1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (c) 2015-2018 Oracle. All rights reserved. 4 * 5 * Support for reverse-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 /** 14 * svc_rdma_handle_bc_reply - Process incoming backchannel Reply 15 * @rqstp: resources for handling the Reply 16 * @rctxt: Received message 17 * 18 */ 19 void svc_rdma_handle_bc_reply(struct svc_rqst *rqstp, 20 struct svc_rdma_recv_ctxt *rctxt) 21 { 22 struct svc_xprt *sxprt = rqstp->rq_xprt; 23 struct rpc_xprt *xprt = sxprt->xpt_bc_xprt; 24 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 25 struct xdr_buf *rcvbuf = &rqstp->rq_arg; 26 struct kvec *dst, *src = &rcvbuf->head[0]; 27 __be32 *rdma_resp = rctxt->rc_recv_buf; 28 struct rpc_rqst *req; 29 u32 credits; 30 31 spin_lock(&xprt->queue_lock); 32 req = xprt_lookup_rqst(xprt, *rdma_resp); 33 if (!req) 34 goto out_unlock; 35 36 dst = &req->rq_private_buf.head[0]; 37 memcpy(&req->rq_private_buf, &req->rq_rcv_buf, sizeof(struct xdr_buf)); 38 if (dst->iov_len < src->iov_len) 39 goto out_unlock; 40 memcpy(dst->iov_base, src->iov_base, src->iov_len); 41 xprt_pin_rqst(req); 42 spin_unlock(&xprt->queue_lock); 43 44 credits = be32_to_cpup(rdma_resp + 2); 45 if (credits == 0) 46 credits = 1; /* don't deadlock */ 47 else if (credits > r_xprt->rx_buf.rb_bc_max_requests) 48 credits = r_xprt->rx_buf.rb_bc_max_requests; 49 spin_lock(&xprt->transport_lock); 50 xprt->cwnd = credits << RPC_CWNDSHIFT; 51 spin_unlock(&xprt->transport_lock); 52 53 spin_lock(&xprt->queue_lock); 54 xprt_complete_rqst(req->rq_task, rcvbuf->len); 55 xprt_unpin_rqst(req); 56 rcvbuf->len = 0; 57 58 out_unlock: 59 spin_unlock(&xprt->queue_lock); 60 } 61 62 /* Send a reverse-direction RPC Call. 63 * 64 * Caller holds the connection's mutex and has already marshaled 65 * the RPC/RDMA request. 66 * 67 * This is similar to svc_rdma_send_reply_msg, but takes a struct 68 * rpc_rqst instead, does not support chunks, and avoids blocking 69 * memory allocation. 70 * 71 * XXX: There is still an opportunity to block in svc_rdma_send() 72 * if there are no SQ entries to post the Send. This may occur if 73 * the adapter has a small maximum SQ depth. 74 */ 75 static int svc_rdma_bc_sendto(struct svcxprt_rdma *rdma, 76 struct rpc_rqst *rqst, 77 struct svc_rdma_send_ctxt *sctxt) 78 { 79 struct svc_rdma_pcl empty_pcl; 80 int ret; 81 82 pcl_init(&empty_pcl); 83 ret = svc_rdma_map_reply_msg(rdma, sctxt, &empty_pcl, &empty_pcl, 84 &rqst->rq_snd_buf); 85 if (ret < 0) 86 return -EIO; 87 88 /* Bump page refcnt so Send completion doesn't release 89 * the rq_buffer before all retransmits are complete. 90 */ 91 get_page(virt_to_page(rqst->rq_buffer)); 92 sctxt->sc_send_wr.opcode = IB_WR_SEND; 93 return svc_rdma_send(rdma, sctxt); 94 } 95 96 /* Server-side transport endpoint wants a whole page for its send 97 * buffer. The client RPC code constructs the RPC header in this 98 * buffer before it invokes ->send_request. 99 */ 100 static int 101 xprt_rdma_bc_allocate(struct rpc_task *task) 102 { 103 struct rpc_rqst *rqst = task->tk_rqstp; 104 size_t size = rqst->rq_callsize; 105 struct page *page; 106 107 if (size > PAGE_SIZE) { 108 WARN_ONCE(1, "svcrdma: large bc buffer request (size %zu)\n", 109 size); 110 return -EINVAL; 111 } 112 113 page = alloc_page(GFP_NOIO | __GFP_NOWARN); 114 if (!page) 115 return -ENOMEM; 116 rqst->rq_buffer = page_address(page); 117 118 rqst->rq_rbuffer = kmalloc(rqst->rq_rcvsize, GFP_NOIO | __GFP_NOWARN); 119 if (!rqst->rq_rbuffer) { 120 put_page(page); 121 return -ENOMEM; 122 } 123 return 0; 124 } 125 126 static void 127 xprt_rdma_bc_free(struct rpc_task *task) 128 { 129 struct rpc_rqst *rqst = task->tk_rqstp; 130 131 put_page(virt_to_page(rqst->rq_buffer)); 132 kfree(rqst->rq_rbuffer); 133 } 134 135 static int 136 rpcrdma_bc_send_request(struct svcxprt_rdma *rdma, struct rpc_rqst *rqst) 137 { 138 struct rpc_xprt *xprt = rqst->rq_xprt; 139 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 140 struct svc_rdma_send_ctxt *ctxt; 141 __be32 *p; 142 int rc; 143 144 ctxt = svc_rdma_send_ctxt_get(rdma); 145 if (!ctxt) 146 goto drop_connection; 147 148 p = xdr_reserve_space(&ctxt->sc_stream, RPCRDMA_HDRLEN_MIN); 149 if (!p) 150 goto put_ctxt; 151 *p++ = rqst->rq_xid; 152 *p++ = rpcrdma_version; 153 *p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_max_requests); 154 *p++ = rdma_msg; 155 *p++ = xdr_zero; 156 *p++ = xdr_zero; 157 *p = xdr_zero; 158 159 rqst->rq_xtime = ktime_get(); 160 rc = svc_rdma_bc_sendto(rdma, rqst, ctxt); 161 if (rc) 162 goto put_ctxt; 163 return 0; 164 165 put_ctxt: 166 svc_rdma_send_ctxt_put(rdma, ctxt); 167 168 drop_connection: 169 return -ENOTCONN; 170 } 171 172 /** 173 * xprt_rdma_bc_send_request - Send a reverse-direction Call 174 * @rqst: rpc_rqst containing Call message to be sent 175 * 176 * Return values: 177 * %0 if the message was sent successfully 178 * %ENOTCONN if the message was not sent 179 */ 180 static int xprt_rdma_bc_send_request(struct rpc_rqst *rqst) 181 { 182 struct svc_xprt *sxprt = rqst->rq_xprt->bc_xprt; 183 struct svcxprt_rdma *rdma = 184 container_of(sxprt, struct svcxprt_rdma, sc_xprt); 185 int ret; 186 187 if (test_bit(XPT_DEAD, &sxprt->xpt_flags)) 188 return -ENOTCONN; 189 190 ret = rpcrdma_bc_send_request(rdma, rqst); 191 if (ret == -ENOTCONN) 192 svc_xprt_close(sxprt); 193 return ret; 194 } 195 196 static void 197 xprt_rdma_bc_close(struct rpc_xprt *xprt) 198 { 199 xprt_disconnect_done(xprt); 200 xprt->cwnd = RPC_CWNDSHIFT; 201 } 202 203 static void 204 xprt_rdma_bc_put(struct rpc_xprt *xprt) 205 { 206 xprt_rdma_free_addresses(xprt); 207 xprt_free(xprt); 208 } 209 210 static const struct rpc_xprt_ops xprt_rdma_bc_procs = { 211 .reserve_xprt = xprt_reserve_xprt_cong, 212 .release_xprt = xprt_release_xprt_cong, 213 .alloc_slot = xprt_alloc_slot, 214 .free_slot = xprt_free_slot, 215 .release_request = xprt_release_rqst_cong, 216 .buf_alloc = xprt_rdma_bc_allocate, 217 .buf_free = xprt_rdma_bc_free, 218 .send_request = xprt_rdma_bc_send_request, 219 .wait_for_reply_request = xprt_wait_for_reply_request_def, 220 .close = xprt_rdma_bc_close, 221 .destroy = xprt_rdma_bc_put, 222 .print_stats = xprt_rdma_print_stats 223 }; 224 225 static const struct rpc_timeout xprt_rdma_bc_timeout = { 226 .to_initval = 60 * HZ, 227 .to_maxval = 60 * HZ, 228 }; 229 230 /* It shouldn't matter if the number of backchannel session slots 231 * doesn't match the number of RPC/RDMA credits. That just means 232 * one or the other will have extra slots that aren't used. 233 */ 234 static struct rpc_xprt * 235 xprt_setup_rdma_bc(struct xprt_create *args) 236 { 237 struct rpc_xprt *xprt; 238 struct rpcrdma_xprt *new_xprt; 239 240 if (args->addrlen > sizeof(xprt->addr)) 241 return ERR_PTR(-EBADF); 242 243 xprt = xprt_alloc(args->net, sizeof(*new_xprt), 244 RPCRDMA_MAX_BC_REQUESTS, 245 RPCRDMA_MAX_BC_REQUESTS); 246 if (!xprt) 247 return ERR_PTR(-ENOMEM); 248 249 xprt->timeout = &xprt_rdma_bc_timeout; 250 xprt_set_bound(xprt); 251 xprt_set_connected(xprt); 252 xprt->bind_timeout = 0; 253 xprt->reestablish_timeout = 0; 254 xprt->idle_timeout = 0; 255 256 xprt->prot = XPRT_TRANSPORT_BC_RDMA; 257 xprt->ops = &xprt_rdma_bc_procs; 258 259 memcpy(&xprt->addr, args->dstaddr, args->addrlen); 260 xprt->addrlen = args->addrlen; 261 xprt_rdma_format_addresses(xprt, (struct sockaddr *)&xprt->addr); 262 xprt->resvport = 0; 263 264 xprt->max_payload = xprt_rdma_max_inline_read; 265 266 new_xprt = rpcx_to_rdmax(xprt); 267 new_xprt->rx_buf.rb_bc_max_requests = xprt->max_reqs; 268 269 xprt_get(xprt); 270 args->bc_xprt->xpt_bc_xprt = xprt; 271 xprt->bc_xprt = args->bc_xprt; 272 273 /* Final put for backchannel xprt is in __svc_rdma_free */ 274 xprt_get(xprt); 275 return xprt; 276 } 277 278 struct xprt_class xprt_rdma_bc = { 279 .list = LIST_HEAD_INIT(xprt_rdma_bc.list), 280 .name = "rdma backchannel", 281 .owner = THIS_MODULE, 282 .ident = XPRT_TRANSPORT_BC_RDMA, 283 .setup = xprt_setup_rdma_bc, 284 }; 285