Date:   Fri, 25 Feb 2022 22:45:27 +0100
From:   Philipp Rossak <embed3d@...il.com>
To:     Arnaud POULIQUEN <arnaud.pouliquen@...s.st.com>,
        Bjorn Andersson <bjorn.andersson@...aro.org>,
        Mathieu Poirier <mathieu.poirier@...aro.org>
Cc:     linux-remoteproc@...r.kernel.org, linux-kernel@...r.kernel.org,
        linux-stm32@...md-mailman.stormreply.com, julien.massot@....bzh
Subject: Re: [PATCH v9 00/11] Restructure the rpmsg_char driver and introduce
 rpmsg_ctrl driver

Hi Arnaud,

On 24.02.22 09:29, Arnaud POULIQUEN wrote:
> Hi Philipp,
> 
> On 2/23/22 22:28, Philipp Rossak wrote:
>> Hi Arnaud,
>>
>> thanks for working on this! I'm currently testing/using this patch
>> series on my imx7d project because it adds the capability that the
>> remote processor can register it's endpoints dynamically (as mentioned
>> in the objectives).
> 
> Thanks for your feedback on this work!
> Don't hesitate to add your tested-by, this help maintainers for the reviews.
> 
I will do this.
>>
>> After a few tests, debugging, and checking the openamp specification [1]
>> I think that you missed the second ns_announcement that should be sent
>> from linux master to the slave after it created the channel/endpoint.
>> Without this second announcement the remote processor is not able to
>> send messages to the linux master because it doesn't know the
>> destination address until it receives a message from the linux master.
> 
> Yes I detected this issues, it is not related to the series
> but to the remoteproc_virtio backend.
> 
> As you mentioned, after the ns announcement from Linux, the remote processor
> send first messages. But the Linux virtio does not do the match between the
> local channel created and the remote endpoint.
> 

I'm not sure if we talk about the same. I'm basically talking about the 
dynamic binding, not dynamic endpoint creation.
I think I already found the issue. I will try to get a bit more into detail.

1. Linux: starts co-processor via remoteproc
2. co-processor: boots and reaches the point where it creates the 
endpoint like it is done in this ST example[1].
Be aware the src address is RPMSG_ADDR_ANY
3. co-processor: reaches the point where it sends the ns_announcement to 
linux ns endpoint
4. linux: receives the ns announcment, creates the channel, bindes the 
endpoint and checks here [2] if the source address is not RPMSG_ADDR_ANY 
and in this case it is not sending a ns_announcement (that's the issue 
when we use dynamic endpoints)
5. linux: according the openamp spec [3] it should now send the 
ns_announcement to the co-processor (slave)
6. co-processor: should receive the ns announcement and binds now the 
endpoint
7. co-processor: can now send messages to linux

This is basically what I'm expecting.


Do you think this is a bug or is the dynamic endpoint binding not 
handled? This line is there since ever [4] ...

Any other thoughts about this?

> This is a feature that is missing in the rpmsg virtio, and perhaps in rpmsg protocol
> itself (a ns annoucement ack message or something similar).
> 
> 
> A fix for the remoteproc virtio is available here:
> https://github.com/arnopo/meta-st-stm32mp-oss/commit/3e57fe73bd19c9bb835ac5a118e50727758b0b96
> 
> Don't hesitate to give me feedback on the fix, if you test it.

I added it to my branch and till now I don't see any side effects
> 
> I plan to propose the fix after this series.
> 
> Thanks,
> Arnaud
> 
>>
>> Cheers,
>> Philipp
>>

Cheers,
Philipp

[1]: 
https://github.com/STMicroelectronics/STM32CubeMP1/blob/master/Projects/STM32MP157C-DK2/Applications/OpenAMP/OpenAMP_raw/Src/openamp.c#L242

[2]: 
https://elixir.bootlin.com/linux/v5.17-rc5/source/drivers/rpmsg/virtio_rpmsg_bus.c#L425

[3]: 
https://github.com/OpenAMP/open-amp/blob/main/docs/img/coprocessor-rpmsg-ns.png

[4]: 
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=bcabbccabffe7326f046f25737ba1084f463c65c

Powered by blists - more mailing lists