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 (server-side). 6 */ 7 8 #include <linux/module.h> 9 #include <linux/sunrpc/svc_rdma.h> 10 #include "xprt_rdma.h" 11 12 #define RPCDBG_FACILITY RPCDBG_SVCXPRT 13 14 #undef SVCRDMA_BACKCHANNEL_DEBUG 15 16 /** 17 * svc_rdma_handle_bc_reply - Process incoming backchannel reply 18 * @xprt: controlling backchannel transport 19 * @rdma_resp: pointer to incoming transport header 20 * @rcvbuf: XDR buffer into which to decode the reply 21 * 22 * Returns: 23 * %0 if @rcvbuf is filled in, xprt_complete_rqst called, 24 * %-EAGAIN if server should call ->recvfrom again. 25 */ 26 int svc_rdma_handle_bc_reply(struct rpc_xprt *xprt, __be32 *rdma_resp, 27 struct xdr_buf *rcvbuf) 28 { 29 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 30 struct kvec *dst, *src = &rcvbuf->head[0]; 31 struct rpc_rqst *req; 32 unsigned long cwnd; 33 u32 credits; 34 size_t len; 35 __be32 xid; 36 __be32 *p; 37 int ret; 38 39 p = (__be32 *)src->iov_base; 40 len = src->iov_len; 41 xid = *rdma_resp; 42 43 #ifdef SVCRDMA_BACKCHANNEL_DEBUG 44 pr_info("%s: xid=%08x, length=%zu\n", 45 __func__, be32_to_cpu(xid), len); 46 pr_info("%s: RPC/RDMA: %*ph\n", 47 __func__, (int)RPCRDMA_HDRLEN_MIN, rdma_resp); 48 pr_info("%s: RPC: %*ph\n", 49 __func__, (int)len, p); 50 #endif 51 52 ret = -EAGAIN; 53 if (src->iov_len < 24) 54 goto out_shortreply; 55 56 spin_lock(&xprt->recv_lock); 57 req = xprt_lookup_rqst(xprt, xid); 58 if (!req) 59 goto out_notfound; 60 61 dst = &req->rq_private_buf.head[0]; 62 memcpy(&req->rq_private_buf, &req->rq_rcv_buf, sizeof(struct xdr_buf)); 63 if (dst->iov_len < len) 64 goto out_unlock; 65 memcpy(dst->iov_base, p, len); 66 67 credits = be32_to_cpup(rdma_resp + 2); 68 if (credits == 0) 69 credits = 1; /* don't deadlock */ 70 else if (credits > r_xprt->rx_buf.rb_bc_max_requests) 71 credits = r_xprt->rx_buf.rb_bc_max_requests; 72 73 spin_lock_bh(&xprt->transport_lock); 74 cwnd = xprt->cwnd; 75 xprt->cwnd = credits << RPC_CWNDSHIFT; 76 if (xprt->cwnd > cwnd) 77 xprt_release_rqst_cong(req->rq_task); 78 spin_unlock_bh(&xprt->transport_lock); 79 80 81 ret = 0; 82 xprt_complete_rqst(req->rq_task, rcvbuf->len); 83 rcvbuf->len = 0; 84 85 out_unlock: 86 spin_unlock(&xprt->recv_lock); 87 out: 88 return ret; 89 90 out_shortreply: 91 dprintk("svcrdma: short bc reply: xprt=%p, len=%zu\n", 92 xprt, src->iov_len); 93 goto out; 94 95 out_notfound: 96 dprintk("svcrdma: unrecognized bc reply: xprt=%p, xid=%08x\n", 97 xprt, be32_to_cpu(xid)); 98 99 goto out_unlock; 100 } 101 102 /* Send a backwards direction RPC call. 103 * 104 * Caller holds the connection's mutex and has already marshaled 105 * the RPC/RDMA request. 106 * 107 * This is similar to svc_rdma_send_reply_msg, but takes a struct 108 * rpc_rqst instead, does not support chunks, and avoids blocking 109 * memory allocation. 110 * 111 * XXX: There is still an opportunity to block in svc_rdma_send() 112 * if there are no SQ entries to post the Send. This may occur if 113 * the adapter has a small maximum SQ depth. 114 */ 115 static int svc_rdma_bc_sendto(struct svcxprt_rdma *rdma, 116 struct rpc_rqst *rqst) 117 { 118 struct svc_rdma_op_ctxt *ctxt; 119 int ret; 120 121 ctxt = svc_rdma_get_context(rdma); 122 123 /* rpcrdma_bc_send_request builds the transport header and 124 * the backchannel RPC message in the same buffer. Thus only 125 * one SGE is needed to send both. 126 */ 127 ret = svc_rdma_map_reply_hdr(rdma, ctxt, rqst->rq_buffer, 128 rqst->rq_snd_buf.len); 129 if (ret < 0) 130 goto out_err; 131 132 ret = svc_rdma_repost_recv(rdma, GFP_NOIO); 133 if (ret) 134 goto out_err; 135 136 /* Bump page refcnt so Send completion doesn't release 137 * the rq_buffer before all retransmits are complete. 138 */ 139 get_page(virt_to_page(rqst->rq_buffer)); 140 ret = svc_rdma_post_send_wr(rdma, ctxt, 1, 0); 141 if (ret) 142 goto out_unmap; 143 144 out_err: 145 dprintk("svcrdma: %s returns %d\n", __func__, ret); 146 return ret; 147 148 out_unmap: 149 svc_rdma_unmap_dma(ctxt); 150 svc_rdma_put_context(ctxt, 1); 151 ret = -EIO; 152 goto out_err; 153 } 154 155 /* Server-side transport endpoint wants a whole page for its send 156 * buffer. The client RPC code constructs the RPC header in this 157 * buffer before it invokes ->send_request. 158 */ 159 static int 160 xprt_rdma_bc_allocate(struct rpc_task *task) 161 { 162 struct rpc_rqst *rqst = task->tk_rqstp; 163 size_t size = rqst->rq_callsize; 164 struct page *page; 165 166 if (size > PAGE_SIZE) { 167 WARN_ONCE(1, "svcrdma: large bc buffer request (size %zu)\n", 168 size); 169 return -EINVAL; 170 } 171 172 page = alloc_page(RPCRDMA_DEF_GFP); 173 if (!page) 174 return -ENOMEM; 175 rqst->rq_buffer = page_address(page); 176 177 rqst->rq_rbuffer = kmalloc(rqst->rq_rcvsize, RPCRDMA_DEF_GFP); 178 if (!rqst->rq_rbuffer) { 179 put_page(page); 180 return -ENOMEM; 181 } 182 return 0; 183 } 184 185 static void 186 xprt_rdma_bc_free(struct rpc_task *task) 187 { 188 struct rpc_rqst *rqst = task->tk_rqstp; 189 190 put_page(virt_to_page(rqst->rq_buffer)); 191 kfree(rqst->rq_rbuffer); 192 } 193 194 static int 195 rpcrdma_bc_send_request(struct svcxprt_rdma *rdma, struct rpc_rqst *rqst) 196 { 197 struct rpc_xprt *xprt = rqst->rq_xprt; 198 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 199 __be32 *p; 200 int rc; 201 202 /* Space in the send buffer for an RPC/RDMA header is reserved 203 * via xprt->tsh_size. 204 */ 205 p = rqst->rq_buffer; 206 *p++ = rqst->rq_xid; 207 *p++ = rpcrdma_version; 208 *p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_max_requests); 209 *p++ = rdma_msg; 210 *p++ = xdr_zero; 211 *p++ = xdr_zero; 212 *p = xdr_zero; 213 214 #ifdef SVCRDMA_BACKCHANNEL_DEBUG 215 pr_info("%s: %*ph\n", __func__, 64, rqst->rq_buffer); 216 #endif 217 218 rc = svc_rdma_bc_sendto(rdma, rqst); 219 if (rc) 220 goto drop_connection; 221 return rc; 222 223 drop_connection: 224 dprintk("svcrdma: failed to send bc call\n"); 225 xprt_disconnect_done(xprt); 226 return -ENOTCONN; 227 } 228 229 /* Send an RPC call on the passive end of a transport 230 * connection. 231 */ 232 static int 233 xprt_rdma_bc_send_request(struct rpc_task *task) 234 { 235 struct rpc_rqst *rqst = task->tk_rqstp; 236 struct svc_xprt *sxprt = rqst->rq_xprt->bc_xprt; 237 struct svcxprt_rdma *rdma; 238 int ret; 239 240 dprintk("svcrdma: sending bc call with xid: %08x\n", 241 be32_to_cpu(rqst->rq_xid)); 242 243 if (!mutex_trylock(&sxprt->xpt_mutex)) { 244 rpc_sleep_on(&sxprt->xpt_bc_pending, task, NULL); 245 if (!mutex_trylock(&sxprt->xpt_mutex)) 246 return -EAGAIN; 247 rpc_wake_up_queued_task(&sxprt->xpt_bc_pending, task); 248 } 249 250 ret = -ENOTCONN; 251 rdma = container_of(sxprt, struct svcxprt_rdma, sc_xprt); 252 if (!test_bit(XPT_DEAD, &sxprt->xpt_flags)) 253 ret = rpcrdma_bc_send_request(rdma, rqst); 254 255 mutex_unlock(&sxprt->xpt_mutex); 256 257 if (ret < 0) 258 return ret; 259 return 0; 260 } 261 262 static void 263 xprt_rdma_bc_close(struct rpc_xprt *xprt) 264 { 265 dprintk("svcrdma: %s: xprt %p\n", __func__, xprt); 266 } 267 268 static void 269 xprt_rdma_bc_put(struct rpc_xprt *xprt) 270 { 271 dprintk("svcrdma: %s: xprt %p\n", __func__, xprt); 272 273 xprt_free(xprt); 274 module_put(THIS_MODULE); 275 } 276 277 static const struct rpc_xprt_ops xprt_rdma_bc_procs = { 278 .reserve_xprt = xprt_reserve_xprt_cong, 279 .release_xprt = xprt_release_xprt_cong, 280 .alloc_slot = xprt_alloc_slot, 281 .release_request = xprt_release_rqst_cong, 282 .buf_alloc = xprt_rdma_bc_allocate, 283 .buf_free = xprt_rdma_bc_free, 284 .send_request = xprt_rdma_bc_send_request, 285 .set_retrans_timeout = xprt_set_retrans_timeout_def, 286 .close = xprt_rdma_bc_close, 287 .destroy = xprt_rdma_bc_put, 288 .print_stats = xprt_rdma_print_stats 289 }; 290 291 static const struct rpc_timeout xprt_rdma_bc_timeout = { 292 .to_initval = 60 * HZ, 293 .to_maxval = 60 * HZ, 294 }; 295 296 /* It shouldn't matter if the number of backchannel session slots 297 * doesn't match the number of RPC/RDMA credits. That just means 298 * one or the other will have extra slots that aren't used. 299 */ 300 static struct rpc_xprt * 301 xprt_setup_rdma_bc(struct xprt_create *args) 302 { 303 struct rpc_xprt *xprt; 304 struct rpcrdma_xprt *new_xprt; 305 306 if (args->addrlen > sizeof(xprt->addr)) { 307 dprintk("RPC: %s: address too large\n", __func__); 308 return ERR_PTR(-EBADF); 309 } 310 311 xprt = xprt_alloc(args->net, sizeof(*new_xprt), 312 RPCRDMA_MAX_BC_REQUESTS, 313 RPCRDMA_MAX_BC_REQUESTS); 314 if (!xprt) { 315 dprintk("RPC: %s: couldn't allocate rpc_xprt\n", 316 __func__); 317 return ERR_PTR(-ENOMEM); 318 } 319 320 xprt->timeout = &xprt_rdma_bc_timeout; 321 xprt_set_bound(xprt); 322 xprt_set_connected(xprt); 323 xprt->bind_timeout = RPCRDMA_BIND_TO; 324 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO; 325 xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO; 326 327 xprt->prot = XPRT_TRANSPORT_BC_RDMA; 328 xprt->tsh_size = RPCRDMA_HDRLEN_MIN / sizeof(__be32); 329 xprt->ops = &xprt_rdma_bc_procs; 330 331 memcpy(&xprt->addr, args->dstaddr, args->addrlen); 332 xprt->addrlen = args->addrlen; 333 xprt_rdma_format_addresses(xprt, (struct sockaddr *)&xprt->addr); 334 xprt->resvport = 0; 335 336 xprt->max_payload = xprt_rdma_max_inline_read; 337 338 new_xprt = rpcx_to_rdmax(xprt); 339 new_xprt->rx_buf.rb_bc_max_requests = xprt->max_reqs; 340 341 xprt_get(xprt); 342 args->bc_xprt->xpt_bc_xprt = xprt; 343 xprt->bc_xprt = args->bc_xprt; 344 345 if (!try_module_get(THIS_MODULE)) 346 goto out_fail; 347 348 /* Final put for backchannel xprt is in __svc_rdma_free */ 349 xprt_get(xprt); 350 return xprt; 351 352 out_fail: 353 xprt_rdma_free_addresses(xprt); 354 args->bc_xprt->xpt_bc_xprt = NULL; 355 args->bc_xprt->xpt_bc_xps = NULL; 356 xprt_put(xprt); 357 xprt_free(xprt); 358 return ERR_PTR(-EINVAL); 359 } 360 361 struct xprt_class xprt_rdma_bc = { 362 .list = LIST_HEAD_INIT(xprt_rdma_bc.list), 363 .name = "rdma backchannel", 364 .owner = THIS_MODULE, 365 .ident = XPRT_TRANSPORT_BC_RDMA, 366 .setup = xprt_setup_rdma_bc, 367 }; 368