1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (c) 2015-2018 Oracle. All rights reserved. 4 * 5 * Support for backward direction RPCs on RPC/RDMA (server-side). 6 */ 7 8 #include <linux/sunrpc/svc_rdma.h> 9 10 #include "xprt_rdma.h" 11 #include <trace/events/rpcrdma.h> 12 13 #define RPCDBG_FACILITY RPCDBG_SVCXPRT 14 15 #undef SVCRDMA_BACKCHANNEL_DEBUG 16 17 /** 18 * svc_rdma_handle_bc_reply - Process incoming backchannel reply 19 * @xprt: controlling backchannel transport 20 * @rdma_resp: pointer to incoming transport header 21 * @rcvbuf: XDR buffer into which to decode the reply 22 * 23 * Returns: 24 * %0 if @rcvbuf is filled in, xprt_complete_rqst called, 25 * %-EAGAIN if server should call ->recvfrom again. 26 */ 27 int svc_rdma_handle_bc_reply(struct rpc_xprt *xprt, __be32 *rdma_resp, 28 struct xdr_buf *rcvbuf) 29 { 30 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 31 struct kvec *dst, *src = &rcvbuf->head[0]; 32 struct rpc_rqst *req; 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->queue_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 xprt_pin_rqst(req); 67 spin_unlock(&xprt->queue_lock); 68 69 credits = be32_to_cpup(rdma_resp + 2); 70 if (credits == 0) 71 credits = 1; /* don't deadlock */ 72 else if (credits > r_xprt->rx_buf.rb_bc_max_requests) 73 credits = r_xprt->rx_buf.rb_bc_max_requests; 74 75 spin_lock(&xprt->transport_lock); 76 xprt->cwnd = credits << RPC_CWNDSHIFT; 77 spin_unlock(&xprt->transport_lock); 78 79 spin_lock(&xprt->queue_lock); 80 ret = 0; 81 xprt_complete_rqst(req->rq_task, rcvbuf->len); 82 xprt_unpin_rqst(req); 83 rcvbuf->len = 0; 84 85 out_unlock: 86 spin_unlock(&xprt->queue_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 struct svc_rdma_send_ctxt *ctxt) 117 { 118 int ret; 119 120 ret = svc_rdma_map_reply_msg(rdma, ctxt, NULL, &rqst->rq_snd_buf); 121 if (ret < 0) 122 return -EIO; 123 124 /* Bump page refcnt so Send completion doesn't release 125 * the rq_buffer before all retransmits are complete. 126 */ 127 get_page(virt_to_page(rqst->rq_buffer)); 128 ctxt->sc_send_wr.opcode = IB_WR_SEND; 129 return svc_rdma_send(rdma, &ctxt->sc_send_wr); 130 } 131 132 /* Server-side transport endpoint wants a whole page for its send 133 * buffer. The client RPC code constructs the RPC header in this 134 * buffer before it invokes ->send_request. 135 */ 136 static int 137 xprt_rdma_bc_allocate(struct rpc_task *task) 138 { 139 struct rpc_rqst *rqst = task->tk_rqstp; 140 size_t size = rqst->rq_callsize; 141 struct page *page; 142 143 if (size > PAGE_SIZE) { 144 WARN_ONCE(1, "svcrdma: large bc buffer request (size %zu)\n", 145 size); 146 return -EINVAL; 147 } 148 149 page = alloc_page(RPCRDMA_DEF_GFP); 150 if (!page) 151 return -ENOMEM; 152 rqst->rq_buffer = page_address(page); 153 154 rqst->rq_rbuffer = kmalloc(rqst->rq_rcvsize, RPCRDMA_DEF_GFP); 155 if (!rqst->rq_rbuffer) { 156 put_page(page); 157 return -ENOMEM; 158 } 159 return 0; 160 } 161 162 static void 163 xprt_rdma_bc_free(struct rpc_task *task) 164 { 165 struct rpc_rqst *rqst = task->tk_rqstp; 166 167 put_page(virt_to_page(rqst->rq_buffer)); 168 kfree(rqst->rq_rbuffer); 169 } 170 171 static int 172 rpcrdma_bc_send_request(struct svcxprt_rdma *rdma, struct rpc_rqst *rqst) 173 { 174 struct rpc_xprt *xprt = rqst->rq_xprt; 175 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 176 struct svc_rdma_send_ctxt *ctxt; 177 __be32 *p; 178 int rc; 179 180 ctxt = svc_rdma_send_ctxt_get(rdma); 181 if (!ctxt) 182 goto drop_connection; 183 184 p = xdr_reserve_space(&ctxt->sc_stream, RPCRDMA_HDRLEN_MIN); 185 if (!p) 186 goto put_ctxt; 187 *p++ = rqst->rq_xid; 188 *p++ = rpcrdma_version; 189 *p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_max_requests); 190 *p++ = rdma_msg; 191 *p++ = xdr_zero; 192 *p++ = xdr_zero; 193 *p = xdr_zero; 194 195 #ifdef SVCRDMA_BACKCHANNEL_DEBUG 196 pr_info("%s: %*ph\n", __func__, 64, rqst->rq_buffer); 197 #endif 198 199 rqst->rq_xtime = ktime_get(); 200 rc = svc_rdma_bc_sendto(rdma, rqst, ctxt); 201 if (rc) 202 goto put_ctxt; 203 return 0; 204 205 put_ctxt: 206 svc_rdma_send_ctxt_put(rdma, ctxt); 207 208 drop_connection: 209 dprintk("svcrdma: failed to send bc call\n"); 210 return -ENOTCONN; 211 } 212 213 /* Send an RPC call on the passive end of a transport 214 * connection. 215 */ 216 static int 217 xprt_rdma_bc_send_request(struct rpc_rqst *rqst) 218 { 219 struct svc_xprt *sxprt = rqst->rq_xprt->bc_xprt; 220 struct svcxprt_rdma *rdma; 221 int ret; 222 223 dprintk("svcrdma: sending bc call with xid: %08x\n", 224 be32_to_cpu(rqst->rq_xid)); 225 226 mutex_lock(&sxprt->xpt_mutex); 227 228 ret = -ENOTCONN; 229 rdma = container_of(sxprt, struct svcxprt_rdma, sc_xprt); 230 if (!test_bit(XPT_DEAD, &sxprt->xpt_flags)) { 231 ret = rpcrdma_bc_send_request(rdma, rqst); 232 if (ret == -ENOTCONN) 233 svc_close_xprt(sxprt); 234 } 235 236 mutex_unlock(&sxprt->xpt_mutex); 237 238 if (ret < 0) 239 return ret; 240 return 0; 241 } 242 243 static void 244 xprt_rdma_bc_close(struct rpc_xprt *xprt) 245 { 246 dprintk("svcrdma: %s: xprt %p\n", __func__, xprt); 247 248 xprt_disconnect_done(xprt); 249 xprt->cwnd = RPC_CWNDSHIFT; 250 } 251 252 static void 253 xprt_rdma_bc_put(struct rpc_xprt *xprt) 254 { 255 dprintk("svcrdma: %s: xprt %p\n", __func__, xprt); 256 257 xprt_rdma_free_addresses(xprt); 258 xprt_free(xprt); 259 } 260 261 static const struct rpc_xprt_ops xprt_rdma_bc_procs = { 262 .reserve_xprt = xprt_reserve_xprt_cong, 263 .release_xprt = xprt_release_xprt_cong, 264 .alloc_slot = xprt_alloc_slot, 265 .free_slot = xprt_free_slot, 266 .release_request = xprt_release_rqst_cong, 267 .buf_alloc = xprt_rdma_bc_allocate, 268 .buf_free = xprt_rdma_bc_free, 269 .send_request = xprt_rdma_bc_send_request, 270 .wait_for_reply_request = xprt_wait_for_reply_request_def, 271 .close = xprt_rdma_bc_close, 272 .destroy = xprt_rdma_bc_put, 273 .print_stats = xprt_rdma_print_stats 274 }; 275 276 static const struct rpc_timeout xprt_rdma_bc_timeout = { 277 .to_initval = 60 * HZ, 278 .to_maxval = 60 * HZ, 279 }; 280 281 /* It shouldn't matter if the number of backchannel session slots 282 * doesn't match the number of RPC/RDMA credits. That just means 283 * one or the other will have extra slots that aren't used. 284 */ 285 static struct rpc_xprt * 286 xprt_setup_rdma_bc(struct xprt_create *args) 287 { 288 struct rpc_xprt *xprt; 289 struct rpcrdma_xprt *new_xprt; 290 291 if (args->addrlen > sizeof(xprt->addr)) { 292 dprintk("RPC: %s: address too large\n", __func__); 293 return ERR_PTR(-EBADF); 294 } 295 296 xprt = xprt_alloc(args->net, sizeof(*new_xprt), 297 RPCRDMA_MAX_BC_REQUESTS, 298 RPCRDMA_MAX_BC_REQUESTS); 299 if (!xprt) { 300 dprintk("RPC: %s: couldn't allocate rpc_xprt\n", 301 __func__); 302 return ERR_PTR(-ENOMEM); 303 } 304 305 xprt->timeout = &xprt_rdma_bc_timeout; 306 xprt_set_bound(xprt); 307 xprt_set_connected(xprt); 308 xprt->bind_timeout = RPCRDMA_BIND_TO; 309 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO; 310 xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO; 311 312 xprt->prot = XPRT_TRANSPORT_BC_RDMA; 313 xprt->ops = &xprt_rdma_bc_procs; 314 315 memcpy(&xprt->addr, args->dstaddr, args->addrlen); 316 xprt->addrlen = args->addrlen; 317 xprt_rdma_format_addresses(xprt, (struct sockaddr *)&xprt->addr); 318 xprt->resvport = 0; 319 320 xprt->max_payload = xprt_rdma_max_inline_read; 321 322 new_xprt = rpcx_to_rdmax(xprt); 323 new_xprt->rx_buf.rb_bc_max_requests = xprt->max_reqs; 324 325 xprt_get(xprt); 326 args->bc_xprt->xpt_bc_xprt = xprt; 327 xprt->bc_xprt = args->bc_xprt; 328 329 /* Final put for backchannel xprt is in __svc_rdma_free */ 330 xprt_get(xprt); 331 return xprt; 332 } 333 334 struct xprt_class xprt_rdma_bc = { 335 .list = LIST_HEAD_INIT(xprt_rdma_bc.list), 336 .name = "rdma backchannel", 337 .owner = THIS_MODULE, 338 .ident = XPRT_TRANSPORT_BC_RDMA, 339 .setup = xprt_setup_rdma_bc, 340 }; 341