On Fri, Apr 15, 2016 at 05:30:03PM +0000, Hefty, Sean wrote: > > I'm not excited about changing the rdma_cm and uverbs char names, we > > have a big investment in training and userspace around that. I'd > > rather have something that does more than what its name implies, > > so imagine uverbs as hfi if that helps... > > Changing from write to ioctl breaks everything. IMO, a name change > hardly matters, and makes it trivial to see which interface should > be used. No, the dev name is sysadmin visible, the ioctl vs write choice is invisible. We have a big investment in training sysadmins on the security model around /dev/uverbs particularly, deployment of udev rules, distro setups, etc. Needlessly changing that name means work for sysadmins. We should not change it for trivial reasons. > > > int uverbs_fd = ioctl(hsi_fd, RDMA_GET_DRIVER_OPS_FD, "verbs"); > > > int psm2_fd = ioctl(hsi_fd, RDMA_GET_DRIVER_OPS_FD, "psm2.intel.com"); > > > > Again the *DRIVER* word is key, it really is only for driver specific > > stuff. > > One _could_ argue that this should be > "verbs.[device|vendor|technology]", since we ultimately end up > dropping into device specific calls, and this just changes where the > multiplexing is done. That feels like it would be horrible for the user space side? Every ioctl site needs a switch statement for all the possibilities? How would the existing code sharing work on the kernel side? Hurm, I don't know - the only other 'APIs' I can thing of using a techinque like this are Web focused ones, (eg HTTP UPGRADE and others) and I have very little idea on the good/bads over time in that space. One plus side I guess is it is an obvious place to put request_module, which is something our subsystem has historically been lacking. The driver bypass fd was a technique I came up with specifically to deal with the qib/hfi issue.. IMHO, it is ugly, but the problem is also very ugly, so maybe it is OK. It is less ugly than every driver having a dedicated cdev.. > > I see no big problem with somewhat orthogonal interfaces as groups of > > ioctls on the same fd. libfabric shows that is basically a workable > > model. Sharing security/discovery/etc is a reasonable enough commonality. > > If we consider a completely non-verbs device, what unrelated code > and modules must be loaded for that device to export its interfaces > to user space? Force loading the infiniband stack for a non-IB > device seems like the wrong approach. I'm not even slightly concerned about that. This is HPC, if you care about a 100k kernel module you are doing something wrong. If someone really cares then design a CONFIG option to compile out the stuff they don't want. And again.. The non-verbs community hasn't really standardized on *anything* hardware facing. It is just impossible, as a kernel community to have any rational discussion on what to provide as a common API when we have no idea what even the requirements are. The two banner non-verbs hardwares (usNIC and qib/hfi) seem *totally* different and, perhaps tellingly, currently, don't need the kernel do anything beyond administrate secure access to some proprietary hardware. Jason -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html