Re: [RESEND PATCH v18 0/3] userspace MHI client interface driver

[Bjorn Andersson]

On Wed 10 Feb 12:41 CST 2021, Jakub Kicinski wrote:

> On Wed, 10 Feb 2021 11:55:31 +0530 Manivannan Sadhasivam wrote:
> > On Tue, Feb 09, 2021 at 08:17:44AM -0800, Jakub Kicinski wrote:
> > > On Tue, 9 Feb 2021 10:20:30 +0100 Aleksander Morgado wrote:  
> > > > This may be a stupid suggestion, but would the integration look less a
> > > > backdoor if it would have been named "mhi_wwan" and it exposed already
> > > > all the AT+DIAG+QMI+MBIM+NMEA possible channels as chardevs, not just
> > > > QMI?  
> > > 
> > > What's DIAG? Who's going to remember that this is a backdoor driver 
> > > a year from now when Qualcomm sends a one liner patches which just 
> > > adds a single ID to open another channel?  
> > 
> > I really appreciate your feedback on this driver eventhough I'm not
> > inclined with you calling this driver a "backdoor interface". But can
> > you please propose a solution on how to make this driver a good one as
> > per your thoughts?
> > 
> > I really don't know what bothers you even if the userspace tools making
> > use of these chardevs are available openly (you can do the audit and see
> > if anything wrong we are doing).
> 
> What bothers me is maintaining shim drivers which just shuttle opaque
> messages between user space and firmware. One of which definitely is,
> and the other may well be, proprietary. This is an open source project,
> users are supposed to be able to meaningfully change the behavior of
> the system.
> 

You're absolutely right in that we in general don't like shim drivers
and there are several examples of proper MHI drivers - for e.g.
networking, WiFi

Technically we could fork/reimplement
https://github.com/freedesktop/libqmi, https://github.com/andersson/diag
and https://github.com/andersson/qdl in the kernel as "proper drivers" -
each one exposing their own userspace ABI.

But to leave these in userspace and rely on something that looks exactly
like USBDEVFS seems like a much better strategy.

> What bothers me is that we have 3 WWAN vendors all doing their own
> thing and no common Linux API for WWAN. It may have been fine 10 years
> ago, but WWAN is increasingly complex and important.
> 

We had a deep discussion and a few prototypes for a WWAN framework going
around 1-1.5 years ago. Unfortunately, what did fit Intel's view of what
a WWAN device is didn't fit at all with what's run and exposed by the
"modem" DSP in a Qualcomm platform. After trying to find various
contrived ways to model this we gave up.

> > And exposing the raw access to the
> > hardware is not a new thing in kernel. There are several existing
> > subsystems/drivers does this as pointed out by Bjorn. Moreover we don't
> > have in-kernel APIs for the functionalities exposed by this driver and
> > creating one is not feasible as explained by many.
> > 
> > So please let us know the path forward on this series. We are open to
> > any suggestions but you haven't provided one till now.
> 
> Well. You sure know how to aggravate people. I said clearly that you
> can move forward on purpose build drivers (e.g. for WWAN). There is no
> way forward on this common shim driver as far as I'm concerned.

But what is a WWAN device? What features does it have? What kind of APIs
does it expose?


Note that in this sense "QMI" really is a "binary equivalent" of AT
commands, the data flows over a DMA engine, which is not part of the
"WWAN device" and other services, such as GPS, already has specific
transports available upstream.

Regards,
Bjorn