1 /* 2 * Copyright (c) 2015 Oracle. All rights reserved. 3 * 4 * Support for backward direction RPCs on RPC/RDMA. 5 */ 6 7 #include <linux/module.h> 8 #include <linux/sunrpc/xprt.h> 9 #include <linux/sunrpc/svc.h> 10 #include <linux/sunrpc/svc_xprt.h> 11 12 #include "xprt_rdma.h" 13 14 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) 15 # define RPCDBG_FACILITY RPCDBG_TRANS 16 #endif 17 18 #undef RPCRDMA_BACKCHANNEL_DEBUG 19 20 static void rpcrdma_bc_free_rqst(struct rpcrdma_xprt *r_xprt, 21 struct rpc_rqst *rqst) 22 { 23 struct rpcrdma_buffer *buf = &r_xprt->rx_buf; 24 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 25 26 spin_lock(&buf->rb_reqslock); 27 list_del(&req->rl_all); 28 spin_unlock(&buf->rb_reqslock); 29 30 rpcrdma_destroy_req(req); 31 32 kfree(rqst); 33 } 34 35 static int rpcrdma_bc_setup_rqst(struct rpcrdma_xprt *r_xprt, 36 struct rpc_rqst *rqst) 37 { 38 struct rpcrdma_regbuf *rb; 39 struct rpcrdma_req *req; 40 size_t size; 41 42 req = rpcrdma_create_req(r_xprt); 43 if (IS_ERR(req)) 44 return PTR_ERR(req); 45 req->rl_backchannel = true; 46 47 rb = rpcrdma_alloc_regbuf(RPCRDMA_HDRBUF_SIZE, 48 DMA_TO_DEVICE, GFP_KERNEL); 49 if (IS_ERR(rb)) 50 goto out_fail; 51 req->rl_rdmabuf = rb; 52 xdr_buf_init(&req->rl_hdrbuf, rb->rg_base, rdmab_length(rb)); 53 54 size = r_xprt->rx_data.inline_rsize; 55 rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, GFP_KERNEL); 56 if (IS_ERR(rb)) 57 goto out_fail; 58 req->rl_sendbuf = rb; 59 xdr_buf_init(&rqst->rq_snd_buf, rb->rg_base, 60 min_t(size_t, size, PAGE_SIZE)); 61 rpcrdma_set_xprtdata(rqst, req); 62 return 0; 63 64 out_fail: 65 rpcrdma_bc_free_rqst(r_xprt, rqst); 66 return -ENOMEM; 67 } 68 69 /* Allocate and add receive buffers to the rpcrdma_buffer's 70 * existing list of rep's. These are released when the 71 * transport is destroyed. 72 */ 73 static int rpcrdma_bc_setup_reps(struct rpcrdma_xprt *r_xprt, 74 unsigned int count) 75 { 76 struct rpcrdma_rep *rep; 77 int rc = 0; 78 79 while (count--) { 80 rep = rpcrdma_create_rep(r_xprt); 81 if (IS_ERR(rep)) { 82 pr_err("RPC: %s: reply buffer alloc failed\n", 83 __func__); 84 rc = PTR_ERR(rep); 85 break; 86 } 87 88 rpcrdma_recv_buffer_put(rep); 89 } 90 91 return rc; 92 } 93 94 /** 95 * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests 96 * @xprt: transport associated with these backchannel resources 97 * @reqs: number of concurrent incoming requests to expect 98 * 99 * Returns 0 on success; otherwise a negative errno 100 */ 101 int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs) 102 { 103 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 104 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf; 105 struct rpc_rqst *rqst; 106 unsigned int i; 107 int rc; 108 109 /* The backchannel reply path returns each rpc_rqst to the 110 * bc_pa_list _after_ the reply is sent. If the server is 111 * faster than the client, it can send another backward 112 * direction request before the rpc_rqst is returned to the 113 * list. The client rejects the request in this case. 114 * 115 * Twice as many rpc_rqsts are prepared to ensure there is 116 * always an rpc_rqst available as soon as a reply is sent. 117 */ 118 if (reqs > RPCRDMA_BACKWARD_WRS >> 1) 119 goto out_err; 120 121 for (i = 0; i < (reqs << 1); i++) { 122 rqst = kzalloc(sizeof(*rqst), GFP_KERNEL); 123 if (!rqst) 124 goto out_free; 125 126 dprintk("RPC: %s: new rqst %p\n", __func__, rqst); 127 128 rqst->rq_xprt = &r_xprt->rx_xprt; 129 INIT_LIST_HEAD(&rqst->rq_list); 130 INIT_LIST_HEAD(&rqst->rq_bc_list); 131 132 if (rpcrdma_bc_setup_rqst(r_xprt, rqst)) 133 goto out_free; 134 135 spin_lock_bh(&xprt->bc_pa_lock); 136 list_add(&rqst->rq_bc_pa_list, &xprt->bc_pa_list); 137 spin_unlock_bh(&xprt->bc_pa_lock); 138 } 139 140 rc = rpcrdma_bc_setup_reps(r_xprt, reqs); 141 if (rc) 142 goto out_free; 143 144 rc = rpcrdma_ep_post_extra_recv(r_xprt, reqs); 145 if (rc) 146 goto out_free; 147 148 buffer->rb_bc_srv_max_requests = reqs; 149 request_module("svcrdma"); 150 151 return 0; 152 153 out_free: 154 xprt_rdma_bc_destroy(xprt, reqs); 155 156 out_err: 157 pr_err("RPC: %s: setup backchannel transport failed\n", __func__); 158 return -ENOMEM; 159 } 160 161 /** 162 * xprt_rdma_bc_up - Create transport endpoint for backchannel service 163 * @serv: server endpoint 164 * @net: network namespace 165 * 166 * The "xprt" is an implied argument: it supplies the name of the 167 * backchannel transport class. 168 * 169 * Returns zero on success, negative errno on failure 170 */ 171 int xprt_rdma_bc_up(struct svc_serv *serv, struct net *net) 172 { 173 int ret; 174 175 ret = svc_create_xprt(serv, "rdma-bc", net, PF_INET, 0, 0); 176 if (ret < 0) 177 return ret; 178 return 0; 179 } 180 181 /** 182 * xprt_rdma_bc_maxpayload - Return maximum backchannel message size 183 * @xprt: transport 184 * 185 * Returns maximum size, in bytes, of a backchannel message 186 */ 187 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *xprt) 188 { 189 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 190 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data; 191 size_t maxmsg; 192 193 maxmsg = min_t(unsigned int, cdata->inline_rsize, cdata->inline_wsize); 194 maxmsg = min_t(unsigned int, maxmsg, PAGE_SIZE); 195 return maxmsg - RPCRDMA_HDRLEN_MIN; 196 } 197 198 /** 199 * rpcrdma_bc_marshal_reply - Send backwards direction reply 200 * @rqst: buffer containing RPC reply data 201 * 202 * Returns zero on success. 203 */ 204 int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst) 205 { 206 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt); 207 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 208 __be32 *p; 209 210 rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0); 211 xdr_init_encode(&req->rl_stream, &req->rl_hdrbuf, 212 req->rl_rdmabuf->rg_base); 213 214 p = xdr_reserve_space(&req->rl_stream, 28); 215 if (unlikely(!p)) 216 return -EIO; 217 *p++ = rqst->rq_xid; 218 *p++ = rpcrdma_version; 219 *p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests); 220 *p++ = rdma_msg; 221 *p++ = xdr_zero; 222 *p++ = xdr_zero; 223 *p = xdr_zero; 224 225 if (!rpcrdma_prepare_send_sges(&r_xprt->rx_ia, req, RPCRDMA_HDRLEN_MIN, 226 &rqst->rq_snd_buf, rpcrdma_noch)) 227 return -EIO; 228 return 0; 229 } 230 231 /** 232 * xprt_rdma_bc_destroy - Release resources for handling backchannel requests 233 * @xprt: transport associated with these backchannel resources 234 * @reqs: number of incoming requests to destroy; ignored 235 */ 236 void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs) 237 { 238 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 239 struct rpc_rqst *rqst, *tmp; 240 241 spin_lock_bh(&xprt->bc_pa_lock); 242 list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) { 243 list_del(&rqst->rq_bc_pa_list); 244 spin_unlock_bh(&xprt->bc_pa_lock); 245 246 rpcrdma_bc_free_rqst(r_xprt, rqst); 247 248 spin_lock_bh(&xprt->bc_pa_lock); 249 } 250 spin_unlock_bh(&xprt->bc_pa_lock); 251 } 252 253 /** 254 * xprt_rdma_bc_free_rqst - Release a backchannel rqst 255 * @rqst: request to release 256 */ 257 void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst) 258 { 259 struct rpc_xprt *xprt = rqst->rq_xprt; 260 261 dprintk("RPC: %s: freeing rqst %p (req %p)\n", 262 __func__, rqst, rpcr_to_rdmar(rqst)); 263 264 smp_mb__before_atomic(); 265 WARN_ON_ONCE(!test_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state)); 266 clear_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state); 267 smp_mb__after_atomic(); 268 269 spin_lock_bh(&xprt->bc_pa_lock); 270 list_add_tail(&rqst->rq_bc_pa_list, &xprt->bc_pa_list); 271 spin_unlock_bh(&xprt->bc_pa_lock); 272 } 273 274 /** 275 * rpcrdma_bc_receive_call - Handle a backward direction call 276 * @xprt: transport receiving the call 277 * @rep: receive buffer containing the call 278 * 279 * Operational assumptions: 280 * o Backchannel credits are ignored, just as the NFS server 281 * forechannel currently does 282 * o The ULP manages a replay cache (eg, NFSv4.1 sessions). 283 * No replay detection is done at the transport level 284 */ 285 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt, 286 struct rpcrdma_rep *rep) 287 { 288 struct rpc_xprt *xprt = &r_xprt->rx_xprt; 289 struct svc_serv *bc_serv; 290 struct rpcrdma_req *req; 291 struct rpc_rqst *rqst; 292 struct xdr_buf *buf; 293 size_t size; 294 __be32 *p; 295 296 p = xdr_inline_decode(&rep->rr_stream, 0); 297 size = xdr_stream_remaining(&rep->rr_stream); 298 299 #ifdef RPCRDMA_BACKCHANNEL_DEBUG 300 pr_info("RPC: %s: callback XID %08x, length=%u\n", 301 __func__, be32_to_cpup(p), size); 302 pr_info("RPC: %s: %*ph\n", __func__, size, p); 303 #endif 304 305 /* Grab a free bc rqst */ 306 spin_lock(&xprt->bc_pa_lock); 307 if (list_empty(&xprt->bc_pa_list)) { 308 spin_unlock(&xprt->bc_pa_lock); 309 goto out_overflow; 310 } 311 rqst = list_first_entry(&xprt->bc_pa_list, 312 struct rpc_rqst, rq_bc_pa_list); 313 list_del(&rqst->rq_bc_pa_list); 314 spin_unlock(&xprt->bc_pa_lock); 315 dprintk("RPC: %s: using rqst %p\n", __func__, rqst); 316 317 /* Prepare rqst */ 318 rqst->rq_reply_bytes_recvd = 0; 319 rqst->rq_bytes_sent = 0; 320 rqst->rq_xid = *p; 321 322 rqst->rq_private_buf.len = size; 323 set_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state); 324 325 buf = &rqst->rq_rcv_buf; 326 memset(buf, 0, sizeof(*buf)); 327 buf->head[0].iov_base = p; 328 buf->head[0].iov_len = size; 329 buf->len = size; 330 331 /* The receive buffer has to be hooked to the rpcrdma_req 332 * so that it is not released while the req is pointing 333 * to its buffer, and so that it can be reposted after 334 * the Upper Layer is done decoding it. 335 */ 336 req = rpcr_to_rdmar(rqst); 337 dprintk("RPC: %s: attaching rep %p to req %p\n", 338 __func__, rep, req); 339 req->rl_reply = rep; 340 341 /* Defeat the retransmit detection logic in send_request */ 342 req->rl_connect_cookie = 0; 343 344 /* Queue rqst for ULP's callback service */ 345 bc_serv = xprt->bc_serv; 346 spin_lock(&bc_serv->sv_cb_lock); 347 list_add(&rqst->rq_bc_list, &bc_serv->sv_cb_list); 348 spin_unlock(&bc_serv->sv_cb_lock); 349 350 wake_up(&bc_serv->sv_cb_waitq); 351 352 r_xprt->rx_stats.bcall_count++; 353 return; 354 355 out_overflow: 356 pr_warn("RPC/RDMA backchannel overflow\n"); 357 xprt_disconnect_done(xprt); 358 /* This receive buffer gets reposted automatically 359 * when the connection is re-established. 360 */ 361 return; 362 } 363