit times for given xmit counter
+ * value
+ * @xmit_wait_val: current xmit counter value
+ * @link_width: width of active link
+ * @link_speed: speed of active link
+ * @return: return xmit counter value in flit times.
+ */
+static inline u64 convert_xmit_counter(u64 xmit_wait_val, u16 link_width,
+				       u16 link_speed)
+{
+	return (xmit_wait_val * 2 * (FACTOR_LINK_WIDTH / link_width)
+		 * link_speed) / DECIMAL_FACTORING;
+}
 #endif				/* _HFI1_MAD_H */