On Tue, 2017-08-08 at 15:42 -0700, Bjorn Andersson wrote:
> On Tue 08 Aug 04:02 PDT 2017, Bj?rn Mork wrote:
> > Bjorn Andersson <bjorn.anders...@linaro.org> writes:
> > > This series starts by moving the common definitions of the QMUX
> > > protocol to the
> > > uapi header, as they are shared with clients - both in kernel and
> > > userspace.
> > >
> > > This series then introduces in-kernel helper functions for aiding
> > > the handling
> > > of QMI encoded messages in the kernel. QMI encoding is a wire-
> > > format used in
> > > exchanging messages between the majority of QRTR clients and
> > > services.
> > Interesting! I tried to add some QMI handling in the kernel a few
> > years
> > ago, but was thankfully voted down. See
> > https://www.spinics.net/lists/netdev/msg183101.html and the
> > following
> > discussion. I am convinced that was the right decision, for the
> > client
> > side at least. The protocol is just too extensive and ever-growing
> > to be
> > implemented in the kernel. We would be catching up forever.
> > Note that I had very limited knowledge of the protocol at the time
> > I
> > wrote that driver. Still have, in fact :-)
> Thanks for the pointer, I definitely think there's more work to be
> here to figure out the proper way to interact with these devices.
> But I think that Dan's reply shows a huge source of confusion here;
> acronym "QMI" covers a large amount of different things - and means
> different things for different people.
I would agree, sorry for any confusion caused. Great discussion so
> In the modem world QMI seems to mean a defined set of logical
> that accepts TLV-encoded messages to do modem-related things. But the
> TLV-encoding is used for non-modem related services and the only
> denominator of everything called QMI is the TLV-encoding.
> Due to my limited exposure to the USB attached "QMI thingies" I
> previously looked into the exact differences. The proposed patches
> to support implementing a few non-modem-related clients using
> QMI-encoded messages over ipcrouter.
> Looking at your patch above, and oPhono, seems to highlight a few
> important differences that will take some thinking to overcome.
> = Transport
> The transport header in the USB case is your struct qmux, which
> the type of message (in "flags") and the transaction id. The
> in the QMUX header matches the service id being communicated with.
> in order to communicate with a service it seems like one requests a
> client-id from the control service.
Correct. You cannot talk to a service on the modem without getting an
allocated client ID from the CTL service, which has a well-defined
> In the smartphone world (with shared memory communication) the
> is ipcrouter - with a header very similar to UDP - and there's no
> information about the payload, it provides only the means of
Can you explain a bit about the relationship of SMD to [I/R]PC, qrtr,
and QMI? A couple years ago there was smd_qmi.c (like for the Nexus 4
with APQ8064 and a discrete MDM9215) which from a 10 minute fresh look
appears to just push QMUX+QMI via SMD rather than being backed by the
RPC/IPC stuff. I could be wrong, there's a lot of indirection there
and it may well end up going over the router. But that's buried deeper
than a 10m look for me.
Is it perhaps only with on-chip blocks where the QMUX/QMI/qrtr/irpc
stuff you describe here is used? If so, perhaps that's the distinction
to be made. I'll let you correct me here since you clearly know more
than I about the internals of these devices.
> messages from one address/port to another address/port. A typical
> smartphone has 3-4 nodes (modem, sensors, audio etc) and ports are
> dynamically allocated. The control messages in the QMUX protocol (not
> the same QMUX protocol as in the USB case!) are used for clients to
> the mapping from service id to a port on the given address. The
> port is dynamically allocated in this case.
> = QMI-encoded messages
> The list of TLV-entries have a "QMI header" prepended in both cases,
> in the QMUX case the header consists only of "msgid" and length.
> In the ipcrouter case the transport doesn't carry any information
> regarding the payload, so the header prepended the TLV entries
> "type", transaction id, "msg_id" and length.
I'll assume that in this case, because the client has already found out
how to contact the target service directly, that it has no use for a
"fat" QMUX header that includes the client ID and service stuff.
I don't really have an issue with the kernel doing "thin" QMUX-related
stuff. That's pretty simple.
> It looks as if once past the differences in the transport and QMI
> message header the messages (TLV-encoded data) are the same. But I'm
> yet sure about how we can hide the transport differences.
QMI itself is really just a header + TLVs. So it makes sense that it
would be loosely repurposed since it's pretty generic. All the
interesting stuff is actually in the services themselves and the
messages that the services respond to.
What's confusing me a lot so far is exactly *how* QMI itself gets used
in-kernel. Instead of the samples that you've provided in this
patchset, could you point me to an actual in-kernel driver that would
use something like qmi_send_message()? I can't properly evaluate
whether I have further comments on your approach without some specific