1b2441318SGreg Kroah-Hartman // SPDX-License-Identifier: GPL-2.0 2f531a5dbSChuck Lever /* 3f531a5dbSChuck Lever * Copyright (c) 2015 Oracle. All rights reserved. 4f531a5dbSChuck Lever * 5f531a5dbSChuck Lever * Support for backward direction RPCs on RPC/RDMA. 6f531a5dbSChuck Lever */ 7f531a5dbSChuck Lever 8f531a5dbSChuck Lever #include <linux/module.h> 963cae470SChuck Lever #include <linux/sunrpc/xprt.h> 1063cae470SChuck Lever #include <linux/sunrpc/svc.h> 1176566773SChuck Lever #include <linux/sunrpc/svc_xprt.h> 12f531a5dbSChuck Lever 13f531a5dbSChuck Lever #include "xprt_rdma.h" 14f531a5dbSChuck Lever 15f531a5dbSChuck Lever #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) 16f531a5dbSChuck Lever # define RPCDBG_FACILITY RPCDBG_TRANS 17f531a5dbSChuck Lever #endif 18f531a5dbSChuck Lever 19c8bbe0c7SChuck Lever #undef RPCRDMA_BACKCHANNEL_DEBUG 2063cae470SChuck Lever 21f531a5dbSChuck Lever static void rpcrdma_bc_free_rqst(struct rpcrdma_xprt *r_xprt, 22f531a5dbSChuck Lever struct rpc_rqst *rqst) 23f531a5dbSChuck Lever { 24f531a5dbSChuck Lever struct rpcrdma_buffer *buf = &r_xprt->rx_buf; 25f531a5dbSChuck Lever struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 26f531a5dbSChuck Lever 27f531a5dbSChuck Lever spin_lock(&buf->rb_reqslock); 28f531a5dbSChuck Lever list_del(&req->rl_all); 29f531a5dbSChuck Lever spin_unlock(&buf->rb_reqslock); 30f531a5dbSChuck Lever 3113650c23SChuck Lever rpcrdma_destroy_req(req); 32f531a5dbSChuck Lever } 33f531a5dbSChuck Lever 34edb41e61SChuck Lever static int rpcrdma_bc_setup_reqs(struct rpcrdma_xprt *r_xprt, 35edb41e61SChuck Lever unsigned int count) 36f531a5dbSChuck Lever { 37edb41e61SChuck Lever struct rpc_xprt *xprt = &r_xprt->rx_xprt; 38edb41e61SChuck Lever struct rpc_rqst *rqst; 39edb41e61SChuck Lever unsigned int i; 40edb41e61SChuck Lever 41edb41e61SChuck Lever for (i = 0; i < (count << 1); i++) { 42f531a5dbSChuck Lever struct rpcrdma_regbuf *rb; 43f531a5dbSChuck Lever struct rpcrdma_req *req; 44f531a5dbSChuck Lever size_t size; 45f531a5dbSChuck Lever 46f531a5dbSChuck Lever req = rpcrdma_create_req(r_xprt); 47abfb6897SDan Carpenter if (IS_ERR(req)) 48abfb6897SDan Carpenter return PTR_ERR(req); 49edb41e61SChuck Lever rqst = &req->rl_slot; 50edb41e61SChuck Lever 51edb41e61SChuck Lever rqst->rq_xprt = xprt; 52edb41e61SChuck Lever INIT_LIST_HEAD(&rqst->rq_list); 53edb41e61SChuck Lever INIT_LIST_HEAD(&rqst->rq_bc_list); 54edb41e61SChuck Lever __set_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state); 55edb41e61SChuck Lever spin_lock_bh(&xprt->bc_pa_lock); 56edb41e61SChuck Lever list_add(&rqst->rq_bc_pa_list, &xprt->bc_pa_list); 57edb41e61SChuck Lever spin_unlock_bh(&xprt->bc_pa_lock); 58f531a5dbSChuck Lever 5908cf2efdSChuck Lever size = r_xprt->rx_data.inline_rsize; 6013650c23SChuck Lever rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, GFP_KERNEL); 61f531a5dbSChuck Lever if (IS_ERR(rb)) 62f531a5dbSChuck Lever goto out_fail; 63f531a5dbSChuck Lever req->rl_sendbuf = rb; 6462aee0e3SChuck Lever xdr_buf_init(&rqst->rq_snd_buf, rb->rg_base, 6562aee0e3SChuck Lever min_t(size_t, size, PAGE_SIZE)); 66edb41e61SChuck Lever } 67f531a5dbSChuck Lever return 0; 68f531a5dbSChuck Lever 69f531a5dbSChuck Lever out_fail: 70f531a5dbSChuck Lever rpcrdma_bc_free_rqst(r_xprt, rqst); 71f531a5dbSChuck Lever return -ENOMEM; 72f531a5dbSChuck Lever } 73f531a5dbSChuck Lever 74f531a5dbSChuck Lever /** 75f531a5dbSChuck Lever * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests 76f531a5dbSChuck Lever * @xprt: transport associated with these backchannel resources 77f531a5dbSChuck Lever * @reqs: number of concurrent incoming requests to expect 78f531a5dbSChuck Lever * 79f531a5dbSChuck Lever * Returns 0 on success; otherwise a negative errno 80f531a5dbSChuck Lever */ 81f531a5dbSChuck Lever int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs) 82f531a5dbSChuck Lever { 83f531a5dbSChuck Lever struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 84f531a5dbSChuck Lever int rc; 85f531a5dbSChuck Lever 86f531a5dbSChuck Lever /* The backchannel reply path returns each rpc_rqst to the 87f531a5dbSChuck Lever * bc_pa_list _after_ the reply is sent. If the server is 88f531a5dbSChuck Lever * faster than the client, it can send another backward 89f531a5dbSChuck Lever * direction request before the rpc_rqst is returned to the 90f531a5dbSChuck Lever * list. The client rejects the request in this case. 91f531a5dbSChuck Lever * 92f531a5dbSChuck Lever * Twice as many rpc_rqsts are prepared to ensure there is 93f531a5dbSChuck Lever * always an rpc_rqst available as soon as a reply is sent. 94f531a5dbSChuck Lever */ 95124fa17dSChuck Lever if (reqs > RPCRDMA_BACKWARD_WRS >> 1) 96124fa17dSChuck Lever goto out_err; 97124fa17dSChuck Lever 98edb41e61SChuck Lever rc = rpcrdma_bc_setup_reqs(r_xprt, reqs); 99edb41e61SChuck Lever if (rc) 100f531a5dbSChuck Lever goto out_free; 101d2c23c00SMarkus Elfring 102edb41e61SChuck Lever r_xprt->rx_buf.rb_bc_srv_max_requests = reqs; 103f531a5dbSChuck Lever request_module("svcrdma"); 104fc1eb807SChuck Lever trace_xprtrdma_cb_setup(r_xprt, reqs); 105f531a5dbSChuck Lever return 0; 106f531a5dbSChuck Lever 107f531a5dbSChuck Lever out_free: 108f531a5dbSChuck Lever xprt_rdma_bc_destroy(xprt, reqs); 109f531a5dbSChuck Lever 110124fa17dSChuck Lever out_err: 111f531a5dbSChuck Lever pr_err("RPC: %s: setup backchannel transport failed\n", __func__); 112f531a5dbSChuck Lever return -ENOMEM; 113f531a5dbSChuck Lever } 114f531a5dbSChuck Lever 115f531a5dbSChuck Lever /** 11676566773SChuck Lever * xprt_rdma_bc_up - Create transport endpoint for backchannel service 11776566773SChuck Lever * @serv: server endpoint 11876566773SChuck Lever * @net: network namespace 11976566773SChuck Lever * 12076566773SChuck Lever * The "xprt" is an implied argument: it supplies the name of the 12176566773SChuck Lever * backchannel transport class. 12276566773SChuck Lever * 12376566773SChuck Lever * Returns zero on success, negative errno on failure 12476566773SChuck Lever */ 12576566773SChuck Lever int xprt_rdma_bc_up(struct svc_serv *serv, struct net *net) 12676566773SChuck Lever { 12776566773SChuck Lever int ret; 12876566773SChuck Lever 12976566773SChuck Lever ret = svc_create_xprt(serv, "rdma-bc", net, PF_INET, 0, 0); 13076566773SChuck Lever if (ret < 0) 13176566773SChuck Lever return ret; 13276566773SChuck Lever return 0; 13376566773SChuck Lever } 13476566773SChuck Lever 13576566773SChuck Lever /** 1366b26cc8cSChuck Lever * xprt_rdma_bc_maxpayload - Return maximum backchannel message size 1376b26cc8cSChuck Lever * @xprt: transport 1386b26cc8cSChuck Lever * 1396b26cc8cSChuck Lever * Returns maximum size, in bytes, of a backchannel message 1406b26cc8cSChuck Lever */ 1416b26cc8cSChuck Lever size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *xprt) 1426b26cc8cSChuck Lever { 1436b26cc8cSChuck Lever struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 1446b26cc8cSChuck Lever struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data; 1456b26cc8cSChuck Lever size_t maxmsg; 1466b26cc8cSChuck Lever 1476b26cc8cSChuck Lever maxmsg = min_t(unsigned int, cdata->inline_rsize, cdata->inline_wsize); 14862aee0e3SChuck Lever maxmsg = min_t(unsigned int, maxmsg, PAGE_SIZE); 1496b26cc8cSChuck Lever return maxmsg - RPCRDMA_HDRLEN_MIN; 1506b26cc8cSChuck Lever } 1516b26cc8cSChuck Lever 152cf73daf5SChuck Lever static int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst) 15383128a60SChuck Lever { 1547ec910e7SChuck Lever struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt); 15583128a60SChuck Lever struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 1567ec910e7SChuck Lever __be32 *p; 15783128a60SChuck Lever 1587ec910e7SChuck Lever rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0); 1597ec910e7SChuck Lever xdr_init_encode(&req->rl_stream, &req->rl_hdrbuf, 1607ec910e7SChuck Lever req->rl_rdmabuf->rg_base); 1617ec910e7SChuck Lever 1627ec910e7SChuck Lever p = xdr_reserve_space(&req->rl_stream, 28); 1637ec910e7SChuck Lever if (unlikely(!p)) 1647ec910e7SChuck Lever return -EIO; 1657ec910e7SChuck Lever *p++ = rqst->rq_xid; 1667ec910e7SChuck Lever *p++ = rpcrdma_version; 1677ec910e7SChuck Lever *p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests); 1687ec910e7SChuck Lever *p++ = rdma_msg; 1697ec910e7SChuck Lever *p++ = xdr_zero; 1707ec910e7SChuck Lever *p++ = xdr_zero; 1717ec910e7SChuck Lever *p = xdr_zero; 17283128a60SChuck Lever 173857f9acaSChuck Lever if (rpcrdma_prepare_send_sges(r_xprt, req, RPCRDMA_HDRLEN_MIN, 174655fec69SChuck Lever &rqst->rq_snd_buf, rpcrdma_noch)) 17554cbd6b0SChuck Lever return -EIO; 176fc1eb807SChuck Lever 177fc1eb807SChuck Lever trace_xprtrdma_cb_reply(rqst); 178655fec69SChuck Lever return 0; 17983128a60SChuck Lever } 18083128a60SChuck Lever 18183128a60SChuck Lever /** 182cf73daf5SChuck Lever * xprt_rdma_bc_send_reply - marshal and send a backchannel reply 183cf73daf5SChuck Lever * @rqst: RPC rqst with a backchannel RPC reply in rq_snd_buf 184cf73daf5SChuck Lever * 185cf73daf5SChuck Lever * Caller holds the transport's write lock. 186cf73daf5SChuck Lever * 187cf73daf5SChuck Lever * Returns: 188cf73daf5SChuck Lever * %0 if the RPC message has been sent 189cf73daf5SChuck Lever * %-ENOTCONN if the caller should reconnect and call again 190cf73daf5SChuck Lever * %-EIO if a permanent error occurred and the request was not 191cf73daf5SChuck Lever * sent. Do not try to send this message again. 192cf73daf5SChuck Lever */ 193cf73daf5SChuck Lever int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst) 194cf73daf5SChuck Lever { 195cf73daf5SChuck Lever struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt); 196cf73daf5SChuck Lever struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 197cf73daf5SChuck Lever int rc; 198cf73daf5SChuck Lever 199cf73daf5SChuck Lever if (!xprt_connected(rqst->rq_xprt)) 200cf73daf5SChuck Lever goto drop_connection; 201cf73daf5SChuck Lever 202cf73daf5SChuck Lever rc = rpcrdma_bc_marshal_reply(rqst); 203cf73daf5SChuck Lever if (rc < 0) 204cf73daf5SChuck Lever goto failed_marshal; 205cf73daf5SChuck Lever 206*7c8d9e7cSChuck Lever rpcrdma_post_recvs(r_xprt, true); 207cf73daf5SChuck Lever if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req)) 208cf73daf5SChuck Lever goto drop_connection; 209cf73daf5SChuck Lever return 0; 210cf73daf5SChuck Lever 211cf73daf5SChuck Lever failed_marshal: 212cf73daf5SChuck Lever if (rc != -ENOTCONN) 213cf73daf5SChuck Lever return rc; 214cf73daf5SChuck Lever drop_connection: 215cf73daf5SChuck Lever xprt_disconnect_done(rqst->rq_xprt); 216cf73daf5SChuck Lever return -ENOTCONN; 217cf73daf5SChuck Lever } 218cf73daf5SChuck Lever 219cf73daf5SChuck Lever /** 220f531a5dbSChuck Lever * xprt_rdma_bc_destroy - Release resources for handling backchannel requests 221f531a5dbSChuck Lever * @xprt: transport associated with these backchannel resources 222f531a5dbSChuck Lever * @reqs: number of incoming requests to destroy; ignored 223f531a5dbSChuck Lever */ 224f531a5dbSChuck Lever void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs) 225f531a5dbSChuck Lever { 226f531a5dbSChuck Lever struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 227f531a5dbSChuck Lever struct rpc_rqst *rqst, *tmp; 228f531a5dbSChuck Lever 229f531a5dbSChuck Lever spin_lock_bh(&xprt->bc_pa_lock); 230f531a5dbSChuck Lever list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) { 231f531a5dbSChuck Lever list_del(&rqst->rq_bc_pa_list); 232f531a5dbSChuck Lever spin_unlock_bh(&xprt->bc_pa_lock); 233f531a5dbSChuck Lever 234f531a5dbSChuck Lever rpcrdma_bc_free_rqst(r_xprt, rqst); 235f531a5dbSChuck Lever 236f531a5dbSChuck Lever spin_lock_bh(&xprt->bc_pa_lock); 237f531a5dbSChuck Lever } 238f531a5dbSChuck Lever spin_unlock_bh(&xprt->bc_pa_lock); 239f531a5dbSChuck Lever } 240f531a5dbSChuck Lever 241f531a5dbSChuck Lever /** 242f531a5dbSChuck Lever * xprt_rdma_bc_free_rqst - Release a backchannel rqst 243f531a5dbSChuck Lever * @rqst: request to release 244f531a5dbSChuck Lever */ 245f531a5dbSChuck Lever void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst) 246f531a5dbSChuck Lever { 247*7c8d9e7cSChuck Lever struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 248f531a5dbSChuck Lever struct rpc_xprt *xprt = rqst->rq_xprt; 249f531a5dbSChuck Lever 250c8bbe0c7SChuck Lever dprintk("RPC: %s: freeing rqst %p (req %p)\n", 251*7c8d9e7cSChuck Lever __func__, rqst, req); 252*7c8d9e7cSChuck Lever 253*7c8d9e7cSChuck Lever rpcrdma_recv_buffer_put(req->rl_reply); 254*7c8d9e7cSChuck Lever req->rl_reply = NULL; 255c8bbe0c7SChuck Lever 256f531a5dbSChuck Lever spin_lock_bh(&xprt->bc_pa_lock); 257f531a5dbSChuck Lever list_add_tail(&rqst->rq_bc_pa_list, &xprt->bc_pa_list); 258f531a5dbSChuck Lever spin_unlock_bh(&xprt->bc_pa_lock); 259f531a5dbSChuck Lever } 26063cae470SChuck Lever 26163cae470SChuck Lever /** 26263cae470SChuck Lever * rpcrdma_bc_receive_call - Handle a backward direction call 2639ab6d89eSChuck Lever * @r_xprt: transport receiving the call 26463cae470SChuck Lever * @rep: receive buffer containing the call 26563cae470SChuck Lever * 26663cae470SChuck Lever * Operational assumptions: 26763cae470SChuck Lever * o Backchannel credits are ignored, just as the NFS server 26863cae470SChuck Lever * forechannel currently does 26963cae470SChuck Lever * o The ULP manages a replay cache (eg, NFSv4.1 sessions). 27063cae470SChuck Lever * No replay detection is done at the transport level 27163cae470SChuck Lever */ 27263cae470SChuck Lever void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt, 27363cae470SChuck Lever struct rpcrdma_rep *rep) 27463cae470SChuck Lever { 27563cae470SChuck Lever struct rpc_xprt *xprt = &r_xprt->rx_xprt; 27663cae470SChuck Lever struct svc_serv *bc_serv; 27763cae470SChuck Lever struct rpcrdma_req *req; 27863cae470SChuck Lever struct rpc_rqst *rqst; 27963cae470SChuck Lever struct xdr_buf *buf; 28063cae470SChuck Lever size_t size; 28163cae470SChuck Lever __be32 *p; 28263cae470SChuck Lever 28341c8f70fSChuck Lever p = xdr_inline_decode(&rep->rr_stream, 0); 28441c8f70fSChuck Lever size = xdr_stream_remaining(&rep->rr_stream); 28541c8f70fSChuck Lever 28663cae470SChuck Lever #ifdef RPCRDMA_BACKCHANNEL_DEBUG 28763cae470SChuck Lever pr_info("RPC: %s: callback XID %08x, length=%u\n", 28841c8f70fSChuck Lever __func__, be32_to_cpup(p), size); 28941c8f70fSChuck Lever pr_info("RPC: %s: %*ph\n", __func__, size, p); 29063cae470SChuck Lever #endif 29163cae470SChuck Lever 29263cae470SChuck Lever /* Grab a free bc rqst */ 29363cae470SChuck Lever spin_lock(&xprt->bc_pa_lock); 29463cae470SChuck Lever if (list_empty(&xprt->bc_pa_list)) { 29563cae470SChuck Lever spin_unlock(&xprt->bc_pa_lock); 29663cae470SChuck Lever goto out_overflow; 29763cae470SChuck Lever } 29863cae470SChuck Lever rqst = list_first_entry(&xprt->bc_pa_list, 29963cae470SChuck Lever struct rpc_rqst, rq_bc_pa_list); 30063cae470SChuck Lever list_del(&rqst->rq_bc_pa_list); 30163cae470SChuck Lever spin_unlock(&xprt->bc_pa_lock); 30263cae470SChuck Lever 30363cae470SChuck Lever /* Prepare rqst */ 30463cae470SChuck Lever rqst->rq_reply_bytes_recvd = 0; 30563cae470SChuck Lever rqst->rq_bytes_sent = 0; 30641c8f70fSChuck Lever rqst->rq_xid = *p; 3079f74660bSChuck Lever 3089f74660bSChuck Lever rqst->rq_private_buf.len = size; 30963cae470SChuck Lever 31063cae470SChuck Lever buf = &rqst->rq_rcv_buf; 31163cae470SChuck Lever memset(buf, 0, sizeof(*buf)); 31263cae470SChuck Lever buf->head[0].iov_base = p; 31363cae470SChuck Lever buf->head[0].iov_len = size; 31463cae470SChuck Lever buf->len = size; 31563cae470SChuck Lever 31663cae470SChuck Lever /* The receive buffer has to be hooked to the rpcrdma_req 31741c8f70fSChuck Lever * so that it is not released while the req is pointing 31841c8f70fSChuck Lever * to its buffer, and so that it can be reposted after 31941c8f70fSChuck Lever * the Upper Layer is done decoding it. 32063cae470SChuck Lever */ 32163cae470SChuck Lever req = rpcr_to_rdmar(rqst); 32263cae470SChuck Lever req->rl_reply = rep; 323fc1eb807SChuck Lever trace_xprtrdma_cb_call(rqst); 32463cae470SChuck Lever 32563cae470SChuck Lever /* Queue rqst for ULP's callback service */ 32663cae470SChuck Lever bc_serv = xprt->bc_serv; 32763cae470SChuck Lever spin_lock(&bc_serv->sv_cb_lock); 32863cae470SChuck Lever list_add(&rqst->rq_bc_list, &bc_serv->sv_cb_list); 32963cae470SChuck Lever spin_unlock(&bc_serv->sv_cb_lock); 33063cae470SChuck Lever 33163cae470SChuck Lever wake_up(&bc_serv->sv_cb_waitq); 33263cae470SChuck Lever 33363cae470SChuck Lever r_xprt->rx_stats.bcall_count++; 33463cae470SChuck Lever return; 33563cae470SChuck Lever 33663cae470SChuck Lever out_overflow: 33763cae470SChuck Lever pr_warn("RPC/RDMA backchannel overflow\n"); 33863cae470SChuck Lever xprt_disconnect_done(xprt); 33963cae470SChuck Lever /* This receive buffer gets reposted automatically 34063cae470SChuck Lever * when the connection is re-established. 34163cae470SChuck Lever */ 34263cae470SChuck Lever return; 34363cae470SChuck Lever } 344