On 12/24/18 2:56 AM, Trond Myklebust wrote: > On Sat, 2018-12-22 at 20:46 +0300, Vasily Averin wrote: >> On 12/21/18 4:00 AM, bfields@xxxxxxxxxxxx wrote: >>> On Tue, Dec 18, 2018 at 02:55:15PM +0000, Trond Myklebust wrote: >>>> No. We don't care about xpt_flags for the back channel because >>>> there is >>>> no "server transport". The actual transport is stored in the >>>> 'struct >>>> rpc_rqst', and is the struct rpc_xprt corresponding to the client >>>> socket or RDMA channel. >>>> >>>> IOW: All we really need in svc_process_common() is to be able to >>>> run >>>> rqstp->rq_xprt->xpt_ops->xpo_prep_reply_hdr(), and that can be >>>> passed >>>> either as a pointer to the struct svc_xprt_ops itself. >>> >>> For what it's worth, I'd rather get rid of that op--it's an awfully >>> roundabout way just to do "svc_putnl(resv, 0);" in the tcp case. >> >> Do you mean that svc_create_xprt(serv, "tcp-bc", ...) was used ONLY >> to call >> svc_tcp_prep_reply_hdr() in svc_process_common() ? >> And according call for rdma-bc does nothing useful at all? >> >> I've just tried to remove svc_create_xprt() from xs_tcp_bc_up() and >> just >> provide pointer to svc_tcp_prep_reply_hdr() in svc_process_common() >> via per-netns sunrpc_net -- and seems it was enough, my testcase >> worked correctly. > > I don't see how that function is related to net namespaces. As far as I > can tell, it only signals whether or not the type of transport uses the > TCP record marking scheme. We need to know which kind of transport is used in specified net namespace, for example init_ns can use RDMA transport and netns "second" can use TCP transport at the same time. If you do not like an idea to use function pointer as a mark -- ok I can save only some boolean flag on sunrpc_net, check it in svc_process_common() and if it is set -- call svc_tcp_prep_reply_hdr() directly. Is it acceptable for you? > IOW: it depends on whether the client is using a stream based protocol > like TCP, or a datagram-like protocol like UDP, or RDMA. Whether that > use is occurring in a private net namespace or in the init process > namespace would be irrelevant. > >> Am I missed something probably? >> Should we really remove svc_create_xprt( "tcp/rdma-bc"...) related >> stuff? ? > > Agreed. The 'bc_up' callback in struct rpc_xprt_ops serves no > discernible purpose, and can be removed. >
