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.
6  */
7 
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>
12 
13 #include "xprt_rdma.h"
14 #include <trace/events/rpcrdma.h>
15 
16 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
17 # define RPCDBG_FACILITY	RPCDBG_TRANS
18 #endif
19 
20 #undef RPCRDMA_BACKCHANNEL_DEBUG
21 
22 static int rpcrdma_bc_setup_reqs(struct rpcrdma_xprt *r_xprt,
23 				 unsigned int count)
24 {
25 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
26 	struct rpcrdma_req *req;
27 	struct rpc_rqst *rqst;
28 	unsigned int i;
29 
30 	for (i = 0; i < (count << 1); i++) {
31 		struct rpcrdma_regbuf *rb;
32 		size_t size;
33 
34 		req = rpcrdma_create_req(r_xprt);
35 		if (IS_ERR(req))
36 			return PTR_ERR(req);
37 		rqst = &req->rl_slot;
38 
39 		rqst->rq_xprt = 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);
45 
46 		size = r_xprt->rx_data.inline_rsize;
47 		rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, GFP_KERNEL);
48 		if (IS_ERR(rb))
49 			goto out_fail;
50 		req->rl_sendbuf = rb;
51 		xdr_buf_init(&rqst->rq_snd_buf, rb->rg_base,
52 			     min_t(size_t, size, PAGE_SIZE));
53 	}
54 	return 0;
55 
56 out_fail:
57 	rpcrdma_req_destroy(req);
58 	return -ENOMEM;
59 }
60 
61 /**
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
65  *
66  * Returns 0 on success; otherwise a negative errno
67  */
68 int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs)
69 {
70 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
71 	int rc;
72 
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.
78 	 *
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.
81 	 */
82 	if (reqs > RPCRDMA_BACKWARD_WRS >> 1)
83 		goto out_err;
84 
85 	rc = rpcrdma_bc_setup_reqs(r_xprt, reqs);
86 	if (rc)
87 		goto out_free;
88 
89 	r_xprt->rx_buf.rb_bc_srv_max_requests = reqs;
90 	trace_xprtrdma_cb_setup(r_xprt, reqs);
91 	return 0;
92 
93 out_free:
94 	xprt_rdma_bc_destroy(xprt, reqs);
95 
96 out_err:
97 	pr_err("RPC:       %s: setup backchannel transport failed\n", __func__);
98 	return -ENOMEM;
99 }
100 
101 /**
102  * xprt_rdma_bc_maxpayload - Return maximum backchannel message size
103  * @xprt: transport
104  *
105  * Returns maximum size, in bytes, of a backchannel message
106  */
107 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *xprt)
108 {
109 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
110 	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
111 	size_t maxmsg;
112 
113 	maxmsg = min_t(unsigned int, cdata->inline_rsize, cdata->inline_wsize);
114 	maxmsg = min_t(unsigned int, maxmsg, PAGE_SIZE);
115 	return maxmsg - RPCRDMA_HDRLEN_MIN;
116 }
117 
118 static int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst)
119 {
120 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
121 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
122 	__be32 *p;
123 
124 	rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
125 	xdr_init_encode(&req->rl_stream, &req->rl_hdrbuf,
126 			req->rl_rdmabuf->rg_base, rqst);
127 
128 	p = xdr_reserve_space(&req->rl_stream, 28);
129 	if (unlikely(!p))
130 		return -EIO;
131 	*p++ = rqst->rq_xid;
132 	*p++ = rpcrdma_version;
133 	*p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests);
134 	*p++ = rdma_msg;
135 	*p++ = xdr_zero;
136 	*p++ = xdr_zero;
137 	*p = xdr_zero;
138 
139 	if (rpcrdma_prepare_send_sges(r_xprt, req, RPCRDMA_HDRLEN_MIN,
140 				      &rqst->rq_snd_buf, rpcrdma_noch))
141 		return -EIO;
142 
143 	trace_xprtrdma_cb_reply(rqst);
144 	return 0;
145 }
146 
147 /**
148  * xprt_rdma_bc_send_reply - marshal and send a backchannel reply
149  * @rqst: RPC rqst with a backchannel RPC reply in rq_snd_buf
150  *
151  * Caller holds the transport's write lock.
152  *
153  * Returns:
154  *	%0 if the RPC message has been sent
155  *	%-ENOTCONN if the caller should reconnect and call again
156  *	%-EIO if a permanent error occurred and the request was not
157  *		sent. Do not try to send this message again.
158  */
159 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst)
160 {
161 	struct rpc_xprt *xprt = rqst->rq_xprt;
162 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
163 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
164 	int rc;
165 
166 	if (!xprt_connected(xprt))
167 		return -ENOTCONN;
168 
169 	if (!xprt_request_get_cong(xprt, rqst))
170 		return -EBADSLT;
171 
172 	rc = rpcrdma_bc_marshal_reply(rqst);
173 	if (rc < 0)
174 		goto failed_marshal;
175 
176 	if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
177 		goto drop_connection;
178 	return 0;
179 
180 failed_marshal:
181 	if (rc != -ENOTCONN)
182 		return rc;
183 drop_connection:
184 	xprt_rdma_close(xprt);
185 	return -ENOTCONN;
186 }
187 
188 /**
189  * xprt_rdma_bc_destroy - Release resources for handling backchannel requests
190  * @xprt: transport associated with these backchannel resources
191  * @reqs: number of incoming requests to destroy; ignored
192  */
193 void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs)
194 {
195 	struct rpc_rqst *rqst, *tmp;
196 
197 	spin_lock(&xprt->bc_pa_lock);
198 	list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) {
199 		list_del(&rqst->rq_bc_pa_list);
200 		spin_unlock(&xprt->bc_pa_lock);
201 
202 		rpcrdma_req_destroy(rpcr_to_rdmar(rqst));
203 
204 		spin_lock(&xprt->bc_pa_lock);
205 	}
206 	spin_unlock(&xprt->bc_pa_lock);
207 }
208 
209 /**
210  * xprt_rdma_bc_free_rqst - Release a backchannel rqst
211  * @rqst: request to release
212  */
213 void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst)
214 {
215 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
216 	struct rpc_xprt *xprt = rqst->rq_xprt;
217 
218 	rpcrdma_recv_buffer_put(req->rl_reply);
219 	req->rl_reply = NULL;
220 
221 	spin_lock(&xprt->bc_pa_lock);
222 	list_add_tail(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
223 	spin_unlock(&xprt->bc_pa_lock);
224 }
225 
226 /**
227  * rpcrdma_bc_receive_call - Handle a backward direction call
228  * @r_xprt: transport receiving the call
229  * @rep: receive buffer containing the call
230  *
231  * Operational assumptions:
232  *    o Backchannel credits are ignored, just as the NFS server
233  *      forechannel currently does
234  *    o The ULP manages a replay cache (eg, NFSv4.1 sessions).
235  *      No replay detection is done at the transport level
236  */
237 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt,
238 			     struct rpcrdma_rep *rep)
239 {
240 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
241 	struct svc_serv *bc_serv;
242 	struct rpcrdma_req *req;
243 	struct rpc_rqst *rqst;
244 	struct xdr_buf *buf;
245 	size_t size;
246 	__be32 *p;
247 
248 	p = xdr_inline_decode(&rep->rr_stream, 0);
249 	size = xdr_stream_remaining(&rep->rr_stream);
250 
251 #ifdef RPCRDMA_BACKCHANNEL_DEBUG
252 	pr_info("RPC:       %s: callback XID %08x, length=%u\n",
253 		__func__, be32_to_cpup(p), size);
254 	pr_info("RPC:       %s: %*ph\n", __func__, size, p);
255 #endif
256 
257 	/* Grab a free bc rqst */
258 	spin_lock(&xprt->bc_pa_lock);
259 	if (list_empty(&xprt->bc_pa_list)) {
260 		spin_unlock(&xprt->bc_pa_lock);
261 		goto out_overflow;
262 	}
263 	rqst = list_first_entry(&xprt->bc_pa_list,
264 				struct rpc_rqst, rq_bc_pa_list);
265 	list_del(&rqst->rq_bc_pa_list);
266 	spin_unlock(&xprt->bc_pa_lock);
267 
268 	/* Prepare rqst */
269 	rqst->rq_reply_bytes_recvd = 0;
270 	rqst->rq_xid = *p;
271 
272 	rqst->rq_private_buf.len = size;
273 
274 	buf = &rqst->rq_rcv_buf;
275 	memset(buf, 0, sizeof(*buf));
276 	buf->head[0].iov_base = p;
277 	buf->head[0].iov_len = size;
278 	buf->len = size;
279 
280 	/* The receive buffer has to be hooked to the rpcrdma_req
281 	 * so that it is not released while the req is pointing
282 	 * to its buffer, and so that it can be reposted after
283 	 * the Upper Layer is done decoding it.
284 	 */
285 	req = rpcr_to_rdmar(rqst);
286 	req->rl_reply = rep;
287 	trace_xprtrdma_cb_call(rqst);
288 
289 	/* Queue rqst for ULP's callback service */
290 	bc_serv = xprt->bc_serv;
291 	spin_lock(&bc_serv->sv_cb_lock);
292 	list_add(&rqst->rq_bc_list, &bc_serv->sv_cb_list);
293 	spin_unlock(&bc_serv->sv_cb_lock);
294 
295 	wake_up(&bc_serv->sv_cb_waitq);
296 
297 	r_xprt->rx_stats.bcall_count++;
298 	return;
299 
300 out_overflow:
301 	pr_warn("RPC/RDMA backchannel overflow\n");
302 	xprt_force_disconnect(xprt);
303 	/* This receive buffer gets reposted automatically
304 	 * when the connection is re-established.
305 	 */
306 	return;
307 }
308