Re: [PATCH v2 00/16] introduce generic IOCTL interface for RPMsg channels management

[Arnaud POULIQUEN]

Hi Mathieu,

On 1/13/21 9:31 PM, Mathieu Poirier wrote:
> Hi Arnaud,
> 
> [...]
> 
>>
>> Arnaud Pouliquen (16):
>>   rpmsg: introduce RPMsg control driver for channel creation
>>   rpmsg: add RPMsg control API to register service
>>   rpmsg: add override field in channel info
>>   rpmsg: ctrl: implement the ioctl function to create device
>>   rpmsg: ns: initialize channel info override field
>>   rpmsg: add helper to register the rpmsg ctrl device
>>   rpmsg: char: clean up rpmsg class
>>   rpmsg: char: make char rpmsg a rpmsg device without the control part
>>   rpmsg: char: register RPMsg raw service to the ioctl interface.
>>   rpmsg: char: allow only one endpoint per device
>>   rpmsg: char: check destination address is not null
>>   rpmsg: virtio: use the driver_override in channel creation ops
>>   rpmsg: virtio: probe the rpmsg_ctl device
>>   rpmsg: glink: add create and release rpmsg channel ops
>>   rpmsg: smd: add create and release rpmsg channel ops
>>   rpmsg: replace rpmsg_chrdev_register_device use
>>
>>  drivers/rpmsg/Kconfig             |   8 +
>>  drivers/rpmsg/Makefile            |   1 +
>>  drivers/rpmsg/qcom_glink_native.c |  96 +++++++--
>>  drivers/rpmsg/qcom_smd.c          |  59 +++++-
>>  drivers/rpmsg/rpmsg_char.c        | 246 ++++++-----------------
>>  drivers/rpmsg/rpmsg_ctrl.c        | 320 ++++++++++++++++++++++++++++++
>>  drivers/rpmsg/rpmsg_internal.h    |  14 --
>>  drivers/rpmsg/rpmsg_ns.c          |   1 +
>>  drivers/rpmsg/virtio_rpmsg_bus.c  |  38 +++-
>>  include/linux/rpmsg.h             |  40 ++++
>>  include/uapi/linux/rpmsg.h        |  14 ++
>>  11 files changed, 606 insertions(+), 231 deletions(-)
>>  create mode 100644 drivers/rpmsg/rpmsg_ctrl.c
> 
> I am finally coming around to review this set.  I see that you already had an
> extensive conversation with Bjorn - did you want me to have a look as well or
> should I wait for the next revision?

Based on Bjorn first feedback, my understanding is that the management based on
create/destroy channel does not match with the QCOM RPMsg backend
implementation. I think this is the blocking point of my V2 implementation.

Before sending a new revision i would hope that we have a roundtable discussion
to clarify the direction to move forward, to avoid sending useless revisions.

As discussed in [1], there are different alternatives, that probably depend on
the features we expect to support.
I tried to sum-up the requirement I have in mind in [1].

The 2 main directions I can see are:
- rework the rpmsg_char to match with all rpmsg backend (V2 implementation)
    to be honest i don't know how to move forward in this direction as QCOM and
    virtio backends are rather different.
- not modify the rpmsg_char but create the rpmsg_ctrl (and perhaps also a
rpmsg_raw for a /dev/rpmsg data interface) that would use the create/destroy
channel such as the rpmsg ns (V1 implementation).
    one advantage of this solution is that this does not impact QCOM drivers.
    one drawback is that we duplicate the code.

[1]
https://patchwork.kernel.org/project/linux-remoteproc/patch/20201222105726.16906-5-arnaud.pouliquen@xxxxxxxxxxx/

[2] https://patchwork.kernel.org/project/linux-remoteproc/list/?series=327277

Thanks,
Arnaud

> 
> Thanks,
> Mathieu
> 
>>
>> -- 
>> 2.17.1
>>