1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (c) 2015-2020, Oracle and/or its affiliates. 4 * 5 * Support for reverse-direction RPCs on RPC/RDMA. 6 */ 7 8 #include <linux/sunrpc/xprt.h> 9 #include <linux/sunrpc/svc.h> 10 #include <linux/sunrpc/svc_xprt.h> 11 #include <linux/sunrpc/svc_rdma.h> 12 #include <linux/sunrpc/bc_xprt.h> 13 14 #include "xprt_rdma.h" 15 #include <trace/events/rpcrdma.h> 16 17 #undef RPCRDMA_BACKCHANNEL_DEBUG 18 19 /** 20 * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests 21 * @xprt: transport associated with these backchannel resources 22 * @reqs: number of concurrent incoming requests to expect 23 * 24 * Returns 0 on success; otherwise a negative errno 25 */ 26 int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs) 27 { 28 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 29 30 r_xprt->rx_buf.rb_bc_srv_max_requests = RPCRDMA_BACKWARD_WRS >> 1; 31 trace_xprtrdma_cb_setup(r_xprt, reqs); 32 return 0; 33 } 34 35 /** 36 * xprt_rdma_bc_maxpayload - Return maximum backchannel message size 37 * @xprt: transport 38 * 39 * Returns maximum size, in bytes, of a backchannel message 40 */ 41 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *xprt) 42 { 43 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 44 struct rpcrdma_ep *ep = r_xprt->rx_ep; 45 size_t maxmsg; 46 47 maxmsg = min_t(unsigned int, ep->re_inline_send, ep->re_inline_recv); 48 maxmsg = min_t(unsigned int, maxmsg, PAGE_SIZE); 49 return maxmsg - RPCRDMA_HDRLEN_MIN; 50 } 51 52 unsigned int xprt_rdma_bc_max_slots(struct rpc_xprt *xprt) 53 { 54 return RPCRDMA_BACKWARD_WRS >> 1; 55 } 56 57 static int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst) 58 { 59 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt); 60 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 61 __be32 *p; 62 63 rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0); 64 xdr_init_encode(&req->rl_stream, &req->rl_hdrbuf, 65 rdmab_data(req->rl_rdmabuf), rqst); 66 67 p = xdr_reserve_space(&req->rl_stream, 28); 68 if (unlikely(!p)) 69 return -EIO; 70 *p++ = rqst->rq_xid; 71 *p++ = rpcrdma_version; 72 *p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests); 73 *p++ = rdma_msg; 74 *p++ = xdr_zero; 75 *p++ = xdr_zero; 76 *p = xdr_zero; 77 78 if (rpcrdma_prepare_send_sges(r_xprt, req, RPCRDMA_HDRLEN_MIN, 79 &rqst->rq_snd_buf, rpcrdma_noch_pullup)) 80 return -EIO; 81 82 trace_xprtrdma_cb_reply(r_xprt, rqst); 83 return 0; 84 } 85 86 /** 87 * xprt_rdma_bc_send_reply - marshal and send a backchannel reply 88 * @rqst: RPC rqst with a backchannel RPC reply in rq_snd_buf 89 * 90 * Caller holds the transport's write lock. 91 * 92 * Returns: 93 * %0 if the RPC message has been sent 94 * %-ENOTCONN if the caller should reconnect and call again 95 * %-EIO if a permanent error occurred and the request was not 96 * sent. Do not try to send this message again. 97 */ 98 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst) 99 { 100 struct rpc_xprt *xprt = rqst->rq_xprt; 101 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 102 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 103 int rc; 104 105 if (!xprt_connected(xprt)) 106 return -ENOTCONN; 107 108 if (!xprt_request_get_cong(xprt, rqst)) 109 return -EBADSLT; 110 111 rc = rpcrdma_bc_marshal_reply(rqst); 112 if (rc < 0) 113 goto failed_marshal; 114 115 if (frwr_send(r_xprt, req)) 116 goto drop_connection; 117 return 0; 118 119 failed_marshal: 120 if (rc != -ENOTCONN) 121 return rc; 122 drop_connection: 123 xprt_rdma_close(xprt); 124 return -ENOTCONN; 125 } 126 127 /** 128 * xprt_rdma_bc_destroy - Release resources for handling backchannel requests 129 * @xprt: transport associated with these backchannel resources 130 * @reqs: number of incoming requests to destroy; ignored 131 */ 132 void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs) 133 { 134 struct rpc_rqst *rqst, *tmp; 135 136 spin_lock(&xprt->bc_pa_lock); 137 list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) { 138 list_del(&rqst->rq_bc_pa_list); 139 spin_unlock(&xprt->bc_pa_lock); 140 141 rpcrdma_req_destroy(rpcr_to_rdmar(rqst)); 142 143 spin_lock(&xprt->bc_pa_lock); 144 } 145 spin_unlock(&xprt->bc_pa_lock); 146 } 147 148 /** 149 * xprt_rdma_bc_free_rqst - Release a backchannel rqst 150 * @rqst: request to release 151 */ 152 void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst) 153 { 154 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 155 struct rpcrdma_rep *rep = req->rl_reply; 156 struct rpc_xprt *xprt = rqst->rq_xprt; 157 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 158 159 rpcrdma_rep_put(&r_xprt->rx_buf, rep); 160 req->rl_reply = NULL; 161 162 rpcrdma_req_put(req); 163 xprt_put(xprt); 164 } 165 166 static struct rpc_rqst *rpcrdma_bc_rqst_get(struct rpcrdma_xprt *r_xprt) 167 { 168 struct rpc_xprt *xprt = &r_xprt->rx_xprt; 169 struct rpcrdma_req *req; 170 struct rpc_rqst *rqst; 171 size_t size; 172 173 spin_lock(&xprt->bc_pa_lock); 174 rqst = list_first_entry_or_null(&xprt->bc_pa_list, struct rpc_rqst, 175 rq_bc_pa_list); 176 if (!rqst) 177 goto create_req; 178 list_del(&rqst->rq_bc_pa_list); 179 spin_unlock(&xprt->bc_pa_lock); 180 return rqst; 181 182 create_req: 183 spin_unlock(&xprt->bc_pa_lock); 184 185 /* Set a limit to prevent a remote from overrunning our resources. 186 */ 187 if (xprt->bc_alloc_count >= RPCRDMA_BACKWARD_WRS) 188 return NULL; 189 190 size = min_t(size_t, r_xprt->rx_ep->re_inline_recv, PAGE_SIZE); 191 req = rpcrdma_req_create(r_xprt, size); 192 if (!req) 193 return NULL; 194 if (rpcrdma_req_setup(r_xprt, req)) { 195 rpcrdma_req_destroy(req); 196 return NULL; 197 } 198 199 xprt->bc_alloc_count++; 200 rqst = &req->rl_slot; 201 rqst->rq_xprt = xprt; 202 __set_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state); 203 xdr_buf_init(&rqst->rq_snd_buf, rdmab_data(req->rl_sendbuf), size); 204 kref_init(&req->rl_kref); 205 return rqst; 206 } 207 208 /** 209 * rpcrdma_bc_receive_call - Handle a reverse-direction Call 210 * @r_xprt: transport receiving the call 211 * @rep: receive buffer containing the call 212 * 213 * Operational assumptions: 214 * o Backchannel credits are ignored, just as the NFS server 215 * forechannel currently does 216 * o The ULP manages a replay cache (eg, NFSv4.1 sessions). 217 * No replay detection is done at the transport level 218 */ 219 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt, 220 struct rpcrdma_rep *rep) 221 { 222 struct rpc_xprt *xprt = &r_xprt->rx_xprt; 223 struct rpcrdma_req *req; 224 struct rpc_rqst *rqst; 225 struct xdr_buf *buf; 226 size_t size; 227 __be32 *p; 228 229 p = xdr_inline_decode(&rep->rr_stream, 0); 230 size = xdr_stream_remaining(&rep->rr_stream); 231 232 #ifdef RPCRDMA_BACKCHANNEL_DEBUG 233 pr_info("RPC: %s: callback XID %08x, length=%u\n", 234 __func__, be32_to_cpup(p), size); 235 pr_info("RPC: %s: %*ph\n", __func__, size, p); 236 #endif 237 238 rqst = rpcrdma_bc_rqst_get(r_xprt); 239 if (!rqst) 240 goto out_overflow; 241 242 rqst->rq_reply_bytes_recvd = 0; 243 rqst->rq_xid = *p; 244 245 rqst->rq_private_buf.len = size; 246 247 buf = &rqst->rq_rcv_buf; 248 memset(buf, 0, sizeof(*buf)); 249 buf->head[0].iov_base = p; 250 buf->head[0].iov_len = size; 251 buf->len = size; 252 253 /* The receive buffer has to be hooked to the rpcrdma_req 254 * so that it is not released while the req is pointing 255 * to its buffer, and so that it can be reposted after 256 * the Upper Layer is done decoding it. 257 */ 258 req = rpcr_to_rdmar(rqst); 259 req->rl_reply = rep; 260 trace_xprtrdma_cb_call(r_xprt, rqst); 261 262 /* Queue rqst for ULP's callback service */ 263 xprt_enqueue_bc_request(rqst); 264 265 r_xprt->rx_stats.bcall_count++; 266 return; 267 268 out_overflow: 269 pr_warn("RPC/RDMA backchannel overflow\n"); 270 xprt_force_disconnect(xprt); 271 /* This receive buffer gets reposted automatically 272 * when the connection is re-established. 273 */ 274 return; 275 } 276