Re: [RFC PATCH 0/3] UART slave device bus

From: "H. Nikolaus Schaller" <hns@goldelico.com>
To: One Thousand Gnomes <gnomes@lxorguk.ukuu.org.uk>
Cc: Sebastian Reichel <sre@kernel.org>, Rob Herring <robh@kernel.org>,
	Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
	Marcel Holtmann <marcel@holtmann.org>,
	Jiri Slaby <jslaby@suse.com>, Pavel Machek <pavel@ucw.cz>,
	Peter Hurley <peter@hurleysoftware.com>,
	NeilBrown <neil@brown.name>, Arnd Bergmann <arnd@arndb.de>,
	Linus Walleij <linus.walleij@linaro.org>,
	"open list:BLUETOOTH DRIVERS" <linux-bluetooth@vger.kernel.org>,
	"linux-serial@vger.kernel.org" <linux-serial@vger.kernel.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Subject: Re: [RFC PATCH 0/3] UART slave device bus
Date: Sun, 21 Aug 2016 20:23:10 +0200	[thread overview]
Message-ID: <FF7761E8-377A-43AD-96B2-83BA140E030B@goldelico.com> (raw)
In-Reply-To: <20160821180910.51cf919d@lxorguk.ukuu.org.uk>

> Am 21.08.2016 um 19:09 schrieb One Thousand Gnomes <gnomes@lxorguk.ukuu.org.uk>:
> 
>> Let me ask a question about your centralized and pre-cooked buffering approach.
>> 
>> As far as I see, even then the kernel API must notify the driver at the right moment
>> that a new block has arrived. Right?
> 
> The low level driver queues words (data byte, flag byte)
> The buffer processing workqueue picks those bytes from the queue and
> atomically empties the queue

When and how fast is the work queue scheduled?
And by which event?

> The workqueue involves the receive handler.

This should be faster than if a driver directly processes incoming bytes?

> 
>> But how does the kernel API know how long such a block is?
> 
> It's as long as the data that has arrived in that time.

Which means the work queue handler have to decide if it is enough for a
frame to decode and if not, wait a little until more arrives.

Or you have to assemble chunks into a frame, i.e. copy data around.

Both seems a waste of scarce cpu cycles in high-speed situations to me.

> 
>> Usually there is a start byte/character, sometimes a length indicator, then payload data,
>> some checksum and finally a stop byte/character. For NMEA it is $, no length, * and \r\n.
>> For other serial protocols it might be AT, no length, and \r. Or something different.
>> HCI seems to use 2 byte op-code or 1 byte event code and 1 byte parameter length.
> 
> It doesn't look for any kind of protocol block headers.

Which might become the pitfall of the design because as I have described it is an
essential part of processing UART based protocols. You seem to focus on efficiently
buffering only but not about efficiently processing the queued data.

> The routine
> invoked by the work queue does any frame recovery.

> 
>> So I would even conclude that you usually can't even use DMA based UART receive
>> processing for arbitrary and not well-defined protocols. Or have to assume that the
> 
> We do, today for bluetooth and other protocols just fine

I think it works (even with user-space HCI daemon) because bluetooth HCI is slow (<300kByte/s).

> - it's all about
> data flows not about framing in the protocol sense.

Yes, but you should also take framing into account for a solution that helps to implement
UART slave devices. That is my concern.

BR,
Nikolaus