oops, I did not reply-all... adding the ML back although Yishai replied/answered most of this anyway. On Tue, Nov 14, 2017 at 4:57 AM, Jason Gunthorpe <jgg@xxxxxxxx> wrote: > > On Mon, Nov 13, 2017 at 10:32:06PM +0200, Alex Rosenbaum wrote: > > > > I assume you'll also add a create_cq_ex to accept a PD instead of a > > > device? > > > currently we're mainly focusing on improving the inter-QP locks. > > we did not plan to improve the CQ related locking at this point so we > > did not plan to add parent domain to the create CQ in this series. > > once Mellanox (or someone else) will improve the CQ related locks it > > will definitely be the first step. > > I would like to see the evolve into a more complete solution, not just > for simple locking but also NUMA issues as well. > > Eg I would like to see it replace the rather limited comp_vector in the > CQ. > > Adding APIs like > > ibv_td_set_cpuset() > * Tell the library the application has bound the calling threads to > certain CPU(s). This may help the library & kernel make a NUMA > decision, eg by choosing the right MSI-X or memory allocation strategy. > > ibv_td_set_locking_mode() > * Disable locking inside the library assuming the caller guarentees > external locking or single-threaded operation. > > etc > > And given Doug's preference to keep the API style, lets use some > comp_mask approach and a ib_td_modify & create whatsit. > > 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
