On Wed, Aug 07, 2024 at 10:44:21AM +0300, Oded Gabbay wrote: > Disconnecting the RPCs from the drivers and providing a generic interface will > allow to have a single user-space library which will be able to communicate > with the firmware for all the IPs in the device. In Habana's case, it was > mainly for monitoring and debugging purposes. Yeah, monitoring and debugging is definately a key use case. > I do have one question about the rpc ioctl. Are you assuming that the rpc > is synchronous (we send a message to the firmware and block until we get the > reply)? If so, what happen if we have an async RPC implementation > inside the driver? How would you recommend to handle it? Yes, this is all simplified so the ioctl system call is synchronous. mlx5 is async under the hood too, it just launches an async work and waits on a completion before returning from the system call. Userspace could multi-thread it, and If people were really keen maybe there is some io_uring approach. Thanks, Jason
