This is a bad idea. Messages sent from ISR silently vanish if the receiver is not waiting (in msg_receive()), or if there's a message queue and that is full. Both will happen once there's load, so it's difficult to track down.

The effect would be easy to show with a unit test that calls bhp_msg_isr_cb() twice from an actual ISR. Without a queue, the second message will be lost. Or with a queue of size N, call bhp_msg_isr_cb` N+1 times.

I'm beginning to think we should deprecate msg_send() from ISRs and only allow msg_try_send(), which msg_send() will use anyways when called from ISR, eating the result code.

As discussed in many old PRs, I was never a fan of ISR offloading using IPC and I 100% agree with what you describe. I plan to migrate all the remaining IPC based IRQ Offloading modules to Events or Thread Flags at some point.
I only introduce this module to give support for the existing modules (concretely LWIP), since they already use IPC.
IMO it shouldn't make the situation worse but at least it should speed up the Radio HAL migration.

OK I see, this is needed because LWIP uses msg already for netdev events and this will integrate with it.

done!

Please squash!

squashed!