Re: [PATCH v5 net-next 2/4] net: ethernet: ti: am65-cpsw: add mqprio qdisc offload in channel mode

[Roger Quadros]

On 05/10/2023 12:01, Vladimir Oltean wrote:
> On Wed, Sep 27, 2023 at 10:27:39AM +0300, Roger Quadros wrote:
>> diff --git a/drivers/net/ethernet/ti/am65-cpsw-qos.c b/drivers/net/ethernet/ti/am65-cpsw-qos.c
>> index f91137d8e73b..991221d8f148 100644
>> --- a/drivers/net/ethernet/ti/am65-cpsw-qos.c
>> +++ b/drivers/net/ethernet/ti/am65-cpsw-qos.c
>> @@ -16,10 +16,19 @@
>>  #include "cpsw_ale.h"
>>  
>>  #define AM65_CPSW_REG_CTL			0x004
>> +#define AM65_CPSW_P0_REG_TX_PRI_MAP		0x018
>> +#define AM65_CPSW_P0_REG_RX_PRI_MAP		0x020
>> +#define AM65_CPSW_P0_REG_FIFO_STATUS		0x050
>> +#define AM65_CPSW_P0_REG_PRI_CIR(pri)		(0x140 + 4 * (pri))
>> +#define AM65_CPSW_P0_REG_PRI_EIR(pri)		(0x160 + 4 * (pri))
>> +
>>  #define AM65_CPSW_PN_REG_CTL			0x004
>> +#define AM65_CPSW_PN_REG_TX_PRI_MAP		0x018
>> +#define AM65_CPSW_PN_REG_RX_PRI_MAP		0x020
>>  #define AM65_CPSW_PN_REG_FIFO_STATUS		0x050
>>  #define AM65_CPSW_PN_REG_EST_CTL		0x060
>>  #define AM65_CPSW_PN_REG_PRI_CIR(pri)		(0x140 + 4 * (pri))
>> +#define AM65_CPSW_PN_REG_PRI_EIR(pri)		(0x160 + 4 * (pri))
>>  
>>  /* AM65_CPSW_REG_CTL register fields */
>>  #define AM65_CPSW_CTL_EST_EN			BIT(18)
>> @@ -50,12 +59,18 @@
>>  #define AM65_CPSW_FETCH_ALLOW_MSK		GENMASK(7, 0)
>>  #define AM65_CPSW_FETCH_ALLOW_MAX		AM65_CPSW_FETCH_ALLOW_MSK
>>  
>> +#define TO_MBPS(x)	((x) * 8 / 1000000)
>> +
>>  enum timer_act {
>>  	TACT_PROG,		/* need program timer */
>>  	TACT_NEED_STOP,		/* need stop first */
>>  	TACT_SKIP_PROG,		/* just buffer can be updated */
>>  };
>>  
>> +/* number of priority queues per port FIFO */
>> +#define AM65_CPSW_PN_FIFO_PRIO_NUM		8
>> +#define AM65_CPSW_PN_TX_PRI_MAP_DEFAULT		0x76543210
> 
>> +static void am65_cpsw_reset_tc_mqprio(struct net_device *ndev)
>> +{
>> +	struct am65_cpsw_port *port = am65_ndev_to_port(ndev);
>> +	struct am65_cpsw_mqprio *p_mqprio = &port->qos.mqprio;
>> +	struct am65_cpsw_common *common = port->common;
>> +
>> +	p_mqprio->shaper_en = false;
>> +	p_mqprio->max_rate_total = 0;
>> +
>> +	am65_cpsw_tx_pn_shaper_reset(port);
>> +	netdev_reset_tc(ndev);
>> +	netif_set_real_num_tx_queues(ndev, common->tx_ch_num);
> 
> If this never gets changed from the value set by am65_cpsw_nuss_ndo_slave_open(),
> then there is no reason to call netif_set_real_num_tx_queues() here.
> 
>> +
>> +	/* Reset all Queue priorities to 0 */
>> +	writel(0, port->port_base + AM65_CPSW_PN_REG_TX_PRI_MAP);
> 
> To 0 or to AM65_CPSW_PN_TX_PRI_MAP_DEFAULT (which is now unused)?
> Also, do I understand correctly that immediately after boot, TX_PRI_MAP
> will have a different value than after the deletion of the mqprio root qdisc
> (if AM65_CPSW_PN_TX_PRI_MAP_DEFAULT represents the hardware default value)?
> The behavior needs to be identical in the 2 cases, since both represent
> "no mqprio offload".

Good catch. I'll get rid of AM65_CPSW_PN_TX_PRI_MAP_DEFAULT and set this to
0 at init time.

> 
>> +}
>> +
>> +static int am65_cpsw_setup_mqprio(struct net_device *ndev, void *type_data)
>> +{
>> +	struct am65_cpsw_port *port = am65_ndev_to_port(ndev);
>> +	struct am65_cpsw_mqprio *p_mqprio = &port->qos.mqprio;
>> +	struct tc_mqprio_qopt_offload *mqprio = type_data;
>> +	struct am65_cpsw_common *common = port->common;
>> +	struct tc_mqprio_qopt *qopt = &mqprio->qopt;
>> +	int tc, offset, count, ret, prio;
>> +	u8 num_tc = qopt->num_tc;
>> +	u32 tx_prio_map = 0;
>> +	int i;
>> +
>> +	memcpy(&p_mqprio->mqprio_hw, mqprio, sizeof(*mqprio));
>> +
>> +	ret = pm_runtime_get_sync(common->dev);
>> +	if (ret < 0) {
>> +		pm_runtime_put_noidle(common->dev);
>> +		return ret;
>> +	}
>> +
>> +	if (!num_tc) {
>> +		am65_cpsw_reset_tc_mqprio(ndev);
>> +		goto exit_put;
>> +	}
>> +
>> +	ret = am65_cpsw_mqprio_verify_shaper(port, mqprio);
>> +	if (ret)
>> +		goto exit_put;
> 
> At "exit_put" we have "return 0" and this discards the value of "ret".

Will fix.

> 
>> +
>> +	netdev_set_num_tc(ndev, num_tc);
>> +
>> +	/* Multiple Linux priorities can map to a Traffic Class
>> +	 * A Traffic Class can have multiple contiguous Queues,
>> +	 * Queues get mapped to Channels (thread_id),
>> +	 *	if not VLAN tagged, thread_id is used as packet_priority
>> +	 *	if VLAN tagged. VLAN priority is used as packet_priorit
>> +	 * packet_priority gets mapped to header_priority in p0_rx_pri_map,
>> +	 * header_priority gets mapped to switch_priority in pn_tx_pri_map.
>> +	 * As p0_rx_pri_map is left at defaults (0x76543210), we can
>> +	 * assume that Queue_n gets mapped to header_priority_n. We can then
>> +	 * set the switch priority in pn_tx_pri_map.
>> +	 */
>> +
>> +	for (tc = 0; tc < num_tc; tc++) {
>> +		prio = tc;
>> +
>> +		/* For simplicity we assign the same priority (TCn) to
>> +		 * all queues of a Traffic Class.
>> +		 */
>> +		for (i = qopt->offset[tc]; i < qopt->offset[tc] + qopt->count[tc]; i++)
>> +			tx_prio_map |= prio << (4 * i);
>> +
>> +		count = qopt->count[tc];
>> +		offset = qopt->offset[tc];
>> +		netdev_set_tc_queue(ndev, tc, count, offset);
>> +	}
> 
> I think this is okay.
> 
>> +
>> +	writel(tx_prio_map,
>> +	       port->port_base + AM65_CPSW_PN_REG_TX_PRI_MAP);
> 
> Nit: This can be written on a single line.
> 
OK.

>> +
>> +	am65_cpsw_tx_pn_shaper_apply(port);
>> +
>> +exit_put:
>> +	pm_runtime_put(common->dev);
>> +	return 0;
>> +}
>> +
>>  int am65_cpsw_qos_ndo_setup_tc(struct net_device *ndev, enum tc_setup_type type,
>>  			       void *type_data)
>>  {

-- 
cheers,
-roger