Re: [PATCHv4] serial: imx: Add DMA buffer configuration via sysfs

[Uwe Kleine-König]

Hello Greg,

[Cc += dt maintainers]

On Tue, Apr 06, 2021 at 09:13:04AM +0200, Greg Kroah-Hartman wrote:
> On Mon, Apr 05, 2021 at 11:44:46PM +0200, Sebastian Reichel wrote:
> > On Fri, Mar 05, 2021 at 01:42:52PM +0100, Sebastian Reichel wrote:
> > > On Fri, Mar 05, 2021 at 01:06:12PM +0100, Greg Kroah-Hartman wrote:
> > > > On Fri, Mar 05, 2021 at 12:50:58PM +0100, Sebastian Reichel wrote:
> > > > > From: Fabien Lahoudere <fabien.lahoudere@xxxxxxxxxxxxx>
> > > > > 
> > > > > In order to optimize serial communication (performance/throughput VS
> > > > > latency), we may need to tweak DMA period number and size. This adds
> > > > > sysfs attributes to configure those values before initialising DMA.
> > > > > The defaults will stay the same as before (16 buffers with a size of
> > > > > 1024 bytes). Afterwards the values can be read/write with the
> > > > > following sysfs files:
> > > > > 
> > > > > /sys/class/tty/ttymxc*/dma_buffer_size
> > > > > /sys/class/tty/ttymxc*/dma_buffer_count
> > > > 
> > > > Ick no.  Custom sysfs attributes for things like serial ports are crazy.
> > > > 
> > > > > This is mainly needed for GEHC CS ONE (arch/arm/boot/dts/imx53-ppd.dts),
> > > > > which has multiple microcontrollers connected via UART controlling. One
> > > > > of the UARTs is connected to an on-board microcontroller at 19200 baud,
> > > > > which constantly pushes critical data (so aging character detect
> > > > > interrupt will never trigger). This data must be processed at 50-200 Hz,
> > > > > so UART should return data in less than 5-20ms. With 1024 byte DMA
> > > > > buffer (and a constant data stream) the read operation instead needs
> > > > > 1024 byte / 19200 baud = 53.333ms, which is way too long (note: Worst
> > > > > Case would be remote processor sending data with short pauses <= 7
> > > > > characters, which would further increase this number). The current
> > > > > downstream kernel instead configures 24 bytes resulting in 1.25ms,
> > > > > but that is obviously not sensible for normal UART use cases and cannot
> > > > > be used as new default.
> > > > 
> > > > Why can't this be a device tree attribute? Why does this have to be a
> > > > sysfs thing that no one will know how to tune and set over time.  This
> > > > hardware should not force a user to manually tune it to get it to work
> > > > properly, this isn't the 1990's anymore :(
> > > > 
> > > > Please never force a user to choose stuff like this, they never will
> > > > know what to do.
> > > 
> > > This used to be a DT attribute in PATCHv1. It has been moved over to
> > > sysfs since PATCHv2, since it does not describe the hardware, but
> > > configuration. Unfortunately lore.kernel.org does not have the full
> > > thread, but this is the discussion:
> > > 
> > > https://lore.kernel.org/linux-serial/20170629182618.jpahpmuq364ldcv2@xxxxxxxxxxxxxx/
> > > 
> > > From downstream POV this can be done either by adding a DT property
> > > to the UART node, or by adding a udev rule.
> > > 
> > > From my POV there is not a huge difference. In both cases we will
> > > be bound by an ABI afterwards, in both cases people will usually
> > > stick to the default value and in both cases people that do deviate
> > > from the default probably ran into problems and started to look
> > > for a solution.
> > 
> > ping? It's not very nice to get a rejected in cycles :(
> 
> I recommend working with the DT people here, as custom sysfs attributes
> for things like this that are really just describing the hardware is
> crazy.

I was one who expressed concerns in the earlier rounds that dt isn't the
right place for this. dt is about hardware description, but choosing
a good value for the dma buffer size is driver tuning and depends on the
individual requirements. (latency, throughput, memory consumption,
robustness under system load). I can even imagine use cases where the
settings should be changed dynamically, which cannot (easily) be done
using dt.

While I see your point that a driver specific sysfs property is
unusual/strange/whatever every downside you mentioned also applies to a
dt property (or a custom ioctl).

Among the solutions I can imagine, my preference order is:

 - automatic tuning
 - sysfs property
 - further discussion
 - dt property
 - custom ioctl

I wonder if there is a sensible way to implement a automatic tuning. In
the use case mentioned in the commit log, Sebastian's need is low
latency for a constantly sending microcontroller on the other side. Is
it sensible to make the used dma buffers smaller if we have a certain
throughput? Or is that too magic and doomed to fail covering most use
cases? If that doesn't work, I support Sebastian's approach to do that
in a sysfs property.

Sebastian, have you evaluated just not using dma for these UARTs?

Best regards
Uwe

-- 
Pengutronix e.K.                           | Uwe Kleine-König            |
Industrial Linux Solutions                 | https://www.pengutronix.de/ |

Attachment: signature.asc
Description: PGP signature