Re: [PATCH 0/3] serdev support for n_gsm

From: Johan Hovold
Date: Thu Jan 24 2019 - 11:39:45 EST

Hi Tony,

and sorry about the reply latency. This is quite a lot to think about.

I'm also adding Kishon and Mark on CC (e.g. for the phy and ASoC bits).

On Mon, Jan 21, 2019 at 09:01:16AM -0800, Tony Lindgren wrote:
> Hi,
> * Johan Hovold <johan@xxxxxxxxxx> [190121 10:57]:
> > Adding Marcel on CC.
> >
> > On Fri, Jan 18, 2019 at 12:59:58PM +0100, Greg Kroah-Hartman wrote:
> > > On Sun, Jan 13, 2019 at 05:25:25PM -0800, Tony Lindgren wrote:
> > > > Hi all,
> > > >
> > > > Here's a series of patches to add initial serdev support to n_gsm
> > > > TS 27.010 line discipline.
> > > >
> > > > This allows handling vendor specific protocols on top of TS 27.010 and
> > > > allows creating simple serdev drivers where it makes sense. So far I've
> > > > tested it with droid 4 for it's modem to provide char devices for AT
> > > > ports, modem PM support, and serdev drivers for GNSS and Alsa ASoC.
> > > >
> > > > I'll be posting the related MFD, GNSS and Alsa ASoC drivers separately.
> > > > For reference, the MFD driver is at [0], the GNSS driver at [1], and
> > > > the Alsa ASoC driver at [2] below.
> > >
> > > I have applied the first two patches to my tree, as those are nice
> > > cleanups.
> > >
> > > The last one I want some feedback from the serdev developers to verify
> > > all is set up properly, and Johan, to see if this ends up conflicting
> > > with the gnss code, as that would not be good.
> >
> > I think we need to have a discussion about how to model modems generally
> > before getting into implementation details.
> >
> > Modems are currently managed by user space (e.g. through ofono) and
> > I'm not sure that moving only parts of its responsibilities into the
> > kernel is necessarily the right thing to do. You still need coordination
> > between the various components for things like power management.
> At least now we do have the option of doing kernel drivers or user
> space apps whichever way we want to :)
> And we can now do the user space apps without having to implement
> any of the Motorola custom packet numbering layer on top of TS 27.010
> for each app.
> For some user space examples, I have posted scripts to send and receive
> SMS at [3], and Pavel has ofono patches [4] below. Seems like we can
> also add support to ModemManager along the similar lines. And for the
> serdev drivers, those implement standard Linux interfaces for apps
> to use.
> For PM, about a year ago I tried making things work with a user space
> solution and it sucked big time[5]. The power management makes sense
> to do in the kernel driver at least in this case as there are shared
> GPIO pins between the USB PHY and TS 27.010 UART. The shared GPIOs
> are handled by the phy-mapphone-mdm6600 driver.
> With the serdev n_gsm MFD driver, the only thing that needs to be done
> to idle the modem is to enable autosuspend for the OHCI interface. So
> no spefific coordination between various components is needed for PM
> beyond that. Things idle just fine using PM runtime.

Yeah, I don't envy you trying to get this to work (and now I'm getting
dragged into it ;) ).

It would really help with a high-level outline of the modem and its
components. I've done my best to derive it from these patches and the
code you link to, but that info needs to go in the patch descriptions
(or cover letter).

This series adds a new interface (gsm_serdev) but no users, which is not
something we normally accept. I think you need to post the lot, at least
as an RFC in order for the full picture to be visible.

Similarly, including an example device tree would also help with the
overall picture. I've been able to derive the following (from your code
and already merged phy driver binding doc):

mdm6600_phy: usb-phy {
compatible = "motorola,mapphone-mdm6600";
enable-gpios = <&gpio3 31 GPIO_ACTIVE_LOW>;
power-gpios = <&gpio2 22 GPIO_ACTIVE_HIGH>;
reset-gpios = <&gpio2 17 GPIO_ACTIVE_HIGH>;
motorola,mode-gpios = <&gpio5 20 GPIO_ACTIVE_HIGH>,
<&gpio5 21 GPIO_ACTIVE_HIGH>;
motorola,cmd-gpios = <&gpio4 7 GPIO_ACTIVE_HIGH>,
<&gpio4 8 GPIO_ACTIVE_HIGH>,
<&gpio5 14 GPIO_ACTIVE_HIGH>;
motorola,status-gpios = <&gpio2 20 GPIO_ACTIVE_HIGH>,
<&gpio2 21 GPIO_ACTIVE_HIGH>,
<&gpio2 23 GPIO_ACTIVE_HIGH>;
#phy-cells = <0>;

&uart {
gsm-mux? {
compatible = "motorola,mapphone-mdm6600-serdev";
phy = <&mdm6600_phy>;

gnss {
compatible = "motorola,mapphone-mdm6600-gnss";

audio-codec {


&ohci {
phys = <&mdm6600_phy>;

Which brings us back to the question of how to model modems. You've
already added a phy-driver for something which really is some sort of
modem representation.

I'm not saying I have a solution for this, but again, I think this needs
to be discussed before merging more code.

> > TS 27.010 may make it seem like we can move everything into the kernel,
> > but Tony's to-be-posted Motorola MFD driver is still exposing character
> > devices for most of the muxed ports. If I understand things correctly,
> > there also still needs to be some coordination with USB over which some
> > channels are handled (e.g. IP over USB, gnss over muxed UART).
> Hmm yes now we can do either user space daemons or kernel serdev
> drivers.
> For USB, the modem data connection already works with USB OHCI over
> QMI. So it's is already handled and separated out of this. The USB
> PHY and TS 27.010 UART have shared GPIO pins handled by the USB PHY
> driver. The USB PHY is integrated into the modem with the shared
> GPIO pins controlling the PHY and the TS 27.010 UART PM..
> But it's working for PM and like I mentioned modem PM works as long
> as OHCI is set to autosuspend. Well the modem also wants to see
> TS 27.010 connected before idling.
> > Instead of adding these extra layers, only to export most ports to user
> > space again, it may be better to hook into the various kernel subsystems
> > through dedicated user-space-implementation interfaces such as the
> > suggested ugnss interface, which means that user space feeds gnss data
> > into the kernel which in turn makes it available through a standard
> > interface.
> Sure that's doable. But notice that we actually need to kick the
> serdev GNSS interface to get more data. It's not a passive GNSS
> data feed in this case. So it's not going to be just a case of
> cat /dev/motmdm4 > /dev/ugnss. Without the serdev GNSS driver,
> it would be some-custom-app -i /dev/motmdm4 -o /dev/ugnss.

Yeah, I remember us discussing that briefly off list.

> And without the n_gsm serdev support, it's a mess of some app
> similar to [5] initializing n_gsm, trying to deal with the USB
> PHY PM, dealing with Motorola custom packet numbering, kicking
> GNSS device, feeding data to /dev/ugnss. Hmm I think I've already
> been there just to be able to type AT commands to the modem and
> it did not work :)

It's a mess indeed, but I'd rather see user-space dealing with until we
figure out how best to do it in the kernel. ;)

> Anyways, for the serdev kernel drivers, the criteria I've tried
> to follow is: "Can this serdev device driver make user space
> apps use standard Linux interfaces for the hardware?"
> So for the serdev Alsa ASoC driver, user space can use the standard
> Alsa interface for setting voice call volume. And for the serdev
> GNSS driver, user space can use /dev/gnss0.

I understand. Both drivers appears to be using AT commands for control.
It would be interesting to hear what Mark has to say about the codec
driver too. Moving AT handling into the kernel scares me a bit. If we
already have a telephony stack to deal with it in user-space, my
inclination is to let it continue to handle it.

Modem-managed GNSS is also different from receivers connected directly
to the host. It's really the modem that drives the GNSS receiver, and
offers a higher-level interface to the host, for example, by buffering
output which the host can later request. It may or may not be the
kernel's job to periodically poll the modem to recreate an NMEA feed so
to speak.

But the end-result of having it accessible through a standard interface
is of course appealing.

I have some more comments on the gsm_serdev interface and motorola mfd,
but let's start with the above.


> [0]
> [1]
> [2]
> [3]
> [4]
> [5]