On Fri, Feb 22, 2019 at 10:48 PM Pierre-Louis Bossart pierre-louis.bossart@linux.intel.com wrote:
On 2/22/19 2:32 AM, xiang xiao wrote:
On Thu, Feb 21, 2019 at 11:27 PM Pierre-Louis Bossart pierre-louis.bossart@linux.intel.com wrote:
Should we utilize official IPC frameowrk instead reinverting the wheel? 1.Load firmware by drivers/remoteproc https://www.kernel.org/doc/Documentation/remoteproc.txt 2.Do the comunication through drivers/rpmsg https://www.kernel.org/doc/Documentation/rpmsg.txt Many vendor(TI, Qualcomm, ST, NXP, Xilinx...) migrate to remoteproc/rpmsg, why Intel provide an other IPC mechanism?
It definitely makes more sense to use rpmsg for Generic IPC driver here.
Qualcomm DSP audio drivers (non SOF) already use rpmsg. This will definitely help everyone in future while immigrating to SOF.
Actually, Xiaomi also build DSP audio driver on top of rpmsg, but fully integrate with the ASoC topology framework, and the firmware is base on FreeRTOS and OpenMAX. SOF initiative is very good and exciting, our team members(include me) spend a couple weeks to study the current code base on both firmware and kernel side, we even port SOF to our DSP/MCU and make it run, but I have to point out that: SOF IPC is too simple and rigid, tightly couple with Intel platform and audio domain, which make: a.It's difficult to integrate with other vendor SoC, especially if other vendor already adopt remote/rpmsg(this is already a trend!). b.It's difficult to add other IPC services for example: i.Audio DSP talk to power MCU to adjust clock and voltage ii.Export ultrasonic distance measurement to IIO subsystem
The IPC scheme suggested in this patchset is only a first pass that works on 3 generations on Intel platforms + the QEMU parts. There are no claims that the current solution is set-in-stone, and this is already an area where things are already changing to support notifications and low-power transitions.
There will clearly be evolutions to make the IPC more flexible/generic, but we've got to start somewhere and bear in mind that we also have to support memory-constrained legacy devices where such generic frameworks aren't needed or even implementable. Some of your proposals such as changing power/clocks with a firmware request aren't necessarily possible or recommended on all platforms - i can already hear security folks howling, this was already mentioned in the GitHub thread.
Rather than evolve the IPC, i would say it makes more sense that we "reuse" existing upstream frameworks.. As given below by xiang this seems to have support for RTOSes (see point 4 below) and looking at below it seems to have much better coverage across systems.
This should also help in easy adoption of SoF for non Intel people...
Also looking at it, lot of IPC code, DSP loading etc would go away making SoF code lesser in footprint.
I think benefits outweigh the effort of porting to a framework which is already upstream and used on many platforms for different vendors!
There is no free lunch. There are 'features' of RPMsg which aren't necessarily great for all platforms, e.g. the concepts of virtio-like rings for IPC with available/used buffers for both directions are not a good match or replacement for the memory-window-based IPC on Intel platforms, where there is no DDR access, a small window allocated by firmware and only a couple of doorbell registers for essentially serial communication.
rpmsg support to define the custom mechanism(see rpmsg_endpoint_ops in drivers\rpmsg\rpmsg_internal.h) but keep the upper layer API, qcomm utilize this for glink and smd actually.
That's interesting. Can anyone at Qualcomm/Linaro point to actual examples of the implementation, so that we get a better picture of the split between 'upper layer API' and 'custom mechanism'?
The resources embedded in a firmware file is another capability that doesn't align with the way the SOF firmware is generated. I also don't know where the topology file would be handled, nor how to deal with suspend-resume where the DSP needs to be restarted. For folks who need an introduction to RPMsg, the link [1] is the best I found to scope out the work required.
We can share our rpmsg based topology implementation as reference which: 1.About 2500 lines(much less than SOF) 2.Support pcm and compress playback/capture 3.No any vendor dependence(thanks for rpmsg/remoteproc)
Sure. Where's the code? What's the license?
The code is base on 4.19 kernel, I could upstream the code basing on the latest kernel in the next couple days for reference. the license is GPL, of course.
Most of the SOF code is really in hardware-specific .ops callbacks and topology handling, the generic IPC layer is only ~800 lines of code. rpmsg would allow for easier portability but a significant reduction of the code size is unlikely.
The reduce come from: 1.Move firmware load and dsp start/stop to remoteproc layer. 2.Move IPC buffer/mailbox to rpmsg layer. 3.Reuse ASoC topology parser to generate the audio graph. 4.Reuse ASoC DAMP to control the graph node state change(run/stop/pause/resume). 5.Use the general machine driver glue all individual components.