1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2015 Oracle. All rights reserved.
5 * Support for backward direction RPCs on RPC/RDMA.
8 #include <linux/sunrpc/xprt.h>
9 #include <linux/sunrpc/svc.h>
10 #include <linux/sunrpc/svc_xprt.h>
11 #include <linux/sunrpc/svc_rdma.h>
13 #include "xprt_rdma.h"
14 #include <trace/events/rpcrdma.h>
16 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
17 # define RPCDBG_FACILITY RPCDBG_TRANS
20 #undef RPCRDMA_BACKCHANNEL_DEBUG
22 static int rpcrdma_bc_setup_reqs(struct rpcrdma_xprt *r_xprt,
25 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
26 struct rpcrdma_req *req;
27 struct rpc_rqst *rqst;
30 for (i = 0; i < (count << 1); i++) {
31 struct rpcrdma_regbuf *rb;
34 req = rpcrdma_create_req(r_xprt);
40 INIT_LIST_HEAD(&rqst->rq_bc_list);
41 __set_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state);
42 spin_lock(&xprt->bc_pa_lock);
43 list_add(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
44 spin_unlock(&xprt->bc_pa_lock);
46 size = r_xprt->rx_data.inline_rsize;
47 rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, GFP_KERNEL);
51 xdr_buf_init(&rqst->rq_snd_buf, rb->rg_base,
52 min_t(size_t, size, PAGE_SIZE));
57 rpcrdma_req_destroy(req);
62 * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests
63 * @xprt: transport associated with these backchannel resources
64 * @reqs: number of concurrent incoming requests to expect
66 * Returns 0 on success; otherwise a negative errno
68 int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs)
70 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
73 /* The backchannel reply path returns each rpc_rqst to the
74 * bc_pa_list _after_ the reply is sent. If the server is
75 * faster than the client, it can send another backward
76 * direction request before the rpc_rqst is returned to the
77 * list. The client rejects the request in this case.
79 * Twice as many rpc_rqsts are prepared to ensure there is
80 * always an rpc_rqst available as soon as a reply is sent.
82 if (reqs > RPCRDMA_BACKWARD_WRS >> 1)
85 rc = rpcrdma_bc_setup_reqs(r_xprt, reqs);
89 r_xprt->rx_buf.rb_bc_srv_max_requests = reqs;
90 trace_xprtrdma_cb_setup(r_xprt, reqs);
94 xprt_rdma_bc_destroy(xprt, reqs);
97 pr_err("RPC: %s: setup backchannel transport failed\n", __func__);
102 * xprt_rdma_bc_up - Create transport endpoint for backchannel service
103 * @serv: server endpoint
104 * @net: network namespace
106 * The "xprt" is an implied argument: it supplies the name of the
107 * backchannel transport class.
109 * Returns zero on success, negative errno on failure
111 int xprt_rdma_bc_up(struct svc_serv *serv, struct net *net)
115 ret = svc_create_xprt(serv, "rdma-bc", net, PF_INET, 0, 0);
122 * xprt_rdma_bc_maxpayload - Return maximum backchannel message size
125 * Returns maximum size, in bytes, of a backchannel message
127 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *xprt)
129 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
130 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
133 maxmsg = min_t(unsigned int, cdata->inline_rsize, cdata->inline_wsize);
134 maxmsg = min_t(unsigned int, maxmsg, PAGE_SIZE);
135 return maxmsg - RPCRDMA_HDRLEN_MIN;
138 static int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst)
140 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
141 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
144 rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
145 xdr_init_encode(&req->rl_stream, &req->rl_hdrbuf,
146 req->rl_rdmabuf->rg_base);
148 p = xdr_reserve_space(&req->rl_stream, 28);
152 *p++ = rpcrdma_version;
153 *p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests);
159 if (rpcrdma_prepare_send_sges(r_xprt, req, RPCRDMA_HDRLEN_MIN,
160 &rqst->rq_snd_buf, rpcrdma_noch))
163 trace_xprtrdma_cb_reply(rqst);
168 * xprt_rdma_bc_send_reply - marshal and send a backchannel reply
169 * @rqst: RPC rqst with a backchannel RPC reply in rq_snd_buf
171 * Caller holds the transport's write lock.
174 * %0 if the RPC message has been sent
175 * %-ENOTCONN if the caller should reconnect and call again
176 * %-EIO if a permanent error occurred and the request was not
177 * sent. Do not try to send this message again.
179 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst)
181 struct rpc_xprt *xprt = rqst->rq_xprt;
182 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
183 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
186 if (!xprt_connected(xprt))
189 if (!xprt_request_get_cong(xprt, rqst))
192 rc = rpcrdma_bc_marshal_reply(rqst);
196 if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
197 goto drop_connection;
204 xprt_rdma_close(xprt);
209 * xprt_rdma_bc_destroy - Release resources for handling backchannel requests
210 * @xprt: transport associated with these backchannel resources
211 * @reqs: number of incoming requests to destroy; ignored
213 void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs)
215 struct rpc_rqst *rqst, *tmp;
217 spin_lock(&xprt->bc_pa_lock);
218 list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) {
219 list_del(&rqst->rq_bc_pa_list);
220 spin_unlock(&xprt->bc_pa_lock);
222 rpcrdma_req_destroy(rpcr_to_rdmar(rqst));
224 spin_lock(&xprt->bc_pa_lock);
226 spin_unlock(&xprt->bc_pa_lock);
230 * xprt_rdma_bc_free_rqst - Release a backchannel rqst
231 * @rqst: request to release
233 void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst)
235 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
236 struct rpc_xprt *xprt = rqst->rq_xprt;
238 rpcrdma_recv_buffer_put(req->rl_reply);
239 req->rl_reply = NULL;
241 spin_lock(&xprt->bc_pa_lock);
242 list_add_tail(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
243 spin_unlock(&xprt->bc_pa_lock);
247 * rpcrdma_bc_receive_call - Handle a backward direction call
248 * @r_xprt: transport receiving the call
249 * @rep: receive buffer containing the call
251 * Operational assumptions:
252 * o Backchannel credits are ignored, just as the NFS server
253 * forechannel currently does
254 * o The ULP manages a replay cache (eg, NFSv4.1 sessions).
255 * No replay detection is done at the transport level
257 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt,
258 struct rpcrdma_rep *rep)
260 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
261 struct svc_serv *bc_serv;
262 struct rpcrdma_req *req;
263 struct rpc_rqst *rqst;
268 p = xdr_inline_decode(&rep->rr_stream, 0);
269 size = xdr_stream_remaining(&rep->rr_stream);
271 #ifdef RPCRDMA_BACKCHANNEL_DEBUG
272 pr_info("RPC: %s: callback XID %08x, length=%u\n",
273 __func__, be32_to_cpup(p), size);
274 pr_info("RPC: %s: %*ph\n", __func__, size, p);
277 /* Grab a free bc rqst */
278 spin_lock(&xprt->bc_pa_lock);
279 if (list_empty(&xprt->bc_pa_list)) {
280 spin_unlock(&xprt->bc_pa_lock);
283 rqst = list_first_entry(&xprt->bc_pa_list,
284 struct rpc_rqst, rq_bc_pa_list);
285 list_del(&rqst->rq_bc_pa_list);
286 spin_unlock(&xprt->bc_pa_lock);
289 rqst->rq_reply_bytes_recvd = 0;
290 rqst->rq_bytes_sent = 0;
293 rqst->rq_private_buf.len = size;
295 buf = &rqst->rq_rcv_buf;
296 memset(buf, 0, sizeof(*buf));
297 buf->head[0].iov_base = p;
298 buf->head[0].iov_len = size;
301 /* The receive buffer has to be hooked to the rpcrdma_req
302 * so that it is not released while the req is pointing
303 * to its buffer, and so that it can be reposted after
304 * the Upper Layer is done decoding it.
306 req = rpcr_to_rdmar(rqst);
308 trace_xprtrdma_cb_call(rqst);
310 /* Queue rqst for ULP's callback service */
311 bc_serv = xprt->bc_serv;
312 spin_lock(&bc_serv->sv_cb_lock);
313 list_add(&rqst->rq_bc_list, &bc_serv->sv_cb_list);
314 spin_unlock(&bc_serv->sv_cb_lock);
316 wake_up(&bc_serv->sv_cb_waitq);
318 r_xprt->rx_stats.bcall_count++;
322 pr_warn("RPC/RDMA backchannel overflow\n");
323 xprt_force_disconnect(xprt);
324 /* This receive buffer gets reposted automatically
325 * when the connection is re-established.