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 goto out_unlock; 99 } 100 101 /* Send a backwards direction RPC call. 102 * 103 * Caller holds the connection's mutex and has already marshaled 104 * the RPC/RDMA request. 105 * 106 * This is similar to svc_rdma_send_reply_msg, but takes a struct 107 * rpc_rqst instead, does not support chunks, and avoids blocking 108 * memory allocation. 109 * 110 * XXX: There is still an opportunity to block in svc_rdma_send() 111 * if there are no SQ entries to post the Send. This may occur if 112 * the adapter has a small maximum SQ depth. 113 */ 114 static int svc_rdma_bc_sendto(struct svcxprt_rdma *rdma, 115 struct rpc_rqst *rqst) 116 { 117 struct svc_rdma_op_ctxt *ctxt; 118 int ret; 119 120 ctxt = svc_rdma_get_context(rdma); 121 122 /* rpcrdma_bc_send_request builds the transport header and 123 * the backchannel RPC message in the same buffer. Thus only 124 * one SGE is needed to send both. 125 */ 126 ret = svc_rdma_map_reply_hdr(rdma, ctxt, rqst->rq_buffer, 127 rqst->rq_snd_buf.len); 128 if (ret < 0) 129 goto out_err; 130 131 /* Bump page refcnt so Send completion doesn't release 132 * the rq_buffer before all retransmits are complete. 133 */ 134 get_page(virt_to_page(rqst->rq_buffer)); 135 ret = svc_rdma_post_send_wr(rdma, ctxt, 1, 0); 136 if (ret) 137 goto out_unmap; 138 139 out_err: 140 dprintk("svcrdma: %s returns %d\n", __func__, ret); 141 return ret; 142 143 out_unmap: 144 svc_rdma_unmap_dma(ctxt); 145 svc_rdma_put_context(ctxt, 1); 146 ret = -EIO; 147 goto out_err; 148 } 149 150 /* Server-side transport endpoint wants a whole page for its send 151 * buffer. The client RPC code constructs the RPC header in this 152 * buffer before it invokes ->send_request. 153 */ 154 static int 155 xprt_rdma_bc_allocate(struct rpc_task *task) 156 { 157 struct rpc_rqst *rqst = task->tk_rqstp; 158 size_t size = rqst->rq_callsize; 159 struct page *page; 160 161 if (size > PAGE_SIZE) { 162 WARN_ONCE(1, "svcrdma: large bc buffer request (size %zu)\n", 163 size); 164 return -EINVAL; 165 } 166 167 page = alloc_page(RPCRDMA_DEF_GFP); 168 if (!page) 169 return -ENOMEM; 170 rqst->rq_buffer = page_address(page); 171 172 rqst->rq_rbuffer = kmalloc(rqst->rq_rcvsize, RPCRDMA_DEF_GFP); 173 if (!rqst->rq_rbuffer) { 174 put_page(page); 175 return -ENOMEM; 176 } 177 return 0; 178 } 179 180 static void 181 xprt_rdma_bc_free(struct rpc_task *task) 182 { 183 struct rpc_rqst *rqst = task->tk_rqstp; 184 185 put_page(virt_to_page(rqst->rq_buffer)); 186 kfree(rqst->rq_rbuffer); 187 } 188 189 static int 190 rpcrdma_bc_send_request(struct svcxprt_rdma *rdma, struct rpc_rqst *rqst) 191 { 192 struct rpc_xprt *xprt = rqst->rq_xprt; 193 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 194 __be32 *p; 195 int rc; 196 197 /* Space in the send buffer for an RPC/RDMA header is reserved 198 * via xprt->tsh_size. 199 */ 200 p = rqst->rq_buffer; 201 *p++ = rqst->rq_xid; 202 *p++ = rpcrdma_version; 203 *p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_max_requests); 204 *p++ = rdma_msg; 205 *p++ = xdr_zero; 206 *p++ = xdr_zero; 207 *p = xdr_zero; 208 209 #ifdef SVCRDMA_BACKCHANNEL_DEBUG 210 pr_info("%s: %*ph\n", __func__, 64, rqst->rq_buffer); 211 #endif 212 213 rc = svc_rdma_bc_sendto(rdma, rqst); 214 if (rc) 215 goto drop_connection; 216 return rc; 217 218 drop_connection: 219 dprintk("svcrdma: failed to send bc call\n"); 220 xprt_disconnect_done(xprt); 221 return -ENOTCONN; 222 } 223 224 /* Send an RPC call on the passive end of a transport 225 * connection. 226 */ 227 static int 228 xprt_rdma_bc_send_request(struct rpc_task *task) 229 { 230 struct rpc_rqst *rqst = task->tk_rqstp; 231 struct svc_xprt *sxprt = rqst->rq_xprt->bc_xprt; 232 struct svcxprt_rdma *rdma; 233 int ret; 234 235 dprintk("svcrdma: sending bc call with xid: %08x\n", 236 be32_to_cpu(rqst->rq_xid)); 237 238 if (!mutex_trylock(&sxprt->xpt_mutex)) { 239 rpc_sleep_on(&sxprt->xpt_bc_pending, task, NULL); 240 if (!mutex_trylock(&sxprt->xpt_mutex)) 241 return -EAGAIN; 242 rpc_wake_up_queued_task(&sxprt->xpt_bc_pending, task); 243 } 244 245 ret = -ENOTCONN; 246 rdma = container_of(sxprt, struct svcxprt_rdma, sc_xprt); 247 if (!test_bit(XPT_DEAD, &sxprt->xpt_flags)) 248 ret = rpcrdma_bc_send_request(rdma, rqst); 249 250 mutex_unlock(&sxprt->xpt_mutex); 251 252 if (ret < 0) 253 return ret; 254 return 0; 255 } 256 257 static void 258 xprt_rdma_bc_close(struct rpc_xprt *xprt) 259 { 260 dprintk("svcrdma: %s: xprt %p\n", __func__, xprt); 261 } 262 263 static void 264 xprt_rdma_bc_put(struct rpc_xprt *xprt) 265 { 266 dprintk("svcrdma: %s: xprt %p\n", __func__, xprt); 267 268 xprt_free(xprt); 269 module_put(THIS_MODULE); 270 } 271 272 static const struct rpc_xprt_ops xprt_rdma_bc_procs = { 273 .reserve_xprt = xprt_reserve_xprt_cong, 274 .release_xprt = xprt_release_xprt_cong, 275 .alloc_slot = xprt_alloc_slot, 276 .release_request = xprt_release_rqst_cong, 277 .buf_alloc = xprt_rdma_bc_allocate, 278 .buf_free = xprt_rdma_bc_free, 279 .send_request = xprt_rdma_bc_send_request, 280 .set_retrans_timeout = xprt_set_retrans_timeout_def, 281 .close = xprt_rdma_bc_close, 282 .destroy = xprt_rdma_bc_put, 283 .print_stats = xprt_rdma_print_stats 284 }; 285 286 static const struct rpc_timeout xprt_rdma_bc_timeout = { 287 .to_initval = 60 * HZ, 288 .to_maxval = 60 * HZ, 289 }; 290 291 /* It shouldn't matter if the number of backchannel session slots 292 * doesn't match the number of RPC/RDMA credits. That just means 293 * one or the other will have extra slots that aren't used. 294 */ 295 static struct rpc_xprt * 296 xprt_setup_rdma_bc(struct xprt_create *args) 297 { 298 struct rpc_xprt *xprt; 299 struct rpcrdma_xprt *new_xprt; 300 301 if (args->addrlen > sizeof(xprt->addr)) { 302 dprintk("RPC: %s: address too large\n", __func__); 303 return ERR_PTR(-EBADF); 304 } 305 306 xprt = xprt_alloc(args->net, sizeof(*new_xprt), 307 RPCRDMA_MAX_BC_REQUESTS, 308 RPCRDMA_MAX_BC_REQUESTS); 309 if (!xprt) { 310 dprintk("RPC: %s: couldn't allocate rpc_xprt\n", 311 __func__); 312 return ERR_PTR(-ENOMEM); 313 } 314 315 xprt->timeout = &xprt_rdma_bc_timeout; 316 xprt_set_bound(xprt); 317 xprt_set_connected(xprt); 318 xprt->bind_timeout = RPCRDMA_BIND_TO; 319 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO; 320 xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO; 321 322 xprt->prot = XPRT_TRANSPORT_BC_RDMA; 323 xprt->tsh_size = RPCRDMA_HDRLEN_MIN / sizeof(__be32); 324 xprt->ops = &xprt_rdma_bc_procs; 325 326 memcpy(&xprt->addr, args->dstaddr, args->addrlen); 327 xprt->addrlen = args->addrlen; 328 xprt_rdma_format_addresses(xprt, (struct sockaddr *)&xprt->addr); 329 xprt->resvport = 0; 330 331 xprt->max_payload = xprt_rdma_max_inline_read; 332 333 new_xprt = rpcx_to_rdmax(xprt); 334 new_xprt->rx_buf.rb_bc_max_requests = xprt->max_reqs; 335 336 xprt_get(xprt); 337 args->bc_xprt->xpt_bc_xprt = xprt; 338 xprt->bc_xprt = args->bc_xprt; 339 340 if (!try_module_get(THIS_MODULE)) 341 goto out_fail; 342 343 /* Final put for backchannel xprt is in __svc_rdma_free */ 344 xprt_get(xprt); 345 return xprt; 346 347 out_fail: 348 xprt_rdma_free_addresses(xprt); 349 args->bc_xprt->xpt_bc_xprt = NULL; 350 args->bc_xprt->xpt_bc_xps = NULL; 351 xprt_put(xprt); 352 xprt_free(xprt); 353 return ERR_PTR(-EINVAL); 354 } 355 356 struct xprt_class xprt_rdma_bc = { 357 .list = LIST_HEAD_INIT(xprt_rdma_bc.list), 358 .name = "rdma backchannel", 359 .owner = THIS_MODULE, 360 .ident = XPRT_TRANSPORT_BC_RDMA, 361 .setup = xprt_setup_rdma_bc, 362 }; 363