New serial card development

New serial card development

[Matt Schulte]

Hello, my name is Matt and I have recently developed a PCIe card based
on the Exar 17v35x series of PCIe multiport UART chips.

We have written Linux drivers for our other products in the past but
they have never been what I would call the best practice for Linux
development.

This time I would like to write a driver that uses the best practices
and possibly submit it to the kernel when I'm done.

I am a little bit confused about what method would be best for this
device.   I have been examining the sample driver provided by Exar
and it seems that the only things that they have that are really
different from the generic 8250 serial driver are the interrupt
handler (optimized for their multiple ports), the transmit and
receive character functions (modified to use their 256 byte FIFOs)
and the function that calculates the baud.  Also I have two features
specific to my card that I would likely need to use an ioctl for.

If these are the only differences would I be able to create a driver
like what is seen in the /drivers/tty/serial/8250 directory (8250_dw.c
for example)?  Or would I need to do something similar to what I find in
the /drviers/tty/serial directory?

Thank you for your time.

Matt Schulte

Re: New serial card development

[Theodore Ts'o]

On Tue, Oct 09, 2012 at 01:43:10PM -0500, Matt Schulte wrote:
> Hello, my name is Matt and I have recently developed a PCIe card based
> on the Exar 17v35x series of PCIe multiport UART chips.
> 
> We have written Linux drivers for our other products in the past but
> they have never been what I would call the best practice for Linux
> development.
> 
> This time I would like to write a driver that uses the best practices
> and possibly submit it to the kernel when I'm done.
> 
> I am a little bit confused about what method would be best for this
> device.   I have been examining the sample driver provided by Exar
> and it seems that the only things that they have that are really
> different from the generic 8250 serial driver are the interrupt
> handler (optimized for their multiple ports), the transmit and
> receive character functions (modified to use their 256 byte FIFOs)
> and the function that calculates the baud.  Also I have two features
> specific to my card that I would likely need to use an ioctl for.

Hi Matt,

It's been a while since I've done serial development (probably close
to a decade ago), but way back when I was responsible for most of the
code which is now in drivers/tty/serial/8250/serial.c, I had added
support for the Exar 16C850 UART; since then someone else has added
support for the Exar XR17D15x UART.

The EXAR 16C850 had a 128 byte FIFO; it sounds like this EXAR variant
is similar to the 16C850, although it's been further extended.

> If these are the only differences would I be able to create a driver
> like what is seen in the /drivers/tty/serial/8250 directory (8250_dw.c
> for example)?  Or would I need to do something similar to what I find in
> the /drviers/tty/serial directory?

We support a very large number of 8250-like UART's already.  It
doesn't make sense create a new driver which is essentially a clone of
the 8250.c driver.  Instead, the 8250.c driver should be extended to
support this new Exar UART variant.  The preferred way to do this is
to detect the EXAR UART directly, via the method documented in the
EXAR data sheet, instead of using the PCI id numbers to make this
determination.  This way, we will be able to support that UART on
other PCI serial cards, and not just yours.

If you have some features which are explicitly specific to your card,
and which are not specific to the UART, then that may need to be
separately triggered via the PCI vendor/device id numbers.  What
feature is it, by the way?  If it has to do with controlling the clock
crystal, there's infrastructure to deal with this already that you may
be able to leverage.

Regards,

						- Ted

Re: New serial card development

[Matt Schulte]

On Sun, Oct 14, 2012 at 4:37 AM, Theodore Ts'o <tytso@mit.edu> wrote:
>
> On Tue, Oct 09, 2012 at 01:43:10PM -0500, Matt Schulte wrote:
> > Hello, my name is Matt and I have recently developed a PCIe card based
> > on the Exar 17v35x series of PCIe multiport UART chips.
> >
> > We have written Linux drivers for our other products in the past but
> > they have never been what I would call the best practice for Linux
> > development.
> >
> > This time I would like to write a driver that uses the best practices
> > and possibly submit it to the kernel when I'm done.
> >
> > I am a little bit confused about what method would be best for this
> > device.   I have been examining the sample driver provided by Exar
> > and it seems that the only things that they have that are really
> > different from the generic 8250 serial driver are the interrupt
> > handler (optimized for their multiple ports), the transmit and
> > receive character functions (modified to use their 256 byte FIFOs)
> > and the function that calculates the baud.  Also I have two features
> > specific to my card that I would likely need to use an ioctl for.
>
> Hi Matt,
>
> It's been a while since I've done serial development (probably close
> to a decade ago), but way back when I was responsible for most of the
> code which is now in drivers/tty/serial/8250/serial.c, I had added
> support for the Exar 16C850 UART; since then someone else has added
> support for the Exar XR17D15x UART.
>
> The EXAR 16C850 had a 128 byte FIFO; it sounds like this EXAR variant
> is similar to the 16C850, although it's been further extended.
>
> > If these are the only differences would I be able to create a driver
> > like what is seen in the /drivers/tty/serial/8250 directory (8250_dw.c
> > for example)?  Or would I need to do something similar to what I find in
> > the /drviers/tty/serial directory?
>
> We support a very large number of 8250-like UART's already.  It
> doesn't make sense create a new driver which is essentially a clone of
> the 8250.c driver.  Instead, the 8250.c driver should be extended to
> support this new Exar UART variant.  The preferred way to do this is
> to detect the EXAR UART directly, via the method documented in the
> EXAR data sheet, instead of using the PCI id numbers to make this
> determination.  This way, we will be able to support that UART on
> other PCI serial cards, and not just yours.


That sounds reasonable but I am not sure what you mean my the "method
documented in the Exar data sheet."  I've been through this thing
quite a bit and I don't see anything about how one might detect the
ports without the PCI id numbers.  Could you explain what you mean by
this?
>
>
> If you have some features which are explicitly specific to your card,
> and which are not specific to the UART, then that may need to be
> separately triggered via the PCI vendor/device id numbers.  What
> feature is it, by the way?  If it has to do with controlling the clock
> crystal, there's infrastructure to deal with this already that you may
> be able to leverage.

* This new Exar UART does not use a crystal it uses a clock from the
PCI Express bus.  Also it has a fractional divisor that gets them
1/16th (0.0625) resolution on their divisors.  The normal DLL and DLM
registers contain the integer part of the divisor and the DLD register
provides the fractional component.  The divisor is able to be set to
anything between 1 and (2^16 - 0.0625) in increments of 0.0625.

* These UARTs have 256 byte FIFOs per port per direction.  I would
like to utilize these to their fullest potential.

* I need to toggle 16-bit general purpose IO signals to control two
things on my card: Receive Echo Cancel (useful for 2-wire 485) and the
selectable termination resistance on my receiver chips.

* I need to change two UART specific registers to change the sampling
mode (16x, 8x and 4x).

* I need to figure out if the new RS485 elements that have been added
to 8250 can be used on these Exar parts and if not I need something
that can enable RS485 for us.

* I need to be able to enable 9-bit mode, what they call Normal
Multi-drop as well as the Auto Address Detection Mode used in
conjunction with the Multi-drop mode.

* There is an interruptible hardware timer/counter that I might like
to utilize at some point.

With this laundry list of items that I would like to include, do you
still feel like this is something that should be included in the
kernel's 8250.c/8250_pci.c drivers?

Thanks for helping me think this through.

Matt

Re: New serial card development

[Alan Cox]

> That sounds reasonable but I am not sure what you mean my the "method
> documented in the Exar data sheet."  I've been through this thing
> quite a bit and I don't see anything about how one might detect the
> ports without the PCI id numbers.  Could you explain what you mean by
> this?

I don't believe you can identify the new Exars except by PCI id - which
is fine as the type can be passed via the 8250_pci driver.

> * This new Exar UART does not use a crystal it uses a clock from the
> PCI Express bus.  Also it has a fractional divisor that gets them
> 1/16th (0.0625) resolution on their divisors.  The normal DLL and DLM
> registers contain the integer part of the divisor and the DLD register
> provides the fractional component.  The divisor is able to be set to
> anything between 1 and (2^16 - 0.0625) in increments of 0.0625.

That probably needs a bit of tweaking to getthe best results

> * These UARTs have 256 byte FIFOs per port per direction.  I would
> like to utilize these to their fullest potential.

We already handle other similar fifos so again in theory you just need to
set the size and it works.

> * I need to toggle 16-bit general purpose IO signals to control two
> things on my card: Receive Echo Cancel (useful for 2-wire 485) and the
> selectable termination resistance on my receiver chips.

Ok - we don't have a generalized way of doing that in the RS485 bits but
anything liek this should get added generically (in case other chips want
to do it)
 
> * I need to change two UART specific registers to change the sampling
> mode (16x, 8x and 4x).

What determines this.

> * I need to figure out if the new RS485 elements that have been added
> to 8250 can be used on these Exar parts and if not I need something
> that can enable RS485 for us.

They should provide the basics.

> * I need to be able to enable 9-bit mode, what they call Normal
> Multi-drop as well as the Auto Address Detection Mode used in
> conjunction with the Multi-drop mode.

That may actually be the most "fun" bit - the entire Linux tty layer is
devoutly 5-8bit, although people do some horrible hacks with  "8 plus
parity".

> With this laundry list of items that I would like to include, do you
> still feel like this is something that should be included in the
> kernel's 8250.c/8250_pci.c drivers?

I think so. It's also the path that lets you do basic enabling, get
it working upstream and then extend as is needed for the other features.

The one area 8250 really can't hack properly is DMA driven 8250 devices.

Alan

Re: New serial card development

[Theodore Ts'o]

On Tue, Oct 16, 2012 at 12:26:08AM +0100, Alan Cox wrote:
> > That sounds reasonable but I am not sure what you mean my the "method
> > documented in the Exar data sheet."  I've been through this thing
> > quite a bit and I don't see anything about how one might detect the
> > ports without the PCI id numbers.  Could you explain what you mean by
> > this?
> 
> I don't believe you can identify the new Exars except by PCI id - which
> is fine as the type can be passed via the 8250_pci driver.

Yeah, I just checked out the data sheet, and it looks like there isn't
a way to do this.  That's unfortunate, since board vendors can change
the PCI vendor and device ID.  So if they reprogram the EEPROM to use
their own id numbers, we won't be able to support that new board
without modifying the kernel source.  This is why I had tried to
implement ways where if the PCI class indicated a 16550 compatible
port, we would try to probe the UART registers to figure out exactly
what type of advanced features beyond those supported by the standard
16550A UART.  Sigh, it looks like we won't be able to do things that
way.

> > * I need to change two UART specific registers to change the sampling
> > mode (16x, 8x and 4x).
> 
> What determines this.

We have support for things like this already; see
serial8250_get_divisor and UPF_MAGIC_MULTIPLIER.  In general you only
want to change the sampling mode when you need it (i.e, for the high
speed baud rates).

> They should provide the basics.
> 
> > * I need to be able to enable 9-bit mode, what they call Normal
> > Multi-drop as well as the Auto Address Detection Mode used in
> > conjunction with the Multi-drop mode.
> 
> That may actually be the most "fun" bit - the entire Linux tty layer is
> devoutly 5-8bit, although people do some horrible hacks with  "8 plus
> parity".

That's what is actually recommended by the Exar application note;
since setting the 9th bit is only used for indicating that the lower 8
bits is the "address" bit for demultiplexing destination "drop", you
don't need to do it very often compared to the data bytes.  So
toggling the parity bit is the best way to deal with the fact you can
only use outb to send 8 bits.

						- Ted

Re: New serial card development

[Matt Schulte]

Forgot to reply to all.
---------- Forwarded message ----------
From: Matt Schulte <matts@commtech-fastcom.com>
Date: Wed, Oct 17, 2012 at 2:10 PM
Subject: Re: New serial card development
To: Alan Cox <alan@lxorguk.ukuu.org.uk>


On Mon, Oct 15, 2012 at 6:26 PM, Alan Cox <alan@lxorguk.ukuu.org.uk> wrote:
>> That sounds reasonable but I am not sure what you mean my the "method
>> documented in the Exar data sheet."  I've been through this thing
>> quite a bit and I don't see anything about how one might detect the
>> ports without the PCI id numbers.  Could you explain what you mean by
>> this?
>
> I don't believe you can identify the new Exars except by PCI id - which
> is fine as the type can be passed via the 8250_pci driver.
>
>> * This new Exar UART does not use a crystal it uses a clock from the
>> PCI Express bus.  Also it has a fractional divisor that gets them
>> 1/16th (0.0625) resolution on their divisors.  The normal DLL and DLM
>> registers contain the integer part of the divisor and the DLD register
>> provides the fractional component.  The divisor is able to be set to
>> anything between 1 and (2^16 - 0.0625) in increments of 0.0625.
>
> That probably needs a bit of tweaking to getthe best results

Does this make you lean one way or the other in how I would best
implement this driver?

>> * These UARTs have 256 byte FIFOs per port per direction.  I would
>> like to utilize these to their fullest potential.
>
> We already handle other similar fifos so again in theory you just need to
> set the size and it works.

Good, I'm sure it will be fine ;)

>> * I need to toggle 16-bit general purpose IO signals to control two
>> things on my card: Receive Echo Cancel (useful for 2-wire 485) and the
>> selectable termination resistance on my receiver chips.
>
> Ok - we don't have a generalized way of doing that in the RS485 bits but
> anything liek this should get added generically (in case other chips want
> to do it)

I have two features, one that could be considered part of RS485 and
one that has nothing at all to do with RS485.

>> * I need to change two UART specific registers to change the sampling
>> mode (16x, 8x and 4x).
>
> What determines this.

User would have to decide to turn on non-standard 8x or 4x sampling mode.

>> * I need to figure out if the new RS485 elements that have been added
>> to 8250 can be used on these Exar parts and if not I need something
>> that can enable RS485 for us.
>
> They should provide the basics.

Where will I take the built in 485 flags and turn them into whatever
my board needs?

>> * I need to be able to enable 9-bit mode, what they call Normal
>> Multi-drop as well as the Auto Address Detection Mode used in
>> conjunction with the Multi-drop mode.
>
> That may actually be the most "fun" bit - the entire Linux tty layer is
> devoutly 5-8bit, although people do some horrible hacks with  "8 plus
> parity".

Transmitting is a little bit interesting but manageable.  In the
receive direction the chip can detect the 9th bit and recognize it as
an address bit and then check that byte against a preset station
address that has been specified by the user.

>> With this laundry list of items that I would like to include, do you
>> still feel like this is something that should be included in the
>> kernel's 8250.c/8250_pci.c drivers?
>
> I think so. It's also the path that lets you do basic enabling, get
> it working upstream and then extend as is needed for the other features.

So is there an example of a device already in the kernel that you
think I should try to emulate? Will I be best served with something
like 8250_dw.c or do you think I should try to get all of my things
directly into 8250 and 8250_pci?

Matt Schulte

Re: New serial card development

[Matt Schulte]

On Mon, Oct 15, 2012 at 9:32 PM, Theodore Ts'o <tytso@mit.edu> wrote:
> On Tue, Oct 16, 2012 at 12:26:08AM +0100, Alan Cox wrote:
>> > That sounds reasonable but I am not sure what you mean my the "method
>> > documented in the Exar data sheet."  I've been through this thing
>> > quite a bit and I don't see anything about how one might detect the
>> > ports without the PCI id numbers.  Could you explain what you mean by
>> > this?
>>
>> I don't believe you can identify the new Exars except by PCI id - which
>> is fine as the type can be passed via the 8250_pci driver.
>
> Yeah, I just checked out the data sheet, and it looks like there isn't
> a way to do this.  That's unfortunate, since board vendors can change
> the PCI vendor and device ID.  So if they reprogram the EEPROM to use
> their own id numbers, we won't be able to support that new board
> without modifying the kernel source.  This is why I had tried to
> implement ways where if the PCI class indicated a 16550 compatible
> port, we would try to probe the UART registers to figure out exactly
> what type of advanced features beyond those supported by the standard
> 16550A UART.  Sigh, it looks like we won't be able to do things that
> way.

Which is exactly what I do to the device id :)

>> > * I need to change two UART specific registers to change the sampling
>> > mode (16x, 8x and 4x).
>>
>> What determines this.
>
> We have support for things like this already; see
> serial8250_get_divisor and UPF_MAGIC_MULTIPLIER.  In general you only
> want to change the sampling mode when you need it (i.e, for the high
> speed baud rates).

I'm not sure the UPF_MAGIC_MULTIPLIER is the same as the sampling rate
that I'm looking for, though I might be able to come up with a way to
hijack the flag to do what I want.

>> They should provide the basics.
>>
>> > * I need to be able to enable 9-bit mode, what they call Normal
>> > Multi-drop as well as the Auto Address Detection Mode used in
>> > conjunction with the Multi-drop mode.
>>
>> That may actually be the most "fun" bit - the entire Linux tty layer is
>> devoutly 5-8bit, although people do some horrible hacks with  "8 plus
>> parity".
>
> That's what is actually recommended by the Exar application note;
> since setting the 9th bit is only used for indicating that the lower 8
> bits is the "address" bit for demultiplexing destination "drop", you
> don't need to do it very often compared to the data bytes.  So
> toggling the parity bit is the best way to deal with the fact you can
> only use outb to send 8 bits.

What I would need to have happen is for a flag to be set that says
"enable 9-bit mode transmit" or something.  When set, the first byte
of data of a given write will go out with mark parity, then the rest
of the bytes will have space parity.  The one byte with mark is the
address byte.

In the receive direction I would need to optionally enable a station
address which the chip will use to verify all incoming data and will
reject anything that is not of the appropriate address.

Matt Schulte

Re: New serial card development

[Alan Cox]

> > That probably needs a bit of tweaking to getthe best results
> 
> Does this make you lean one way or the other in how I would best
> implement this driver?

I would always start from what we have.

I wonder actually if there is a bit of a disconnect here from the line
of questions. I know in a lot of environments the model is

	concept design
	internal approval/verification
	consult with other stakeholders
	commit to model	
	build
	release

but in the Linux space the model is

	don't break anyone elses stuff
	get it working
	figure out as you go what is needed
	if it doesn't work out adjust
	polish as it becomes clear what can be done better
	repeat

So even if you submit initial patches to make it work with 8250.c and it
turns out that actually this was the wrong thing everyone will be happy
with the patches that then relocate it.

In addition what tends to occur is that the rules change anyway. You get
half way through submitting support for 256 byte fifos and someone comes
along with another different device that has the feature and so on.

So I would start from 8250.c and get trivial stuff like the fifo changes,
the PCI identifiers and so forth in as a small set of patches and then
work long the rest of it.

Even 'don't break other people's stuff' isn't a hard rule. The real rule
is 'if you break other people's stuff fix it back up' - because it
happens, and it's unavoidable.

Alan

Re: New serial card development

[Theodore Ts'o]

On Wed, Oct 17, 2012 at 03:24:06PM -0500, Matt Schulte wrote:
> 
> What I would need to have happen is for a flag to be set that says
> "enable 9-bit mode transmit" or something.  When set, the first byte
> of data of a given write will go out with mark parity, then the rest
> of the bytes will have space parity.  The one byte with mark is the
> address byte.

Why not just have a userspace library routine which simply uses
tcgetattr/tcsetattr to toggle a bit (probably a new bit in c_cflag) in
struct termios, issue the single byte write, and then toggle the bit
back?

If you really want to implement a special case "the first byte should
treated differently from the rest", that gets problemtic for a number
of reasons.  First of all, it's just a lot more kernel code; you'd
probably need to implement a new line displine, and secondly, if
userspace sends too much data so that it can't fit in the kernel
buffer, you might get a partial write.  In which case, when you send a
second write call to send the rest of the data, you'd have to go
through extra work to make sure the first byte of that second,
follow-up byte doesn't have the high 9th bit set.

Of course, you could hack in even more changes to make the write
system call behave entirely differently in your magic "9th bit mode",
where the write system call becomes a "all or nothing" sort of thing,
but that's even more evil hackery that almost certainly would be
rejected with extreme prejudice by the kernel maintainers.

In Linux, system calls are fast, so using some extra ioctl's to toggle
the termios structure is probably the way to go.  It's also much more
general since it doesn't presume a very specific protocol (i.e., your
magic multi-drop protocol).

> In the receive direction I would need to optionally enable a station
> address which the chip will use to verify all incoming data and will
> reject anything that is not of the appropriate address.

For the receive direction, my suggestion is to encode it as an
extension for how PARMRK does things, and then leave the parsing of
the address bits to userspace.  Otherwise you'll end up needing to
make the kernel know what are "appropriate" addresses, which again is
a lot of very protocol specific knowledge that you would have to push
into the kernel.

Alan's advice to get your card working as a basic serial card is very
good one.  Get basic functionality working, and then you can add the
support for the extra bits later....

						- Ted

Re: New serial card development

[Matt Schulte]

On Fri, Oct 19, 2012 at 4:21 PM, Theodore Ts'o <tytso@mit.edu> wrote:
> On Wed, Oct 17, 2012 at 03:24:06PM -0500, Matt Schulte wrote:
>>
>> What I would need to have happen is for a flag to be set that says
>> "enable 9-bit mode transmit" or something.  When set, the first byte
>> of data of a given write will go out with mark parity, then the rest
>> of the bytes will have space parity.  The one byte with mark is the
>> address byte.
>
> Why not just have a userspace library routine which simply uses
> tcgetattr/tcsetattr to toggle a bit (probably a new bit in c_cflag) in
> struct termios, issue the single byte write, and then toggle the bit
> back?

> In Linux, system calls are fast, so using some extra ioctl's to toggle
> the termios structure is probably the way to go.  It's also much more
> general since it doesn't presume a very specific protocol (i.e., your
> magic multi-drop protocol).

The only reason that I could come up with as to why not to do that is
the probably delay between the first byte and the second byte.
Perhaps the system calls will be so quick that this delay won't be an
issue.  Other than that I think that a new bit in c_cflag would be a
pretty decent solution.

Also this isn't a magic protocol, it is a very old protocol that is
apparently still in use today because I have several customers still
asking for it.

>> In the receive direction I would need to optionally enable a station
>> address which the chip will use to verify all incoming data and will
>> reject anything that is not of the appropriate address.
>
> For the receive direction, my suggestion is to encode it as an
> extension for how PARMRK does things, and then leave the parsing of
> the address bits to userspace.  Otherwise you'll end up needing to
> make the kernel know what are "appropriate" addresses, which again is
> a lot of very protocol specific knowledge that you would have to push
> into the kernel.

The address is something that is decided by the user, it could
literally be anything.  All I need is a way to write a byte to the
appropriate UART register and a bit to turn it on and off.

> Alan's advice to get your card working as a basic serial card is very
> good one.  Get basic functionality working, and then you can add the
> support for the extra bits later....

I can see the logic in getting it working as a basic serial card
first.  I think at minimum I would still need to implement the extra
divisor calculations to get accurate bit rates.

So when it works as a basic serial card, I assume you would want me to
use the default PCI IDs to keep it more generic.  Then would I add my
own PCI IDs and refer them back to the generic port?

Matt

Re: New serial card development

[Matt Schulte]

On Tue, Oct 23, 2012 at 11:27 AM, Matt Schulte
<matts@commtech-fastcom.com> wrote:
> On Fri, Oct 19, 2012 at 4:21 PM, Theodore Ts'o <tytso@mit.edu> wrote:
>> Alan's advice to get your card working as a basic serial card is very
>> good one.  Get basic functionality working, and then you can add the
>> support for the extra bits later....
>
> I can see the logic in getting it working as a basic serial card
> first.  I think at minimum I would still need to implement the extra
> divisor calculations to get accurate bit rates.
>
> So when it works as a basic serial card, I assume you would want me to
> use the default PCI IDs to keep it more generic.  Then would I add my
> own PCI IDs and refer them back to the generic port?
>

As more of a procedural question, when I go to make the patch(es) to
submit which, kernel repo do I start with?  Do I start with Greg KH's
tty repo and then generate the patches and submit them here?

Matt

Re: New serial card development

[Grant Edwards]

On 2012-10-23, Matt Schulte <matts@commtech-fastcom.com> wrote:
> On Fri, Oct 19, 2012 at 4:21 PM, Theodore Ts'o <tytso@mit.edu> wrote:
>> On Wed, Oct 17, 2012 at 03:24:06PM -0500, Matt Schulte wrote:
>>>
>>> What I would need to have happen is for a flag to be set that says
>>> "enable 9-bit mode transmit" or something.  When set, the first byte
>>> of data of a given write will go out with mark parity, then the rest
>>> of the bytes will have space parity.  The one byte with mark is the
>>> address byte.
>
>> Why not just have a userspace library routine which simply uses
>> tcgetattr/tcsetattr to toggle a bit (probably a new bit in c_cflag) in
>> struct termios, issue the single byte write, and then toggle the bit
>> back?

Can it be guaranteed that it's going to be fast enough at high
baud-rates to prevent any gap between the first byte and subsequent
bytes?

>> In Linux, system calls are fast, so using some extra ioctl's to toggle
>> the termios structure is probably the way to go.  It's also much more
>> general since it doesn't presume a very specific protocol (i.e., your
>> magic multi-drop protocol).
>
> Also this isn't a magic protocol, it is a very old protocol that is
> apparently still in use today because I have several customers still
> asking for it.

It's been pretty much a standard in the microctroller world for a
_long_ time.  I remember 9-bit HW support with special interrupt on
the "address" byte from almost 30 years ago: when I first worked with
the Intel 8051 in 1983, it had 9-bit support in it's UART (and we used
it).

I now work for a company that has manufactured PC serial boards for
25+ years, and we still get regular requests for that feature (and our
boards do support it -- though our Linux driver does not).

-- 
Grant Edwards               grant.b.edwards        Yow! I guess you guys got
                                  at               BIG MUSCLES from doing too
                              gmail.com            much STUDYING!

Re: New serial card development

[Alan Cox]

> Can it be guaranteed that it's going to be fast enough at high
> baud-rates to prevent any gap between the first byte and subsequent
> bytes?

Possibly not for some protocols (or worse yet 'almost always' for some
protocols).

It's something to look at once the basic bits are in.

> I now work for a company that has manufactured PC serial boards for
> 25+ years, and we still get regular requests for that feature (and our
> boards do support it -- though our Linux driver does not).

In which case when we get to addressing this it will be good to make
sure we cover your needs as well.

Alan
 

Re: New serial card development

[Greg KH]

On Tue, Oct 23, 2012 at 11:31:41AM -0500, Matt Schulte wrote:
> On Tue, Oct 23, 2012 at 11:27 AM, Matt Schulte
> <matts@commtech-fastcom.com> wrote:
> > On Fri, Oct 19, 2012 at 4:21 PM, Theodore Ts'o <tytso@mit.edu> wrote:
> >> Alan's advice to get your card working as a basic serial card is very
> >> good one.  Get basic functionality working, and then you can add the
> >> support for the extra bits later....
> >
> > I can see the logic in getting it working as a basic serial card
> > first.  I think at minimum I would still need to implement the extra
> > divisor calculations to get accurate bit rates.
> >
> > So when it works as a basic serial card, I assume you would want me to
> > use the default PCI IDs to keep it more generic.  Then would I add my
> > own PCI IDs and refer them back to the generic port?
> >
> 
> As more of a procedural question, when I go to make the patch(es) to
> submit which, kernel repo do I start with?  Do I start with Greg KH's
> tty repo and then generate the patches and submit them here?

That would be a great place to work off of, use the tty-next branch
please.

thanks,

greg k-h

Re: New serial card development

[Grant Edwards]

On 2012-10-23, Alan Cox <alan@lxorguk.ukuu.org.uk> wrote:
>> Can it be guaranteed that it's going to be fast enough at high
>> baud-rates to prevent any gap between the first byte and subsequent
>> bytes?
>
> Possibly not for some protocols (or worse yet 'almost always' for
> some protocols).
>
> It's something to look at once the basic bits are in.
>
>> I now work for a company that has manufactured PC serial boards for
>> 25+ years, and we still get regular requests for that feature (and our
>> boards do support it -- though our Linux driver does not).
>
> In which case when we get to addressing this it will be good to make
> sure we cover your needs as well.

FWIW, in some products we're planning that will require support for
various industrial serial protocols, I'm leaning towards abandoning
the tty driver approach and writing a stand-alone character device
driver.  The byte-stream oriented tty/line-discipline layer just
doesn't fit well when dealing with frame-oriented industrial protocols
that depend on things like 9th bit addressing and detecting
sub-millisecond inter-byte timeouts.  When I add in the lack of
long-term stability in the tty API it seems like it might not be such
a bad idea to give up trying to make the tty abstraction fit a use
case that's just nothing like a teletype.

-- 
Grant Edwards               grant.b.edwards        Yow! Gibble, Gobble, we
                                  at               ACCEPT YOU ...
                              gmail.com            

Re: New serial card development

[Greg KH]

On Tue, Oct 23, 2012 at 06:45:51PM +0000, Grant Edwards wrote:
> On 2012-10-23, Alan Cox <alan@lxorguk.ukuu.org.uk> wrote:
> >> Can it be guaranteed that it's going to be fast enough at high
> >> baud-rates to prevent any gap between the first byte and subsequent
> >> bytes?
> >
> > Possibly not for some protocols (or worse yet 'almost always' for
> > some protocols).
> >
> > It's something to look at once the basic bits are in.
> >
> >> I now work for a company that has manufactured PC serial boards for
> >> 25+ years, and we still get regular requests for that feature (and our
> >> boards do support it -- though our Linux driver does not).
> >
> > In which case when we get to addressing this it will be good to make
> > sure we cover your needs as well.
> 
> FWIW, in some products we're planning that will require support for
> various industrial serial protocols, I'm leaning towards abandoning
> the tty driver approach and writing a stand-alone character device
> driver.  The byte-stream oriented tty/line-discipline layer just
> doesn't fit well when dealing with frame-oriented industrial protocols
> that depend on things like 9th bit addressing and detecting
> sub-millisecond inter-byte timeouts.  When I add in the lack of
> long-term stability in the tty API it seems like it might not be such
> a bad idea to give up trying to make the tty abstraction fit a use
> case that's just nothing like a teletype.

What do you mean "lack of long-term stability"?  The userspace tty api
hasn't ever changed or broken.  Don't focus on in-kernel api, that's
always going to change, no matter what interface you choose to use in
the kernel.

Just get your drivers merged into the tree and you will not have to
worry about it.

thanks,

greg k-h

Re: New serial card development

[Alan Cox]

> FWIW, in some products we're planning that will require support for
> various industrial serial protocols, I'm leaning towards abandoning
> the tty driver approach and writing a stand-alone character device
> driver.  The byte-stream oriented tty/line-discipline layer just
> doesn't fit well when dealing with frame-oriented industrial protocols
> that depend on things like 9th bit addressing and detecting
> sub-millisecond inter-byte timeouts.  When I add in the lack of
> long-term stability in the tty API it seems like it might not be such
> a bad idea to give up trying to make the tty abstraction fit a use
> case that's just nothing like a teletype.

Not unreasonable but we do need to cover it to some extent because there
are a lot of 'multi-use' port types where you need to share the hardware
or switch modes.

Although we don't use it that way its not entirely accidental that the
tty buffer code supports chains of buffers with lengths 8)

Alan

Re: New serial card development

[Grant Edwards]

On 2012-10-23, Greg KH <greg@kroah.com> wrote:
> On Tue, Oct 23, 2012 at 06:45:51PM +0000, Grant Edwards wrote:
>
>> FWIW, in some products we're planning that will require support for
>> various industrial serial protocols, I'm leaning towards abandoning
>> the tty driver approach and writing a stand-alone character device
>> driver.  The byte-stream oriented tty/line-discipline layer just
>> doesn't fit well when dealing with frame-oriented industrial protocols
>> that depend on things like 9th bit addressing and detecting
>> sub-millisecond inter-byte timeouts.  When I add in the lack of
>> long-term stability in the tty API it seems like it might not be such
>> a bad idea to give up trying to make the tty abstraction fit a use
>> case that's just nothing like a teletype.
>
> What do you mean "lack of long-term stability"?  The userspace tty api
> hasn't ever changed or broken.

I meant the in-kernel api.

> Don't focus on in-kernel api,

It's my job to focus on the in-kernel api.

> that's always going to change, no matter what interface you choose to
> use in the kernel.

Maybe it's just my perception, but the the tty API seems to change a
more than the plain character-device API.

-- 
Grant Edwards               grant.b.edwards        Yow! I wish I was on a
                                  at               Cincinnati street corner
                              gmail.com            holding a clean dog!

Re: New serial card development

[Grant Edwards]

On 2012-10-23, Alan Cox <alan@lxorguk.ukuu.org.uk> wrote:
>> FWIW, in some products we're planning that will require support for
>> various industrial serial protocols, I'm leaning towards abandoning
>> the tty driver approach and writing a stand-alone character device
>> driver.  The byte-stream oriented tty/line-discipline layer just
>> doesn't fit well when dealing with frame-oriented industrial protocols
>> that depend on things like 9th bit addressing and detecting
>> sub-millisecond inter-byte timeouts.  When I add in the lack of
>> long-term stability in the tty API it seems like it might not be such
>> a bad idea to give up trying to make the tty abstraction fit a use
>> case that's just nothing like a teletype.
>
> Not unreasonable but we do need to cover it to some extent because there
> are a lot of 'multi-use' port types where you need to share the hardware
> or switch modes.

Agreed. Providing support for things like 9 bit mode, inter-byte
timeouts, arbitrary baud rates, half-duplex mode, and user-selectable
electrical interfaces (232/422/485/etc.) in the standard tty API would
be a good thing. Half-duplex mode (sometimes called RS485 mode) and
arbitrary baud rate are great recent additions.

> Although we don't use it that way its not entirely accidental that the
> tty buffer code supports chains of buffers with lengths 8)

Thanks, that's good to know.

-- 
Grant Edwards               grant.b.edwards        Yow! I can't decide which
                                  at               WRONG TURN to make first!!
                              gmail.com            I wonder if BOB GUCCIONE
                                                   has these problems!

Re: New serial card development

[Greg KH]

On Tue, Oct 23, 2012 at 07:42:28PM +0000, Grant Edwards wrote:
> On 2012-10-23, Greg KH <greg@kroah.com> wrote:
> > On Tue, Oct 23, 2012 at 06:45:51PM +0000, Grant Edwards wrote:
> >
> >> FWIW, in some products we're planning that will require support for
> >> various industrial serial protocols, I'm leaning towards abandoning
> >> the tty driver approach and writing a stand-alone character device
> >> driver.  The byte-stream oriented tty/line-discipline layer just
> >> doesn't fit well when dealing with frame-oriented industrial protocols
> >> that depend on things like 9th bit addressing and detecting
> >> sub-millisecond inter-byte timeouts.  When I add in the lack of
> >> long-term stability in the tty API it seems like it might not be such
> >> a bad idea to give up trying to make the tty abstraction fit a use
> >> case that's just nothing like a teletype.
> >
> > What do you mean "lack of long-term stability"?  The userspace tty api
> > hasn't ever changed or broken.
> 
> I meant the in-kernel api.
> 
> > Don't focus on in-kernel api,
> 
> It's my job to focus on the in-kernel api.

It's my job to ensure that you don't have to.  Why are you caring?  Are
you trying to keep drivers outside of the kernel tree?  If so, there's
nothing we can do, except point out what a bad idea that really is to
try to do.

> > that's always going to change, no matter what interface you choose to
> > use in the kernel.
> 
> Maybe it's just my perception, but the the tty API seems to change a
> more than the plain character-device API.

Recently, yes.  But, once that churn is over, it should settle down.

Oh, and watch out, the "plain" character-device api is going to change
in the next year or so, I've been working on lots of fixups in that
area, and hope to publish something in a month or so if I get it cleaned
up.

My point is, all of the kernel changes, all the time, so don't use the
lack of change, or the rate of change, for a specific api, as any
indication that it will not change again in the future, possibly in very
drastic ways.

thanks,

greg k-h

Re: New serial card development

[Theodore Ts'o]

On Tue, Oct 23, 2012 at 06:45:51PM +0000, Grant Edwards wrote:
> FWIW, in some products we're planning that will require support for
> various industrial serial protocols, I'm leaning towards abandoning
> the tty driver approach and writing a stand-alone character device
> driver.  The byte-stream oriented tty/line-discipline layer just
> doesn't fit well when dealing with frame-oriented industrial protocols
> that depend on things like 9th bit addressing and detecting
> sub-millisecond inter-byte timeouts.  

You might want to take a look at how the ppp line discipline (which is
obviously highly packet oriented) works.  One advantage of doing
things that way is if you have a multiport serial card, where some
ports want to play this packet-oriented approach, and other cards
might be connected to a traditional modem or terminal sort of device,
you won't have to figure out how to share a PCI board between two
different drivers, allocating one port to the traditional tty/serial
driver, and another one to your new driver, etc.

Regards,

						- Ted

Re: New serial card development

[Grant Edwards]

On 2012-10-23, Theodore Ts'o <tytso@mit.edu> wrote:
> On Tue, Oct 23, 2012 at 06:45:51PM +0000, Grant Edwards wrote:

>> FWIW, in some products we're planning that will require support for
>> various industrial serial protocols, I'm leaning towards abandoning
>> the tty driver approach and writing a stand-alone character device
>> driver.  The byte-stream oriented tty/line-discipline layer just
>> doesn't fit well when dealing with frame-oriented industrial protocols
>> that depend on things like 9th bit addressing and detecting
>> sub-millisecond inter-byte timeouts.  
>
> You might want to take a look at how the ppp line discipline (which
> is obviously highly packet oriented) works.  One advantage of doing
> things that way is if you have a multiport serial card, where some
> ports want to play this packet-oriented approach, and other cards
> might be connected to a traditional modem or terminal sort of device,
> you won't have to figure out how to share a PCI board between two
> different drivers, allocating one port to the traditional tty/serial
> driver, and another one to your new driver, etc.

Thanks for the suggestion.  I did look into line-disciplines briefly
quite a while back. Most of the information I found implied that
adding a line disciplilne wasn't something you could do at run-time by
loading a user-built module -- it would require recompiling parts of
the kernel.  Articles made statements about having to modify kernel
include files like tty.h. If that's wrong, then a line discipline
might be a good option.

-- 
Grant Edwards               grant.b.edwards        Yow! I'm having a
                                  at               quadrophonic sensation
                              gmail.com            of two winos alone in a
                                                   steel mill!

Re: New serial card development

[Matt Schulte]

On Tue, Oct 23, 2012 at 1:38 PM, Greg KH <greg@kroah.com> wrote:
> On Tue, Oct 23, 2012 at 11:31:41AM -0500, Matt Schulte wrote:
>> On Tue, Oct 23, 2012 at 11:27 AM, Matt Schulte
>> <matts@commtech-fastcom.com> wrote:
>> > On Fri, Oct 19, 2012 at 4:21 PM, Theodore Ts'o <tytso@mit.edu> wrote:
>> >> Alan's advice to get your card working as a basic serial card is very
>> >> good one.  Get basic functionality working, and then you can add the
>> >> support for the extra bits later....
>> >
>> > I can see the logic in getting it working as a basic serial card
>> > first.  I think at minimum I would still need to implement the extra
>> > divisor calculations to get accurate bit rates.
>> >
>> > So when it works as a basic serial card, I assume you would want me to
>> > use the default PCI IDs to keep it more generic.  Then would I add my
>> > own PCI IDs and refer them back to the generic port?
>> >

I am working to get the Exar devices added to the 8250 driver.  I
think I get most of what I need to make happen until I get to the
detection of the EFR register.  Can someone explain to me what is
going on here and where the Exar device is supposed to end up?  It
seems to me that I would expect them to show up somewhere around :

/*
 * Exar uarts have EFR in a weird location
 */

But instead I am getting my UARTs detected as EFRv1, iir=3 and marked
as type=ST16650.

For reference the Exar chips have EFR located at an address of 1001
(UART_XR_EFR).  There isn't anything special that has to be done to
access this location, just look at 1001.

Matt Schulte

Re: New serial card development

[Matt Schulte]

On Mon, Oct 29, 2012 at 3:04 PM, Matt Schulte
<matts@commtech-fastcom.com> wrote:
> On Tue, Oct 23, 2012 at 1:38 PM, Greg KH <greg@kroah.com> wrote:
>> On Tue, Oct 23, 2012 at 11:31:41AM -0500, Matt Schulte wrote:
>>> On Tue, Oct 23, 2012 at 11:27 AM, Matt Schulte
>>> <matts@commtech-fastcom.com> wrote:
>>> > On Fri, Oct 19, 2012 at 4:21 PM, Theodore Ts'o <tytso@mit.edu> wrote:
>>> >> Alan's advice to get your card working as a basic serial card is very
>>> >> good one.  Get basic functionality working, and then you can add the
>>> >> support for the extra bits later....
>>> >
>>> > I can see the logic in getting it working as a basic serial card
>>> > first.  I think at minimum I would still need to implement the extra
>>> > divisor calculations to get accurate bit rates.
>>> >
>>> > So when it works as a basic serial card, I assume you would want me to
>>> > use the default PCI IDs to keep it more generic.  Then would I add my
>>> > own PCI IDs and refer them back to the generic port?
>>> >
>
> I am working to get the Exar devices added to the 8250 driver.  I
> think I get most of what I need to make happen until I get to the
> detection of the EFR register.  Can someone explain to me what is
> going on here and where the Exar device is supposed to end up?  It
> seems to me that I would expect them to show up somewhere around :
>
> /*
>  * Exar uarts have EFR in a weird location
>  */
>
> But instead I am getting my UARTs detected as EFRv1, iir=3 and marked
> as type=ST16650.
>
> For reference the Exar chips have EFR located at an address of 1001
> (UART_XR_EFR).  There isn't anything special that has to be done to
> access this location, just look at 1001.

I was able to work my way through the funky EFR detection.  Hopefully
you guys will like what I did when you see the patch.

Next I need to know the channel number of a given UART port.  If I
have a 4 port UART I need to know that a given channel is 0,1,2 or 3.
If I have an 8 port UART I need to know if a given channel is
0,1,2,3,4,5,6 or 7.

The only way that I can find to get this number is using the idx
variable that is passed to the port's .setup operation.  Would I be
able to add a member to the uart_port structure so that I could store
this number inside my UARTs setup function?

Something like:

struct uart_port {
.
.
.
unsigned int channel;
};

pci_myuart_setup(struct serial_private *priv,
                const struct pciserial_board *board,
                struct uart_8250_port *port, int idx)
{
...
        port->port.channel = idx;
        return pci_default_setup(priv, board, port, idx);
}

Thank you,
Matt Schulte

Re: New serial card development

[Matt Schulte]

On Wed, Oct 31, 2012 at 4:55 PM, Matt Schulte
<matts@commtech-fastcom.com> wrote:
> On Mon, Oct 29, 2012 at 3:04 PM, Matt Schulte
> <matts@commtech-fastcom.com> wrote:
>> On Tue, Oct 23, 2012 at 1:38 PM, Greg KH <greg@kroah.com> wrote:
>>> On Tue, Oct 23, 2012 at 11:31:41AM -0500, Matt Schulte wrote:
>>>> On Tue, Oct 23, 2012 at 11:27 AM, Matt Schulte
>>>> <matts@commtech-fastcom.com> wrote:
>>>> > On Fri, Oct 19, 2012 at 4:21 PM, Theodore Ts'o <tytso@mit.edu> wrote:
>>>> >> Alan's advice to get your card working as a basic serial card is very
>>>> >> good one.  Get basic functionality working, and then you can add the
>>>> >> support for the extra bits later....
>>>> >
>>>> > I can see the logic in getting it working as a basic serial card
>>>> > first.  I think at minimum I would still need to implement the extra
>>>> > divisor calculations to get accurate bit rates.
>>>> >
>>>> > So when it works as a basic serial card, I assume you would want me to
>>>> > use the default PCI IDs to keep it more generic.  Then would I add my
>>>> > own PCI IDs and refer them back to the generic port?

Hello, the next question that I have is how to handle the interrupts
on this UART.  It has a more complex interrupt scheme than a
traditional 8250 UART.  There is a global ISR register that tells you
which port had the interrupt and then there is another set of
registers that hold the actual interrupts.

In this situation where this UART needs extra code for the handling of
its interrupt, where is the best place to add this special interrupt
handler?  Directly to the 8250.c default handler or somewhere else?

Thank you,
Matt Schulte

Re: New serial card development

[Alan Cox]

> Hello, the next question that I have is how to handle the interrupts
> on this UART.  It has a more complex interrupt scheme than a
> traditional 8250 UART.  There is a global ISR register that tells you
> which port had the interrupt and then there is another set of
> registers that hold the actual interrupts.

port->handle_irq is half of what you need, but you probably need to be
able to hook serial8250_interruot and provide your own alternative. Given
you don't need the irq chain handling either probably the bit that wants
to be hookable is the calls to serial_link_irq_chain and
serial_unlink_irq_chain in the startup/shutdown methods.

In yor case all that and all the irq scanning the chain becomes a simple
read of the port and call of the handlers (serial8250_handlee_irq(port,
iir)



Alan

Re: New serial card development

[Matt Schulte]

On Thu, Nov 1, 2012 at 5:26 PM, Alan Cox <alan@lxorguk.ukuu.org.uk> wrote:
>> Hello, the next question that I have is how to handle the interrupts
>> on this UART.  It has a more complex interrupt scheme than a
>> traditional 8250 UART.  There is a global ISR register that tells you
>> which port had the interrupt and then there is another set of
>> registers that hold the actual interrupts.
>
> port->handle_irq is half of what you need, but you probably need to be
> able to hook serial8250_interruot and provide your own alternative. Given
> you don't need the irq chain handling either probably the bit that wants
> to be hookable is the calls to serial_link_irq_chain and
> serial_unlink_irq_chain in the startup/shutdown methods.
>
> In yor case all that and all the irq scanning the chain becomes a simple
> read of the port and call of the handlers (serial8250_handlee_irq(port,
> iir)

To be clear, I shouldn't put a custom irq handler inside 8250.c, correct?

I would want to do something along the vein of 8250_dw.c with a custom
probe and a custome handle_irq?

Matt Schulte

Re: New serial card development

[Alan Cox]

On Fri, 2 Nov 2012 13:47:25 -0500
Matt Schulte <matts@commtech-fastcom.com> wrote:

> On Thu, Nov 1, 2012 at 5:26 PM, Alan Cox <alan@lxorguk.ukuu.org.uk> wrote:
> >> Hello, the next question that I have is how to handle the interrupts
> >> on this UART.  It has a more complex interrupt scheme than a
> >> traditional 8250 UART.  There is a global ISR register that tells you
> >> which port had the interrupt and then there is another set of
> >> registers that hold the actual interrupts.
> >
> > port->handle_irq is half of what you need, but you probably need to be
> > able to hook serial8250_interruot and provide your own alternative. Given
> > you don't need the irq chain handling either probably the bit that wants
> > to be hookable is the calls to serial_link_irq_chain and
> > serial_unlink_irq_chain in the startup/shutdown methods.
> >
> > In yor case all that and all the irq scanning the chain becomes a simple
> > read of the port and call of the handlers (serial8250_handlee_irq(port,
> > iir)
> 
> To be clear, I shouldn't put a custom irq handler inside 8250.c, correct?
> 
> I would want to do something along the vein of 8250_dw.c with a custom
> probe and a custome handle_irq?

That's my first guess - there is an awful lot of stuff you simply don't
need in this situation.