Re: [PATCH net-next 2/3] net: mscc: ocelot: Implement port policers via tc command

Hi Andrew,

The 05/02/2019 14:32, Andrew Lunn wrote:

External E-Mail


Hi Joergen

quoted

+
+#define MSCC_RC(expr)				\
+	do {					\
+		int __rc__ = (expr);		\
+		if (__rc__ < 0)			\
+			return __rc__;		\
+	}					\
+	while (0)

I'm sure checkpatch warned about this. A return inside a macros is a
bad idea. I inherited code doing this, and broke it when adding
locking, because it was not obvious there was a return.

I saw the warning but I assumed that it wasn't a problem in this small context.
The macro will be removed in v2.

quoted

+
+/* The following two functions do the same as in iproute2 */
+#define TIME_UNITS_PER_SEC	1000000
+static unsigned int tc_core_tick2time(unsigned int tick)
+{
+	return (tick * (u32)PSCHED_TICKS2NS(1)) / 1000;
+}
+
+static unsigned int tc_calc_xmitsize(u64 rate, unsigned int ticks)
+{
+	return div_u64(rate * tc_core_tick2time(ticks), TIME_UNITS_PER_SEC);
+}

Should these but put somewhere others can use them?

It would be nice to put them in a more public place, but I am in doubt where to
put them and what to call them.

Maybe they belong in the new file: include/net/tc_act/tc_police.h.
Would that be ok?

quoted

+
+enum mscc_qos_rate_mode {
+	MSCC_QOS_RATE_MODE_DISABLED, /* Policer/shaper disabled */
+	MSCC_QOS_RATE_MODE_LINE, /* Measure line rate in kbps incl. IPG */
+	MSCC_QOS_RATE_MODE_DATA, /* Measures data rate in kbps excl. IPG */
+	MSCC_QOS_RATE_MODE_FRAME, /* Measures frame rate in fps */
+	__MSCC_QOS_RATE_MODE_END,
+	NUM_MSCC_QOS_RATE_MODE = __MSCC_QOS_RATE_MODE_END,
+	MSCC_QOS_RATE_MODE_MAX = __MSCC_QOS_RATE_MODE_END - 1,
+};
+
+/* Round x divided by y to nearest integer. x and y are integers */
+#define MSCC_ROUNDING_DIVISION(x, y) (((x) + ((y) / 2)) / (y))

linux/kernel.h defines DIV_ROUND_UP(). Maybe add DIV_ROUND_DOWN()?

This macro is currently not used and I will remove it in v2.

quoted

+
+/* Round x divided by y to nearest higher integer. x and y are integers */
+#define MSCC_DIV_ROUND_UP(x, y) (((x) + (y) - 1) / (y))

DIV_ROUND_UP() ?

I will use DIV_ROUND_UP() in v2.

quoted

+	/* Limit to maximum values */
+	pir = min_t(u32, GENMASK(15, 0), pir);
+	cir = min_t(u32, GENMASK(15, 0), cir);
+	pbs = min_t(u32, pbs_max, pbs);
+	cbs = min_t(u32, cbs_max, cbs);

If it does need to limit, maybe return -EOPNOTSUPP?

It seems fine to return -EOPBITSUPP here.
I will do that in v2.

quoted

+int ocelot_port_policer_add(struct ocelot_port *port,
+			    struct tcf_police *p)
+{
+	struct ocelot *ocelot = port->ocelot;
+	struct qos_policer_conf pp;
+
+	if (!p)
+		return -EINVAL;
+
+	netdev_dbg(port->dev,
+		   "result %d ewma_rate %u burst %lld mtu %u mtu_pktoks %lld\n",
+		   p->params->tcfp_result,
+		   p->params->tcfp_ewma_rate,
+		   p->params->tcfp_burst,
+		   p->params->tcfp_mtu,
+		   p->params->tcfp_mtu_ptoks);
+
+	if (p->params->rate_present)
+		netdev_dbg(port->dev,
+			   "rate: rate %llu mult %u over %u link %u shift %u\n",
+			   p->params->rate.rate_bytes_ps,
+			   p->params->rate.mult,
+			   p->params->rate.overhead,
+			   p->params->rate.linklayer,
+			   p->params->rate.shift);
+
+	if (p->params->peak_present)
+		netdev_dbg(port->dev,
+			   "peak: rate %llu mult %u over %u link %u shift %u\n",
+			   p->params->peak.rate_bytes_ps,
+			   p->params->peak.mult,
+			   p->params->peak.overhead,
+			   p->params->peak.linklayer,
+			   p->params->peak.shift);
+
+	memset(&pp, 0, sizeof(pp));

Rather than memset, you can do:

	struct qos_policer_conf pp = { 0 };

I will do as you suggest in v2.

	Andrew

-- 
Joergen Andreasen, Microchip