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