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 53 size = r_xprt->rx_data.inline_rsize; 54 rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, GFP_KERNEL); 55 if (IS_ERR(rb)) 56 goto out_fail; 57 req->rl_sendbuf = rb; 58 xdr_buf_init(&rqst->rq_snd_buf, rb->rg_base, 59 min_t(size_t, size, PAGE_SIZE)); 60 rpcrdma_set_xprtdata(rqst, req); 61 return 0; 62 63 out_fail: 64 rpcrdma_bc_free_rqst(r_xprt, rqst); 65 return -ENOMEM; 66 } 67 68 /* Allocate and add receive buffers to the rpcrdma_buffer's 69 * existing list of rep's. These are released when the 70 * transport is destroyed. 71 */ 72 static int rpcrdma_bc_setup_reps(struct rpcrdma_xprt *r_xprt, 73 unsigned int count) 74 { 75 struct rpcrdma_rep *rep; 76 int rc = 0; 77 78 while (count--) { 79 rep = rpcrdma_create_rep(r_xprt); 80 if (IS_ERR(rep)) { 81 pr_err("RPC: %s: reply buffer alloc failed\n", 82 __func__); 83 rc = PTR_ERR(rep); 84 break; 85 } 86 87 rpcrdma_recv_buffer_put(rep); 88 } 89 90 return rc; 91 } 92 93 /** 94 * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests 95 * @xprt: transport associated with these backchannel resources 96 * @reqs: number of concurrent incoming requests to expect 97 * 98 * Returns 0 on success; otherwise a negative errno 99 */ 100 int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs) 101 { 102 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 103 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf; 104 struct rpc_rqst *rqst; 105 unsigned int i; 106 int rc; 107 108 /* The backchannel reply path returns each rpc_rqst to the 109 * bc_pa_list _after_ the reply is sent. If the server is 110 * faster than the client, it can send another backward 111 * direction request before the rpc_rqst is returned to the 112 * list. The client rejects the request in this case. 113 * 114 * Twice as many rpc_rqsts are prepared to ensure there is 115 * always an rpc_rqst available as soon as a reply is sent. 116 */ 117 if (reqs > RPCRDMA_BACKWARD_WRS >> 1) 118 goto out_err; 119 120 for (i = 0; i < (reqs << 1); i++) { 121 rqst = kzalloc(sizeof(*rqst), GFP_KERNEL); 122 if (!rqst) { 123 pr_err("RPC: %s: Failed to create bc rpc_rqst\n", 124 __func__); 125 goto out_free; 126 } 127 dprintk("RPC: %s: new rqst %p\n", __func__, rqst); 128 129 rqst->rq_xprt = &r_xprt->rx_xprt; 130 INIT_LIST_HEAD(&rqst->rq_list); 131 INIT_LIST_HEAD(&rqst->rq_bc_list); 132 133 if (rpcrdma_bc_setup_rqst(r_xprt, rqst)) 134 goto out_free; 135 136 spin_lock_bh(&xprt->bc_pa_lock); 137 list_add(&rqst->rq_bc_pa_list, &xprt->bc_pa_list); 138 spin_unlock_bh(&xprt->bc_pa_lock); 139 } 140 141 rc = rpcrdma_bc_setup_reps(r_xprt, reqs); 142 if (rc) 143 goto out_free; 144 145 rc = rpcrdma_ep_post_extra_recv(r_xprt, reqs); 146 if (rc) 147 goto out_free; 148 149 buffer->rb_bc_srv_max_requests = reqs; 150 request_module("svcrdma"); 151 152 return 0; 153 154 out_free: 155 xprt_rdma_bc_destroy(xprt, reqs); 156 157 out_err: 158 pr_err("RPC: %s: setup backchannel transport failed\n", __func__); 159 return -ENOMEM; 160 } 161 162 /** 163 * xprt_rdma_bc_up - Create transport endpoint for backchannel service 164 * @serv: server endpoint 165 * @net: network namespace 166 * 167 * The "xprt" is an implied argument: it supplies the name of the 168 * backchannel transport class. 169 * 170 * Returns zero on success, negative errno on failure 171 */ 172 int xprt_rdma_bc_up(struct svc_serv *serv, struct net *net) 173 { 174 int ret; 175 176 ret = svc_create_xprt(serv, "rdma-bc", net, PF_INET, 0, 0); 177 if (ret < 0) 178 return ret; 179 return 0; 180 } 181 182 /** 183 * xprt_rdma_bc_maxpayload - Return maximum backchannel message size 184 * @xprt: transport 185 * 186 * Returns maximum size, in bytes, of a backchannel message 187 */ 188 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *xprt) 189 { 190 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 191 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data; 192 size_t maxmsg; 193 194 maxmsg = min_t(unsigned int, cdata->inline_rsize, cdata->inline_wsize); 195 maxmsg = min_t(unsigned int, maxmsg, PAGE_SIZE); 196 return maxmsg - RPCRDMA_HDRLEN_MIN; 197 } 198 199 /** 200 * rpcrdma_bc_marshal_reply - Send backwards direction reply 201 * @rqst: buffer containing RPC reply data 202 * 203 * Returns zero on success. 204 */ 205 int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst) 206 { 207 struct rpc_xprt *xprt = rqst->rq_xprt; 208 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 209 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 210 struct rpcrdma_msg *headerp; 211 212 headerp = rdmab_to_msg(req->rl_rdmabuf); 213 headerp->rm_xid = rqst->rq_xid; 214 headerp->rm_vers = rpcrdma_version; 215 headerp->rm_credit = 216 cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests); 217 headerp->rm_type = rdma_msg; 218 headerp->rm_body.rm_chunks[0] = xdr_zero; 219 headerp->rm_body.rm_chunks[1] = xdr_zero; 220 headerp->rm_body.rm_chunks[2] = xdr_zero; 221 222 if (!rpcrdma_prepare_send_sges(&r_xprt->rx_ia, req, RPCRDMA_HDRLEN_MIN, 223 &rqst->rq_snd_buf, rpcrdma_noch)) 224 return -EIO; 225 return 0; 226 } 227 228 /** 229 * xprt_rdma_bc_destroy - Release resources for handling backchannel requests 230 * @xprt: transport associated with these backchannel resources 231 * @reqs: number of incoming requests to destroy; ignored 232 */ 233 void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs) 234 { 235 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 236 struct rpc_rqst *rqst, *tmp; 237 238 spin_lock_bh(&xprt->bc_pa_lock); 239 list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) { 240 list_del(&rqst->rq_bc_pa_list); 241 spin_unlock_bh(&xprt->bc_pa_lock); 242 243 rpcrdma_bc_free_rqst(r_xprt, rqst); 244 245 spin_lock_bh(&xprt->bc_pa_lock); 246 } 247 spin_unlock_bh(&xprt->bc_pa_lock); 248 } 249 250 /** 251 * xprt_rdma_bc_free_rqst - Release a backchannel rqst 252 * @rqst: request to release 253 */ 254 void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst) 255 { 256 struct rpc_xprt *xprt = rqst->rq_xprt; 257 258 dprintk("RPC: %s: freeing rqst %p (req %p)\n", 259 __func__, rqst, rpcr_to_rdmar(rqst)); 260 261 smp_mb__before_atomic(); 262 WARN_ON_ONCE(!test_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state)); 263 clear_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state); 264 smp_mb__after_atomic(); 265 266 spin_lock_bh(&xprt->bc_pa_lock); 267 list_add_tail(&rqst->rq_bc_pa_list, &xprt->bc_pa_list); 268 spin_unlock_bh(&xprt->bc_pa_lock); 269 } 270 271 /** 272 * rpcrdma_bc_receive_call - Handle a backward direction call 273 * @xprt: transport receiving the call 274 * @rep: receive buffer containing the call 275 * 276 * Called in the RPC reply handler, which runs in a tasklet. 277 * Be quick about it. 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 rpcrdma_msg *headerp; 290 struct svc_serv *bc_serv; 291 struct rpcrdma_req *req; 292 struct rpc_rqst *rqst; 293 struct xdr_buf *buf; 294 size_t size; 295 __be32 *p; 296 297 headerp = rdmab_to_msg(rep->rr_rdmabuf); 298 #ifdef RPCRDMA_BACKCHANNEL_DEBUG 299 pr_info("RPC: %s: callback XID %08x, length=%u\n", 300 __func__, be32_to_cpu(headerp->rm_xid), rep->rr_len); 301 pr_info("RPC: %s: %*ph\n", __func__, rep->rr_len, headerp); 302 #endif 303 304 /* Sanity check: 305 * Need at least enough bytes for RPC/RDMA header, as code 306 * here references the header fields by array offset. Also, 307 * backward calls are always inline, so ensure there 308 * are some bytes beyond the RPC/RDMA header. 309 */ 310 if (rep->rr_len < RPCRDMA_HDRLEN_MIN + 24) 311 goto out_short; 312 p = (__be32 *)((unsigned char *)headerp + RPCRDMA_HDRLEN_MIN); 313 size = rep->rr_len - RPCRDMA_HDRLEN_MIN; 314 315 /* Grab a free bc rqst */ 316 spin_lock(&xprt->bc_pa_lock); 317 if (list_empty(&xprt->bc_pa_list)) { 318 spin_unlock(&xprt->bc_pa_lock); 319 goto out_overflow; 320 } 321 rqst = list_first_entry(&xprt->bc_pa_list, 322 struct rpc_rqst, rq_bc_pa_list); 323 list_del(&rqst->rq_bc_pa_list); 324 spin_unlock(&xprt->bc_pa_lock); 325 dprintk("RPC: %s: using rqst %p\n", __func__, rqst); 326 327 /* Prepare rqst */ 328 rqst->rq_reply_bytes_recvd = 0; 329 rqst->rq_bytes_sent = 0; 330 rqst->rq_xid = headerp->rm_xid; 331 332 rqst->rq_private_buf.len = size; 333 set_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state); 334 335 buf = &rqst->rq_rcv_buf; 336 memset(buf, 0, sizeof(*buf)); 337 buf->head[0].iov_base = p; 338 buf->head[0].iov_len = size; 339 buf->len = size; 340 341 /* The receive buffer has to be hooked to the rpcrdma_req 342 * so that it can be reposted after the server is done 343 * parsing it but just before sending the backward 344 * direction reply. 345 */ 346 req = rpcr_to_rdmar(rqst); 347 dprintk("RPC: %s: attaching rep %p to req %p\n", 348 __func__, rep, req); 349 req->rl_reply = rep; 350 351 /* Defeat the retransmit detection logic in send_request */ 352 req->rl_connect_cookie = 0; 353 354 /* Queue rqst for ULP's callback service */ 355 bc_serv = xprt->bc_serv; 356 spin_lock(&bc_serv->sv_cb_lock); 357 list_add(&rqst->rq_bc_list, &bc_serv->sv_cb_list); 358 spin_unlock(&bc_serv->sv_cb_lock); 359 360 wake_up(&bc_serv->sv_cb_waitq); 361 362 r_xprt->rx_stats.bcall_count++; 363 return; 364 365 out_overflow: 366 pr_warn("RPC/RDMA backchannel overflow\n"); 367 xprt_disconnect_done(xprt); 368 /* This receive buffer gets reposted automatically 369 * when the connection is re-established. 370 */ 371 return; 372 373 out_short: 374 pr_warn("RPC/RDMA short backward direction call\n"); 375 376 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, rep)) 377 xprt_disconnect_done(xprt); 378 else 379 pr_warn("RPC: %s: reposting rep %p\n", 380 __func__, rep); 381 } 382