Re: [PATCH net-next v3 1/6] net: ethernet: ti: am65-cpsw: Introduce multi queue Rx

[Roger Quadros <rogerq@kernel.org>]

Hi Joe,

On 27/07/2024 09:29, Roger Quadros wrote:
> 
> 
> On 23/07/2024 20:11, Joe Damato wrote:
>> On Wed, Jul 03, 2024 at 04:51:32PM +0300, Roger Quadros wrote:
>>
>> [...]
>>
>>> @@ -699,6 +727,14 @@ static int am65_cpsw_nuss_common_open(struct am65_cpsw_common *common)
>>>  		goto fail_rx;
>>>  	}
>>>  
>>> +	for (i = 0; i < common->rx_ch_num_flows ; i++) {
>>> +		napi_enable(&common->rx_chns.flows[i].napi_rx);
>>> +		if (common->rx_chns.flows[i].irq_disabled) {
>>> +			common->rx_chns.flows[i].irq_disabled = false;
>>
>> Just a minor nit (not a reason to hold this back): I've been
>> encouraging folks to use the new netdev-genl APIs in their drivers
>> to map NAPIs to queues and IRQs if possible because it allows for
>> more expressive and interesting userland applications.
>>
>> You may consider in the future using something vaguely like (this is
>> untested psuedo-code I just typed out):
>>
>>    netif_napi_set_irq(&common->rx_chns.flows[i].napi_rx,
>>                       common->rx_chns.flows[i].irq);
>>
>> and 
>>
>>    netif_queue_set_napi(common->dev, i, NETDEV_QUEUE_TYPE_RX,
>>                         &common->rx_chns.flows[i].napi_rx);

common->dev is not actually struct net_device.
The hardware has one set of queues for more than one network ports.
Currently the NAPI queues are associated only with the first network device
that is brought up.

In such a case, what must be done with netif_queue_set_napi()?
do we still associate the same napi with all the network devices?

>>
>> To link everything together (note that RTNL must be held while doing
>> this -- I haven't checked your code path to see if that is true here).

This happens as part of ndo_open call. We don't explicitly hold RTNL here.

>>
>> For an example, see 64b62146ba9e ("net/mlx4: link NAPI instances to
>> queues and IRQs). 
>>
>> Doing this would allow userland to get data via netlink, which you
>> can examine yourself by using cli.py like this:
>>
>> python3 tools/net/ynl/cli.py \
>>   --spec Documentation/netlink/specs/netdev.yaml \
>>   --dump queue-get
>>
>> python3 tools/net/ynl/cli.py \
>>   --spec Documentation/netlink/specs/netdev.yaml \
>>   --dump napi-get
>>
> 
> Thanks for the pionters. I will check and see if I can incorportate
> this in the next spin.
> 
>>> +			enable_irq(common->rx_chns.flows[i].irq);
>>> +		}
>>> +	}
>>> +
>>>  	for (tx = 0; tx < common->tx_ch_num; tx++) {
>>>  		ret = k3_udma_glue_enable_tx_chn(tx_chn[tx].tx_chn);
>>>  		if (ret) {
>>> @@ -710,12 +746,6 @@ static int am65_cpsw_nuss_common_open(struct am65_cpsw_common *common)
>>>  		napi_enable(&tx_chn[tx].napi_tx);
>>>  	}
>>>  
>>> -	napi_enable(&common->napi_rx);
>>> -	if (common->rx_irq_disabled) {
>>> -		common->rx_irq_disabled = false;
>>> -		enable_irq(rx_chn->irq);
>>> -	}
>>> -
>>>  	dev_dbg(common->dev, "cpsw_nuss started\n");
>>>  	return 0;
>>>  
>>> @@ -726,11 +756,24 @@ static int am65_cpsw_nuss_common_open(struct am65_cpsw_common *common)
>>>  		tx--;
>>>  	}
>>>  
>>> +	for (flow_idx = 0; i < common->rx_ch_num_flows; flow_idx++) {
>>> +		flow = &rx_chn->flows[flow_idx];
>>> +		if (!flow->irq_disabled) {
>>> +			disable_irq(flow->irq);
>>> +			flow->irq_disabled = true;
>>> +		}
>>> +		napi_disable(&flow->napi_rx);
>>> +	}
>>> +
>>>  	k3_udma_glue_disable_rx_chn(rx_chn->rx_chn);
>>>  
>>>  fail_rx:
>>> -	k3_udma_glue_reset_rx_chn(rx_chn->rx_chn, 0, rx_chn,
>>> -				  am65_cpsw_nuss_rx_cleanup, 0);
>>> +	for (i = 0; i < common->rx_ch_num_flows; i--)
>>> +		k3_udma_glue_reset_rx_chn(rx_chn->rx_chn, i, &rx_chn->flows[i],
>>> +					  am65_cpsw_nuss_rx_cleanup, !!i);
>>> +
>>> +	am65_cpsw_destroy_xdp_rxqs(common);
>>> +
>>>  	return ret;
>>>  }
>>>  
>>> @@ -779,12 +822,12 @@ static int am65_cpsw_nuss_common_stop(struct am65_cpsw_common *common)
>>>  			dev_err(common->dev, "rx teardown timeout\n");
>>>  	}
>>>  
>>> -	napi_disable(&common->napi_rx);
>>> -	hrtimer_cancel(&common->rx_hrtimer);
>>> -
>>> -	for (i = 0; i < AM65_CPSW_MAX_RX_FLOWS; i++)
>>> -		k3_udma_glue_reset_rx_chn(rx_chn->rx_chn, i, rx_chn,
>>> +	for (i = 0; i < common->rx_ch_num_flows; i++) {
>>> +		napi_disable(&common->rx_chns.flows[i].napi_rx);
>>
>> The inverse of the above is probably true somewhere around here;
>> again a small piece of psuedo code for illustrative purposes:
>>
>>    netif_queue_set_napi(common->dev, i, NETDEV_QUEUE_TYPE_RX, NULL);
>>
>>> +		hrtimer_cancel(&common->rx_chns.flows[i].rx_hrtimer);
>>> +		k3_udma_glue_reset_rx_chn(rx_chn->rx_chn, i, &rx_chn->flows[i],
>>>  					  am65_cpsw_nuss_rx_cleanup, !!i);
>>> +	}
>>>  
>>>  	k3_udma_glue_disable_rx_chn(rx_chn->rx_chn);
>>>  
> 

-- 
cheers,
-roger