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 13 #include "xprt_rdma.h" 14 #include <trace/events/rpcrdma.h> 15 16 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) 17 # define RPCDBG_FACILITY RPCDBG_TRANS 18 #endif 19 20 #undef RPCRDMA_BACKCHANNEL_DEBUG 21 22 /** 23 * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests 24 * @xprt: transport associated with these backchannel resources 25 * @reqs: number of concurrent incoming requests to expect 26 * 27 * Returns 0 on success; otherwise a negative errno 28 */ 29 int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs) 30 { 31 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 32 33 r_xprt->rx_buf.rb_bc_srv_max_requests = RPCRDMA_BACKWARD_WRS >> 1; 34 trace_xprtrdma_cb_setup(r_xprt, reqs); 35 return 0; 36 } 37 38 /** 39 * xprt_rdma_bc_maxpayload - Return maximum backchannel message size 40 * @xprt: transport 41 * 42 * Returns maximum size, in bytes, of a backchannel message 43 */ 44 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *xprt) 45 { 46 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 47 struct rpcrdma_ep *ep = r_xprt->rx_ep; 48 size_t maxmsg; 49 50 maxmsg = min_t(unsigned int, ep->re_inline_send, ep->re_inline_recv); 51 maxmsg = min_t(unsigned int, maxmsg, PAGE_SIZE); 52 return maxmsg - RPCRDMA_HDRLEN_MIN; 53 } 54 55 unsigned int xprt_rdma_bc_max_slots(struct rpc_xprt *xprt) 56 { 57 return RPCRDMA_BACKWARD_WRS >> 1; 58 } 59 60 static int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst) 61 { 62 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt); 63 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 64 __be32 *p; 65 66 rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0); 67 xdr_init_encode(&req->rl_stream, &req->rl_hdrbuf, 68 rdmab_data(req->rl_rdmabuf), rqst); 69 70 p = xdr_reserve_space(&req->rl_stream, 28); 71 if (unlikely(!p)) 72 return -EIO; 73 *p++ = rqst->rq_xid; 74 *p++ = rpcrdma_version; 75 *p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests); 76 *p++ = rdma_msg; 77 *p++ = xdr_zero; 78 *p++ = xdr_zero; 79 *p = xdr_zero; 80 81 if (rpcrdma_prepare_send_sges(r_xprt, req, RPCRDMA_HDRLEN_MIN, 82 &rqst->rq_snd_buf, rpcrdma_noch_pullup)) 83 return -EIO; 84 85 trace_xprtrdma_cb_reply(r_xprt, rqst); 86 return 0; 87 } 88 89 /** 90 * xprt_rdma_bc_send_reply - marshal and send a backchannel reply 91 * @rqst: RPC rqst with a backchannel RPC reply in rq_snd_buf 92 * 93 * Caller holds the transport's write lock. 94 * 95 * Returns: 96 * %0 if the RPC message has been sent 97 * %-ENOTCONN if the caller should reconnect and call again 98 * %-EIO if a permanent error occurred and the request was not 99 * sent. Do not try to send this message again. 100 */ 101 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst) 102 { 103 struct rpc_xprt *xprt = rqst->rq_xprt; 104 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 105 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 106 int rc; 107 108 if (!xprt_connected(xprt)) 109 return -ENOTCONN; 110 111 if (!xprt_request_get_cong(xprt, rqst)) 112 return -EBADSLT; 113 114 rc = rpcrdma_bc_marshal_reply(rqst); 115 if (rc < 0) 116 goto failed_marshal; 117 118 if (frwr_send(r_xprt, req)) 119 goto drop_connection; 120 return 0; 121 122 failed_marshal: 123 if (rc != -ENOTCONN) 124 return rc; 125 drop_connection: 126 xprt_rdma_close(xprt); 127 return -ENOTCONN; 128 } 129 130 /** 131 * xprt_rdma_bc_destroy - Release resources for handling backchannel requests 132 * @xprt: transport associated with these backchannel resources 133 * @reqs: number of incoming requests to destroy; ignored 134 */ 135 void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs) 136 { 137 struct rpc_rqst *rqst, *tmp; 138 139 spin_lock(&xprt->bc_pa_lock); 140 list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) { 141 list_del(&rqst->rq_bc_pa_list); 142 spin_unlock(&xprt->bc_pa_lock); 143 144 rpcrdma_req_destroy(rpcr_to_rdmar(rqst)); 145 146 spin_lock(&xprt->bc_pa_lock); 147 } 148 spin_unlock(&xprt->bc_pa_lock); 149 } 150 151 /** 152 * xprt_rdma_bc_free_rqst - Release a backchannel rqst 153 * @rqst: request to release 154 */ 155 void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst) 156 { 157 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 158 struct rpcrdma_rep *rep = req->rl_reply; 159 struct rpc_xprt *xprt = rqst->rq_xprt; 160 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 161 162 rpcrdma_rep_put(&r_xprt->rx_buf, rep); 163 req->rl_reply = NULL; 164 165 spin_lock(&xprt->bc_pa_lock); 166 list_add_tail(&rqst->rq_bc_pa_list, &xprt->bc_pa_list); 167 spin_unlock(&xprt->bc_pa_lock); 168 xprt_put(xprt); 169 } 170 171 static struct rpc_rqst *rpcrdma_bc_rqst_get(struct rpcrdma_xprt *r_xprt) 172 { 173 struct rpc_xprt *xprt = &r_xprt->rx_xprt; 174 struct rpcrdma_req *req; 175 struct rpc_rqst *rqst; 176 size_t size; 177 178 spin_lock(&xprt->bc_pa_lock); 179 rqst = list_first_entry_or_null(&xprt->bc_pa_list, struct rpc_rqst, 180 rq_bc_pa_list); 181 if (!rqst) 182 goto create_req; 183 list_del(&rqst->rq_bc_pa_list); 184 spin_unlock(&xprt->bc_pa_lock); 185 return rqst; 186 187 create_req: 188 spin_unlock(&xprt->bc_pa_lock); 189 190 /* Set a limit to prevent a remote from overrunning our resources. 191 */ 192 if (xprt->bc_alloc_count >= RPCRDMA_BACKWARD_WRS) 193 return NULL; 194 195 size = min_t(size_t, r_xprt->rx_ep->re_inline_recv, PAGE_SIZE); 196 req = rpcrdma_req_create(r_xprt, size, GFP_KERNEL); 197 if (!req) 198 return NULL; 199 if (rpcrdma_req_setup(r_xprt, req)) { 200 rpcrdma_req_destroy(req); 201 return NULL; 202 } 203 204 xprt->bc_alloc_count++; 205 rqst = &req->rl_slot; 206 rqst->rq_xprt = xprt; 207 __set_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state); 208 xdr_buf_init(&rqst->rq_snd_buf, rdmab_data(req->rl_sendbuf), size); 209 return rqst; 210 } 211 212 /** 213 * rpcrdma_bc_receive_call - Handle a reverse-direction Call 214 * @r_xprt: transport receiving the call 215 * @rep: receive buffer containing the call 216 * 217 * Operational assumptions: 218 * o Backchannel credits are ignored, just as the NFS server 219 * forechannel currently does 220 * o The ULP manages a replay cache (eg, NFSv4.1 sessions). 221 * No replay detection is done at the transport level 222 */ 223 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt, 224 struct rpcrdma_rep *rep) 225 { 226 struct rpc_xprt *xprt = &r_xprt->rx_xprt; 227 struct svc_serv *bc_serv; 228 struct rpcrdma_req *req; 229 struct rpc_rqst *rqst; 230 struct xdr_buf *buf; 231 size_t size; 232 __be32 *p; 233 234 p = xdr_inline_decode(&rep->rr_stream, 0); 235 size = xdr_stream_remaining(&rep->rr_stream); 236 237 #ifdef RPCRDMA_BACKCHANNEL_DEBUG 238 pr_info("RPC: %s: callback XID %08x, length=%u\n", 239 __func__, be32_to_cpup(p), size); 240 pr_info("RPC: %s: %*ph\n", __func__, size, p); 241 #endif 242 243 rqst = rpcrdma_bc_rqst_get(r_xprt); 244 if (!rqst) 245 goto out_overflow; 246 247 rqst->rq_reply_bytes_recvd = 0; 248 rqst->rq_xid = *p; 249 250 rqst->rq_private_buf.len = size; 251 252 buf = &rqst->rq_rcv_buf; 253 memset(buf, 0, sizeof(*buf)); 254 buf->head[0].iov_base = p; 255 buf->head[0].iov_len = size; 256 buf->len = size; 257 258 /* The receive buffer has to be hooked to the rpcrdma_req 259 * so that it is not released while the req is pointing 260 * to its buffer, and so that it can be reposted after 261 * the Upper Layer is done decoding it. 262 */ 263 req = rpcr_to_rdmar(rqst); 264 req->rl_reply = rep; 265 trace_xprtrdma_cb_call(r_xprt, rqst); 266 267 /* Queue rqst for ULP's callback service */ 268 bc_serv = xprt->bc_serv; 269 xprt_get(xprt); 270 spin_lock(&bc_serv->sv_cb_lock); 271 list_add(&rqst->rq_bc_list, &bc_serv->sv_cb_list); 272 spin_unlock(&bc_serv->sv_cb_lock); 273 274 wake_up(&bc_serv->sv_cb_waitq); 275 276 r_xprt->rx_stats.bcall_count++; 277 return; 278 279 out_overflow: 280 pr_warn("RPC/RDMA backchannel overflow\n"); 281 xprt_force_disconnect(xprt); 282 /* This receive buffer gets reposted automatically 283 * when the connection is re-established. 284 */ 285 return; 286 } 287