Syscall proxying on Zephyr

[rhaberkorn]

Hello!

I am struggling to use syscall (file IO) proxying in a Zephyr firmware running on a slave MCU (M4 core on STM32MP157C). The master will be running Linux.

Unfortunately, the rpc_demo example is not written for Zephyr. So I tried to adapt code from the rpmsg_multi_services.

I have to call rpmsg_rpc_init() and pass plaform_poll which is unfortunately not defined on Zephyr. I believe it should be similar to this:

static int platform_poll(void *priv)
{
	struct remoteproc *rproc = priv;
	int ret;

	while(1) {
		if (metal_io_read32(shm_io, 0)) {
			ret = remoteproc_get_notification(rproc,
							  RSC_NOTIFY_ID_ANY);
			if (ret)
				return ret;
			break;
		}
	}
	return 0;
}

Do I understand correctly, that I will have to implement something like platform_create_proc() as well in order to initialize struct remoteproc?
This leads of course into a rabbit hole, since I will now have to provide remoteproc_ops as well. I feel, I might be misunderstanding something fundamentally.

[arnopo]

Hello @rhaberkorn
on zephyr we use mailbox ( or legacy ipm) for notification.
to notify the remote side we use mailbox_notify()
To receive a message you have to register a callback in rpmsg_create_ep()

[rhaberkorn]

Okay, you say that ipm_send() must be used instead of remoteproc_get_notification()? Actually I wasn't even sure what's the purpose of remoteproc_get_notification(). It's poorly documented.
What exactly the platform_poll() callback is supposed to do - how does it fit into the larger picture?

[tnmysh]

The platform_poll() and remoteproc_get_notifiaction() are used baremetal (mostly for zynqmp platform) and linux apps (linux userspace apps that is using OpenAMP library).

Whole apps directory needs to be refactored.

[rhaberkorn]

That doesn't really explain what they do and what I have to change for a port to Zephyr.

[tnmysh]

platform_poll() is platform specific implementation to wait till kick is received from host.
Once kick is received, remoteproc_get_notifiaction() function starts scanning vring for new buffers.

I don't know what you mean by "port to zephyr". Zephyr doesn't need to implement platform_poll(). This functionality is achieved by using mailbox or ipm driver. Please refer zephyr demo.

Zephyr demo receives messages here: https://github.com/OpenAMP/openamp-system-reference/blob/main/examples/zephyr/rpmsg_multi_services/src/main_remote.c#L471

[rhaberkorn]

@tnmysh If I don't implement platform_poll() on Zephyr, won't these loops in rpmsg_retarget.c simply block indefinitely?

while ((atomic_flag_test_and_set(&rpc->nacked))) {
        if (rpc->poll)
                rpc->poll(rpc->poll_arg);
}

The nacked flag is obviously cleared only in rpmsg_rpc_ept_cb(). So unless that get's called from interrupts or in separate threads, there is no chance this flag could be cleared. And indeed, my code without any poll callback appears to hang.

The rpmsg_multi_services example for Zephyr appears to use semaphores for that purpose (cf. app_rpmsg_raw()), so cooperative threading appears to be involved after all.
I would suppose, that it could be worth a try to implement platform_poll() with a k_yield() or even k_usleep().

[rhaberkorn]

btw. it even seems, I cannot currently enable the Proxy library (WITH_PROXY) from Zephyr. Hell, I don't even see how Zephyr eventually passes down CMake flags to OpenAMP. Apparently I will have to discuss this with the Zephyr people. Apparently they have a Discord channel for OpenAMP...

[rhaberkorn]

Theoretically, it should be possible to do a west build -- -DWITH_PROXY=ON, but it doesn't work. The settings gets reset/overridden in Zephyr's CMake cache.

[rhaberkorn]

This one I solved. It's a problem in Zephyr's binding which silently and mercilessly forced WITH_PROXY=OFF. Should be fixed upstream (in Zephyr).