xref: /linux/net/sunrpc/xprtrdma/backchannel.c (revision b2128290c29902315e632ea59e0504d6bc9e9b42)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2015-2020, Oracle and/or its affiliates.
4  *
5  * Support for reverse-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 #include <linux/sunrpc/bc_xprt.h>
13 
14 #include "xprt_rdma.h"
15 #include <trace/events/rpcrdma.h>
16 
17 #undef RPCRDMA_BACKCHANNEL_DEBUG
18 
19 /**
20  * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests
21  * @xprt: transport associated with these backchannel resources
22  * @reqs: number of concurrent incoming requests to expect
23  *
24  * Returns 0 on success; otherwise a negative errno
25  */
26 int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs)
27 {
28 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
29 
30 	r_xprt->rx_buf.rb_bc_srv_max_requests = RPCRDMA_BACKWARD_WRS >> 1;
31 	trace_xprtrdma_cb_setup(r_xprt, reqs);
32 	return 0;
33 }
34 
35 /**
36  * xprt_rdma_bc_maxpayload - Return maximum backchannel message size
37  * @xprt: transport
38  *
39  * Returns maximum size, in bytes, of a backchannel message
40  */
41 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *xprt)
42 {
43 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
44 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
45 	size_t maxmsg;
46 
47 	maxmsg = min_t(unsigned int, ep->re_inline_send, ep->re_inline_recv);
48 	maxmsg = min_t(unsigned int, maxmsg, PAGE_SIZE);
49 	return maxmsg - RPCRDMA_HDRLEN_MIN;
50 }
51 
52 unsigned int xprt_rdma_bc_max_slots(struct rpc_xprt *xprt)
53 {
54 	return RPCRDMA_BACKWARD_WRS >> 1;
55 }
56 
57 static int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst)
58 {
59 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
60 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
61 	__be32 *p;
62 
63 	rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
64 	xdr_init_encode(&req->rl_stream, &req->rl_hdrbuf,
65 			rdmab_data(req->rl_rdmabuf), rqst);
66 
67 	p = xdr_reserve_space(&req->rl_stream, 28);
68 	if (unlikely(!p))
69 		return -EIO;
70 	*p++ = rqst->rq_xid;
71 	*p++ = rpcrdma_version;
72 	*p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests);
73 	*p++ = rdma_msg;
74 	*p++ = xdr_zero;
75 	*p++ = xdr_zero;
76 	*p = xdr_zero;
77 
78 	if (rpcrdma_prepare_send_sges(r_xprt, req, RPCRDMA_HDRLEN_MIN,
79 				      &rqst->rq_snd_buf, rpcrdma_noch_pullup))
80 		return -EIO;
81 
82 	trace_xprtrdma_cb_reply(r_xprt, rqst);
83 	return 0;
84 }
85 
86 /**
87  * xprt_rdma_bc_send_reply - marshal and send a backchannel reply
88  * @rqst: RPC rqst with a backchannel RPC reply in rq_snd_buf
89  *
90  * Caller holds the transport's write lock.
91  *
92  * Returns:
93  *	%0 if the RPC message has been sent
94  *	%-ENOTCONN if the caller should reconnect and call again
95  *	%-EIO if a permanent error occurred and the request was not
96  *		sent. Do not try to send this message again.
97  */
98 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst)
99 {
100 	struct rpc_xprt *xprt = rqst->rq_xprt;
101 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
102 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
103 	int rc;
104 
105 	if (!xprt_connected(xprt))
106 		return -ENOTCONN;
107 
108 	if (!xprt_request_get_cong(xprt, rqst))
109 		return -EBADSLT;
110 
111 	rc = rpcrdma_bc_marshal_reply(rqst);
112 	if (rc < 0)
113 		goto failed_marshal;
114 
115 	if (frwr_send(r_xprt, req))
116 		goto drop_connection;
117 	return 0;
118 
119 failed_marshal:
120 	if (rc != -ENOTCONN)
121 		return rc;
122 drop_connection:
123 	xprt_rdma_close(xprt);
124 	return -ENOTCONN;
125 }
126 
127 /**
128  * xprt_rdma_bc_destroy - Release resources for handling backchannel requests
129  * @xprt: transport associated with these backchannel resources
130  * @reqs: number of incoming requests to destroy; ignored
131  */
132 void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs)
133 {
134 	struct rpc_rqst *rqst, *tmp;
135 
136 	spin_lock(&xprt->bc_pa_lock);
137 	list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) {
138 		list_del(&rqst->rq_bc_pa_list);
139 		spin_unlock(&xprt->bc_pa_lock);
140 
141 		rpcrdma_req_destroy(rpcr_to_rdmar(rqst));
142 
143 		spin_lock(&xprt->bc_pa_lock);
144 	}
145 	spin_unlock(&xprt->bc_pa_lock);
146 }
147 
148 /**
149  * xprt_rdma_bc_free_rqst - Release a backchannel rqst
150  * @rqst: request to release
151  */
152 void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst)
153 {
154 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
155 	struct rpcrdma_rep *rep = req->rl_reply;
156 	struct rpc_xprt *xprt = rqst->rq_xprt;
157 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
158 
159 	rpcrdma_rep_put(&r_xprt->rx_buf, rep);
160 	req->rl_reply = NULL;
161 
162 	rpcrdma_req_put(req);
163 	xprt_put(xprt);
164 }
165 
166 static struct rpc_rqst *rpcrdma_bc_rqst_get(struct rpcrdma_xprt *r_xprt)
167 {
168 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
169 	struct rpcrdma_req *req;
170 	struct rpc_rqst *rqst;
171 	size_t size;
172 
173 	spin_lock(&xprt->bc_pa_lock);
174 	rqst = list_first_entry_or_null(&xprt->bc_pa_list, struct rpc_rqst,
175 					rq_bc_pa_list);
176 	if (!rqst)
177 		goto create_req;
178 	list_del(&rqst->rq_bc_pa_list);
179 	spin_unlock(&xprt->bc_pa_lock);
180 	return rqst;
181 
182 create_req:
183 	spin_unlock(&xprt->bc_pa_lock);
184 
185 	/* Set a limit to prevent a remote from overrunning our resources.
186 	 */
187 	if (xprt->bc_alloc_count >= RPCRDMA_BACKWARD_WRS)
188 		return NULL;
189 
190 	size = min_t(size_t, r_xprt->rx_ep->re_inline_recv, PAGE_SIZE);
191 	req = rpcrdma_req_create(r_xprt, size);
192 	if (!req)
193 		return NULL;
194 	if (rpcrdma_req_setup(r_xprt, req)) {
195 		rpcrdma_req_destroy(req);
196 		return NULL;
197 	}
198 
199 	xprt->bc_alloc_count++;
200 	rqst = &req->rl_slot;
201 	rqst->rq_xprt = xprt;
202 	__set_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state);
203 	xdr_buf_init(&rqst->rq_snd_buf, rdmab_data(req->rl_sendbuf), size);
204 	kref_init(&req->rl_kref);
205 	return rqst;
206 }
207 
208 /**
209  * rpcrdma_bc_receive_call - Handle a reverse-direction Call
210  * @r_xprt: transport receiving the call
211  * @rep: receive buffer containing the call
212  *
213  * Operational assumptions:
214  *    o Backchannel credits are ignored, just as the NFS server
215  *      forechannel currently does
216  *    o The ULP manages a replay cache (eg, NFSv4.1 sessions).
217  *      No replay detection is done at the transport level
218  */
219 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt,
220 			     struct rpcrdma_rep *rep)
221 {
222 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
223 	struct rpcrdma_req *req;
224 	struct rpc_rqst *rqst;
225 	struct xdr_buf *buf;
226 	size_t size;
227 	__be32 *p;
228 
229 	p = xdr_inline_decode(&rep->rr_stream, 0);
230 	size = xdr_stream_remaining(&rep->rr_stream);
231 
232 #ifdef RPCRDMA_BACKCHANNEL_DEBUG
233 	pr_info("RPC:       %s: callback XID %08x, length=%u\n",
234 		__func__, be32_to_cpup(p), size);
235 	pr_info("RPC:       %s: %*ph\n", __func__, size, p);
236 #endif
237 
238 	rqst = rpcrdma_bc_rqst_get(r_xprt);
239 	if (!rqst)
240 		goto out_overflow;
241 
242 	rqst->rq_reply_bytes_recvd = 0;
243 	rqst->rq_xid = *p;
244 
245 	rqst->rq_private_buf.len = size;
246 
247 	buf = &rqst->rq_rcv_buf;
248 	memset(buf, 0, sizeof(*buf));
249 	buf->head[0].iov_base = p;
250 	buf->head[0].iov_len = size;
251 	buf->len = size;
252 
253 	/* The receive buffer has to be hooked to the rpcrdma_req
254 	 * so that it is not released while the req is pointing
255 	 * to its buffer, and so that it can be reposted after
256 	 * the Upper Layer is done decoding it.
257 	 */
258 	req = rpcr_to_rdmar(rqst);
259 	req->rl_reply = rep;
260 	trace_xprtrdma_cb_call(r_xprt, rqst);
261 
262 	/* Queue rqst for ULP's callback service */
263 	xprt_enqueue_bc_request(rqst);
264 
265 	r_xprt->rx_stats.bcall_count++;
266 	return;
267 
268 out_overflow:
269 	pr_warn("RPC/RDMA backchannel overflow\n");
270 	xprt_force_disconnect(xprt);
271 	/* This receive buffer gets reposted automatically
272 	 * when the connection is re-established.
273 	 */
274 	return;
275 }
276