Re: [PATCH v3 5/6] can: c_can: prepare to up the message objects number - Marc Kleine-Budde

From: Marc Kleine-Budde <mkl@pengutronix.de>
To: Dario Binacchi <dariobin@libero.it>
Cc: Kurt Van Dijck <dev.kurt@vandijck-laurijssen.be>,
	linux-kernel@vger.kernel.org,
	Federico Vaga <federico.vaga@gmail.com>,
	Alexander Stein <alexander.stein@systec-electronic.com>,
	"David S. Miller" <davem@davemloft.net>,
	Jakub Kicinski <kuba@kernel.org>,
	Oliver Hartkopp <socketcan@hartkopp.net>,
	Vincent Mailhol <mailhol.vincent@wanadoo.fr>,
	Wolfgang Grandegger <wg@grandegger.com>,
	YueHaibing <yuehaibing@huawei.com>,
	Zhang Qilong <zhangqilong3@huawei.com>,
	linux-can@vger.kernel.org, netdev@vger.kernel.org
Subject: Re: [PATCH v3 5/6] can: c_can: prepare to up the message objects number
Date: Wed, 3 Mar 2021 10:00:36 +0100	[thread overview]
Message-ID: <20210303090036.aocqk6gp3vqnzaku@pengutronix.de> (raw)
In-Reply-To: <91394876.26757.1614759793793@mail1.libero.it>

[-- Attachment #1: Type: text/plain, Size: 2136 bytes --]

On 03.03.2021 09:23:13, Dario Binacchi wrote:
[...]
> > > @@ -1205,17 +1203,31 @@ static int c_can_close(struct net_device *dev)
> > >  	return 0;
> > >  }
> > >  
> > > -struct net_device *alloc_c_can_dev(void)
> > > +struct net_device *alloc_c_can_dev(int msg_obj_num)
> > >  {
> > >  	struct net_device *dev;
> > >  	struct c_can_priv *priv;
> > > +	int msg_obj_tx_num = msg_obj_num / 2;
> > 
> > IMO, a bigger tx queue is not usefull.
> > A bigger rx queue however is.
> 
> This would not be good for my application. I think it really depends
> on the type of application. We can probably say that being able to
> size rx/tx queue would be a useful feature.

Ok. There is an ethtool interface to configure the size of the RX and TX
queues. In ethtool it's called the RX/TX "ring" size and you can get it
via the -g parameter, e.g. here for by Ethernet interface:

| $ ethtool -g enp0s25
| Ring parameters for enp0s25:
| Pre-set maximums:
| RX:		4096
| RX Mini:	n/a
| RX Jumbo:	n/a
| TX:		4096
| Current hardware settings:
| RX:		256
| RX Mini:	n/a
| RX Jumbo:	n/a
| TX:		256

If I understand correctly patch 6 has some assumptions that RX and TX
are max 32. To support up to 64 RX objects, you have to convert:
- u32 -> u64
- BIT() -> BIT_ULL()
- GENMASK() -> GENMASK_ULL()

The register access has to be converted, too. For performance reasons
you want to do as least as possible. Which is probably the most
complicated.

In the flexcan driver I have a similar problem. The driver keeps masks,
which mailboxes are RX and which TX and I added wrapper functions to
minimize IO access:

https://elixir.bootlin.com/linux/v5.11/source/drivers/net/can/flexcan.c#L904

This should to IMHO into patch 6.

Adding the ethtool support and making the rings configurable would be a
separate patch.

regards,
Marc

-- 
Pengutronix e.K.                 | Marc Kleine-Budde           |
Embedded Linux                   | https://www.pengutronix.de  |
Vertretung West/Dortmund         | Phone: +49-231-2826-924     |
Amtsgericht Hildesheim, HRA 2686 | Fax:   +49-5121-206917-5555 |

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]