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(&r_xprt->rx_ia, 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_ia *ia = &r_xprt->rx_ia; 39 struct rpcrdma_regbuf *rb; 40 struct rpcrdma_req *req; 41 struct xdr_buf *buf; 42 size_t size; 43 44 req = rpcrdma_create_req(r_xprt); 45 if (IS_ERR(req)) 46 return PTR_ERR(req); 47 req->rl_backchannel = true; 48 49 size = RPCRDMA_INLINE_WRITE_THRESHOLD(rqst); 50 rb = rpcrdma_alloc_regbuf(ia, size, GFP_KERNEL); 51 if (IS_ERR(rb)) 52 goto out_fail; 53 req->rl_rdmabuf = rb; 54 55 size += RPCRDMA_INLINE_READ_THRESHOLD(rqst); 56 rb = rpcrdma_alloc_regbuf(ia, size, GFP_KERNEL); 57 if (IS_ERR(rb)) 58 goto out_fail; 59 rb->rg_owner = req; 60 req->rl_sendbuf = rb; 61 /* so that rpcr_to_rdmar works when receiving a request */ 62 rqst->rq_buffer = (void *)req->rl_sendbuf->rg_base; 63 64 buf = &rqst->rq_snd_buf; 65 buf->head[0].iov_base = rqst->rq_buffer; 66 buf->head[0].iov_len = 0; 67 buf->tail[0].iov_base = NULL; 68 buf->tail[0].iov_len = 0; 69 buf->page_len = 0; 70 buf->len = 0; 71 buf->buflen = size; 72 73 return 0; 74 75 out_fail: 76 rpcrdma_bc_free_rqst(r_xprt, rqst); 77 return -ENOMEM; 78 } 79 80 /* Allocate and add receive buffers to the rpcrdma_buffer's 81 * existing list of rep's. These are released when the 82 * transport is destroyed. 83 */ 84 static int rpcrdma_bc_setup_reps(struct rpcrdma_xprt *r_xprt, 85 unsigned int count) 86 { 87 struct rpcrdma_rep *rep; 88 int rc = 0; 89 90 while (count--) { 91 rep = rpcrdma_create_rep(r_xprt); 92 if (IS_ERR(rep)) { 93 pr_err("RPC: %s: reply buffer alloc failed\n", 94 __func__); 95 rc = PTR_ERR(rep); 96 break; 97 } 98 99 rpcrdma_recv_buffer_put(rep); 100 } 101 102 return rc; 103 } 104 105 /** 106 * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests 107 * @xprt: transport associated with these backchannel resources 108 * @reqs: number of concurrent incoming requests to expect 109 * 110 * Returns 0 on success; otherwise a negative errno 111 */ 112 int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs) 113 { 114 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 115 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf; 116 struct rpc_rqst *rqst; 117 unsigned int i; 118 int rc; 119 120 /* The backchannel reply path returns each rpc_rqst to the 121 * bc_pa_list _after_ the reply is sent. If the server is 122 * faster than the client, it can send another backward 123 * direction request before the rpc_rqst is returned to the 124 * list. The client rejects the request in this case. 125 * 126 * Twice as many rpc_rqsts are prepared to ensure there is 127 * always an rpc_rqst available as soon as a reply is sent. 128 */ 129 if (reqs > RPCRDMA_BACKWARD_WRS >> 1) 130 goto out_err; 131 132 for (i = 0; i < (reqs << 1); i++) { 133 rqst = kzalloc(sizeof(*rqst), GFP_KERNEL); 134 if (!rqst) { 135 pr_err("RPC: %s: Failed to create bc rpc_rqst\n", 136 __func__); 137 goto out_free; 138 } 139 dprintk("RPC: %s: new rqst %p\n", __func__, rqst); 140 141 rqst->rq_xprt = &r_xprt->rx_xprt; 142 INIT_LIST_HEAD(&rqst->rq_list); 143 INIT_LIST_HEAD(&rqst->rq_bc_list); 144 145 if (rpcrdma_bc_setup_rqst(r_xprt, rqst)) 146 goto out_free; 147 148 spin_lock_bh(&xprt->bc_pa_lock); 149 list_add(&rqst->rq_bc_pa_list, &xprt->bc_pa_list); 150 spin_unlock_bh(&xprt->bc_pa_lock); 151 } 152 153 rc = rpcrdma_bc_setup_reps(r_xprt, reqs); 154 if (rc) 155 goto out_free; 156 157 rc = rpcrdma_ep_post_extra_recv(r_xprt, reqs); 158 if (rc) 159 goto out_free; 160 161 buffer->rb_bc_srv_max_requests = reqs; 162 request_module("svcrdma"); 163 164 return 0; 165 166 out_free: 167 xprt_rdma_bc_destroy(xprt, reqs); 168 169 out_err: 170 pr_err("RPC: %s: setup backchannel transport failed\n", __func__); 171 return -ENOMEM; 172 } 173 174 /** 175 * xprt_rdma_bc_up - Create transport endpoint for backchannel service 176 * @serv: server endpoint 177 * @net: network namespace 178 * 179 * The "xprt" is an implied argument: it supplies the name of the 180 * backchannel transport class. 181 * 182 * Returns zero on success, negative errno on failure 183 */ 184 int xprt_rdma_bc_up(struct svc_serv *serv, struct net *net) 185 { 186 int ret; 187 188 ret = svc_create_xprt(serv, "rdma-bc", net, PF_INET, 0, 0); 189 if (ret < 0) 190 return ret; 191 return 0; 192 } 193 194 /** 195 * rpcrdma_bc_marshal_reply - Send backwards direction reply 196 * @rqst: buffer containing RPC reply data 197 * 198 * Returns zero on success. 199 */ 200 int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst) 201 { 202 struct rpc_xprt *xprt = rqst->rq_xprt; 203 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 204 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 205 struct rpcrdma_msg *headerp; 206 size_t rpclen; 207 208 headerp = rdmab_to_msg(req->rl_rdmabuf); 209 headerp->rm_xid = rqst->rq_xid; 210 headerp->rm_vers = rpcrdma_version; 211 headerp->rm_credit = 212 cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests); 213 headerp->rm_type = rdma_msg; 214 headerp->rm_body.rm_chunks[0] = xdr_zero; 215 headerp->rm_body.rm_chunks[1] = xdr_zero; 216 headerp->rm_body.rm_chunks[2] = xdr_zero; 217 218 rpclen = rqst->rq_svec[0].iov_len; 219 220 #ifdef RPCRDMA_BACKCHANNEL_DEBUG 221 pr_info("RPC: %s: rpclen %zd headerp 0x%p lkey 0x%x\n", 222 __func__, rpclen, headerp, rdmab_lkey(req->rl_rdmabuf)); 223 pr_info("RPC: %s: RPC/RDMA: %*ph\n", 224 __func__, (int)RPCRDMA_HDRLEN_MIN, headerp); 225 pr_info("RPC: %s: RPC: %*ph\n", 226 __func__, (int)rpclen, rqst->rq_svec[0].iov_base); 227 #endif 228 229 req->rl_send_iov[0].addr = rdmab_addr(req->rl_rdmabuf); 230 req->rl_send_iov[0].length = RPCRDMA_HDRLEN_MIN; 231 req->rl_send_iov[0].lkey = rdmab_lkey(req->rl_rdmabuf); 232 233 req->rl_send_iov[1].addr = rdmab_addr(req->rl_sendbuf); 234 req->rl_send_iov[1].length = rpclen; 235 req->rl_send_iov[1].lkey = rdmab_lkey(req->rl_sendbuf); 236 237 req->rl_niovs = 2; 238 return 0; 239 } 240 241 /** 242 * xprt_rdma_bc_destroy - Release resources for handling backchannel requests 243 * @xprt: transport associated with these backchannel resources 244 * @reqs: number of incoming requests to destroy; ignored 245 */ 246 void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs) 247 { 248 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 249 struct rpc_rqst *rqst, *tmp; 250 251 spin_lock_bh(&xprt->bc_pa_lock); 252 list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) { 253 list_del(&rqst->rq_bc_pa_list); 254 spin_unlock_bh(&xprt->bc_pa_lock); 255 256 rpcrdma_bc_free_rqst(r_xprt, rqst); 257 258 spin_lock_bh(&xprt->bc_pa_lock); 259 } 260 spin_unlock_bh(&xprt->bc_pa_lock); 261 } 262 263 /** 264 * xprt_rdma_bc_free_rqst - Release a backchannel rqst 265 * @rqst: request to release 266 */ 267 void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst) 268 { 269 struct rpc_xprt *xprt = rqst->rq_xprt; 270 271 dprintk("RPC: %s: freeing rqst %p (req %p)\n", 272 __func__, rqst, rpcr_to_rdmar(rqst)); 273 274 smp_mb__before_atomic(); 275 WARN_ON_ONCE(!test_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state)); 276 clear_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state); 277 smp_mb__after_atomic(); 278 279 spin_lock_bh(&xprt->bc_pa_lock); 280 list_add_tail(&rqst->rq_bc_pa_list, &xprt->bc_pa_list); 281 spin_unlock_bh(&xprt->bc_pa_lock); 282 } 283 284 /** 285 * rpcrdma_bc_receive_call - Handle a backward direction call 286 * @xprt: transport receiving the call 287 * @rep: receive buffer containing the call 288 * 289 * Called in the RPC reply handler, which runs in a tasklet. 290 * Be quick about it. 291 * 292 * Operational assumptions: 293 * o Backchannel credits are ignored, just as the NFS server 294 * forechannel currently does 295 * o The ULP manages a replay cache (eg, NFSv4.1 sessions). 296 * No replay detection is done at the transport level 297 */ 298 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt, 299 struct rpcrdma_rep *rep) 300 { 301 struct rpc_xprt *xprt = &r_xprt->rx_xprt; 302 struct rpcrdma_msg *headerp; 303 struct svc_serv *bc_serv; 304 struct rpcrdma_req *req; 305 struct rpc_rqst *rqst; 306 struct xdr_buf *buf; 307 size_t size; 308 __be32 *p; 309 310 headerp = rdmab_to_msg(rep->rr_rdmabuf); 311 #ifdef RPCRDMA_BACKCHANNEL_DEBUG 312 pr_info("RPC: %s: callback XID %08x, length=%u\n", 313 __func__, be32_to_cpu(headerp->rm_xid), rep->rr_len); 314 pr_info("RPC: %s: %*ph\n", __func__, rep->rr_len, headerp); 315 #endif 316 317 /* Sanity check: 318 * Need at least enough bytes for RPC/RDMA header, as code 319 * here references the header fields by array offset. Also, 320 * backward calls are always inline, so ensure there 321 * are some bytes beyond the RPC/RDMA header. 322 */ 323 if (rep->rr_len < RPCRDMA_HDRLEN_MIN + 24) 324 goto out_short; 325 p = (__be32 *)((unsigned char *)headerp + RPCRDMA_HDRLEN_MIN); 326 size = rep->rr_len - RPCRDMA_HDRLEN_MIN; 327 328 /* Grab a free bc rqst */ 329 spin_lock(&xprt->bc_pa_lock); 330 if (list_empty(&xprt->bc_pa_list)) { 331 spin_unlock(&xprt->bc_pa_lock); 332 goto out_overflow; 333 } 334 rqst = list_first_entry(&xprt->bc_pa_list, 335 struct rpc_rqst, rq_bc_pa_list); 336 list_del(&rqst->rq_bc_pa_list); 337 spin_unlock(&xprt->bc_pa_lock); 338 dprintk("RPC: %s: using rqst %p\n", __func__, rqst); 339 340 /* Prepare rqst */ 341 rqst->rq_reply_bytes_recvd = 0; 342 rqst->rq_bytes_sent = 0; 343 rqst->rq_xid = headerp->rm_xid; 344 345 rqst->rq_private_buf.len = size; 346 set_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state); 347 348 buf = &rqst->rq_rcv_buf; 349 memset(buf, 0, sizeof(*buf)); 350 buf->head[0].iov_base = p; 351 buf->head[0].iov_len = size; 352 buf->len = size; 353 354 /* The receive buffer has to be hooked to the rpcrdma_req 355 * so that it can be reposted after the server is done 356 * parsing it but just before sending the backward 357 * direction reply. 358 */ 359 req = rpcr_to_rdmar(rqst); 360 dprintk("RPC: %s: attaching rep %p to req %p\n", 361 __func__, rep, req); 362 req->rl_reply = rep; 363 364 /* Defeat the retransmit detection logic in send_request */ 365 req->rl_connect_cookie = 0; 366 367 /* Queue rqst for ULP's callback service */ 368 bc_serv = xprt->bc_serv; 369 spin_lock(&bc_serv->sv_cb_lock); 370 list_add(&rqst->rq_bc_list, &bc_serv->sv_cb_list); 371 spin_unlock(&bc_serv->sv_cb_lock); 372 373 wake_up(&bc_serv->sv_cb_waitq); 374 375 r_xprt->rx_stats.bcall_count++; 376 return; 377 378 out_overflow: 379 pr_warn("RPC/RDMA backchannel overflow\n"); 380 xprt_disconnect_done(xprt); 381 /* This receive buffer gets reposted automatically 382 * when the connection is re-established. 383 */ 384 return; 385 386 out_short: 387 pr_warn("RPC/RDMA short backward direction call\n"); 388 389 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep)) 390 xprt_disconnect_done(xprt); 391 else 392 pr_warn("RPC: %s: reposting rep %p\n", 393 __func__, rep); 394 } 395