Re: [PATCH 02/15] can: m_can: Wakeup net queue once tx was issued - Markus Schneider-Pargmann

From: Markus Schneider-Pargmann <msp@baylibre.com>
To: Marc Kleine-Budde <mkl@pengutronix.de>
Cc: Chandrasekar Ramakrishnan <rcsekar@samsung.com>,
	Wolfgang Grandegger <wg@grandegger.com>,
	linux-can@vger.kernel.org, netdev@vger.kernel.org,
	linux-kernel@vger.kernel.org
Subject: Re: [PATCH 02/15] can: m_can: Wakeup net queue once tx was issued
Date: Thu, 1 Dec 2022 17:49:02 +0100	[thread overview]
Message-ID: <20221201164902.ipd3ctrtne47jtmv@blmsp> (raw)
In-Reply-To: <20221201091605.jgd7dlswcbxapdy3@pengutronix.de>

Hi Marc,

On Thu, Dec 01, 2022 at 10:16:05AM +0100, Marc Kleine-Budde wrote:
> On 01.12.2022 09:43:02, Markus Schneider-Pargmann wrote:
> > Hi Marc,
> > 
> > On Wed, Nov 30, 2022 at 06:21:00PM +0100, Marc Kleine-Budde wrote:
> > > On 16.11.2022 21:52:55, Markus Schneider-Pargmann wrote:
> > > > Currently the driver waits to wakeup the queue until the interrupt for
> > > > the transmit event is received and acknowledged. If we want to use the
> > > > hardware FIFO, this is too late.
> > > > 
> > > > Instead release the queue as soon as the transmit was transferred into
> > > > the hardware FIFO. We are then ready for the next transmit to be
> > > > transferred.
> > > 
> > > If you want to really speed up the TX path, remove the worker and use
> > > the spi_async() API from the xmit callback, see mcp251xfd_start_xmit().
> > 
> > Good idea. I will check how regmap's async_write works and if it is
> > suitable to do the job. I don't want to drop the regmap usage for this
> > right now.
> 
> IIRC regmap async write still uses mutexes, but sleeping is not allowed
> in the xmit handler. The mcp251xfd driver does the endianness conversion
> (and the optional CRC) manually for the TX path.

I just saw, you can force regmap to use spinlocks as well. But it uses
the same operation for sync operations as well.

> 
> Sending directly from the xmit handler basically eliminates the queuing
> between the network stack and the worker. Getting rid of the worker
> makes life easier and it's faster anyways.

The current implementation of the driver doesn't really queue anything
between the network stack and the worker. It is a queue of size 1 ;).

To be honest I would rather focus on the other things than on getting
rid of the worker completely as this can be done in a separate patch
later as well. Yes I agree it would be nice to get rid of the worker but
it is also probably not a major bottleneck for the performance and in
its current state it works. If I have time left at the end I will be
more than happy to do that. But for the moment I would just keep the
worker as it is. Is that OK for you?

Thanks,
Markus

> 
> > > Extra bonus if you implement xmit_more() and transfer more than 1 skb
> > > per SPI transfer.
> > 
> > That's on my todo list, but I am not sure I will get to it soonish.
> 
> I haven't implemented this for the mcp251xfd, yet, but I have some
> proof-of-concept code somewhere. However, the mcp251xfd driver already
> implemented byte queue limits: 0084e298acfe ("can: mcp251xfd: add BQL
> support").
> 
> 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 |