Hi Loic, On Fri, Jun 04, 2021 at 11:11:45PM +0200, Loic Poulain wrote: > On Wed, 2 Jun 2021 at 20:20, Stephan Gerhold <stephan@xxxxxxxxxxx> wrote: > > I've been thinking about creating some sort of "RPMSG" driver for the > > new WWAN subsystem; this would be used as a QMI/AT channel to the > > integrated modem on some older Qualcomm SoCs such as MSM8916 and MSM8974. > > > > It's easy to confuse all the different approaches that Qualcomm has to > > talk to their modems, so I will first try to briefly give an overview > > about those that I'm familiar with: > > > > --- > > There is USB and MHI that are mainly used to talk to "external" modems. > > > > For the integrated modems in many Qualcomm SoCs there is typically > > a separate control and data path. They are not really related to each > > other (e.g. currently no common parent device in sysfs). > > > > For the data path (network interface) there is "IPA" (drivers/net/ipa) > > on newer SoCs or "BAM-DMUX" on some older SoCs (e.g. MSM8916/MSM8974). > > I have a driver for BAM-DMUX that I hope to finish up and submit soon. > > > > The connection is set up via QMI. The messages are either sent via > > a shared RPMSG channel (net/qrtr sockets in Linux) or via standalone > > SMD/RPMSG channels (e.g. "DATA5_CNTL" for QMI and "DATA1" for AT). > > > > This gives a lot of possible combinations like BAM-DMUX+RPMSG > > (MSM8916, MSM8974), or IPA+QRTR (SDM845) but also other funny > > combinations like IPA+RPMSG (MSM8994) or BAM-DMUX+QRTR (MSM8937). > > > > Simply put, supporting all these in userspace like ModemManager > > is a mess (Aleksander can probably confirm). > > It would be nice if this could be simplified through the WWAN subsystem. > > > > It's not clear to me if or how well QRTR sockets can be mapped to a char > > device for the WWAN subsystem, so for now I'm trying to focus on the > > standalone RPMSG approach (for MSM8916, MSM8974, ...). > > --- > > > > Currently ModemManager uses the RPMSG channels via the rpmsg-chardev > > (drivers/rpmsg/rpmsg_char.c). It wasn't my idea to use it like this, > > I just took that over from someone else. Realistically speaking, the > > current approach isn't too different from the UCI "backdoor interface" > > approach that was rejected for MHI... > > > > I kind of expected that I can just trivially copy some code from > > rpmsg_char.c into a WWAN driver since they both end up as a simple char > > device. But it looks like the abstractions in wwan_core are kind of > > getting in the way here... As far as I can tell, they don't really fit > > together with the RPMSG interface. > > > > For example there is rpmsg_send(...) (blocking) and rpmsg_trysend(...) > > (non-blocking) and even a rpmsg_poll(...) [1] but I don't see a way to > > get notified when the TX queue is full or no longer full so I can call > > wwan_port_txon/off(). > > > > Any suggestions or other thoughts? > > It would be indeed nice to get this in the WWAN framework. > I don't know much about rpmsg but I think it is straightforward for > the RX path, the ept_cb can simply forward the buffers to > wwan_port_rx. Right, that part should be straightforward. > For tx, simply call rpmsg_trysend() in the wwan tx > callback and don't use the txon/off helpers. In short, keep it simple > and check if you observe any issues. > I'm not sure that's a good idea. This sounds like exactly the kind of thing that might explode later just because I don't manage to get the TX queue full in my tests. In that case, writing to the WWAN char dev would not block, even if O_NONBLOCK is not set. But I think you're right that it's probably easiest if I start with that, see if I can get anything working at all ... > And for sure you can propose changes in the WWAN framework if you > think something is missing to support your specific case. > ... and then we can discuss that further on a RFC PATCH or something like that. Does that sound good to you? Thanks! Stephan
