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
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 /**
23 * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests
24 * @xprt: transport associated with these backchannel resources
25 * @reqs: number of concurrent incoming requests to expect
26 *
27 * Returns 0 on success; otherwise a negative errno
28 */
xprt_rdma_bc_setup(struct rpc_xprt * xprt,unsigned int reqs)29 int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs)
30 {
31 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
32
33 r_xprt->rx_buf.rb_bc_srv_max_requests = RPCRDMA_BACKWARD_WRS >> 1;
34 trace_xprtrdma_cb_setup(r_xprt, reqs);
35 return 0;
36 }
37
38 /**
39 * xprt_rdma_bc_maxpayload - Return maximum backchannel message size
40 * @xprt: transport
41 *
42 * Returns maximum size, in bytes, of a backchannel message
43 */
xprt_rdma_bc_maxpayload(struct rpc_xprt * xprt)44 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *xprt)
45 {
46 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
47 struct rpcrdma_ep *ep = r_xprt->rx_ep;
48 size_t maxmsg;
49
50 maxmsg = min_t(unsigned int, ep->re_inline_send, ep->re_inline_recv);
51 maxmsg = min_t(unsigned int, maxmsg, PAGE_SIZE);
52 return maxmsg - RPCRDMA_HDRLEN_MIN;
53 }
54
xprt_rdma_bc_max_slots(struct rpc_xprt * xprt)55 unsigned int xprt_rdma_bc_max_slots(struct rpc_xprt *xprt)
56 {
57 return RPCRDMA_BACKWARD_WRS >> 1;
58 }
59
rpcrdma_bc_marshal_reply(struct rpc_rqst * rqst)60 static int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst)
61 {
62 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
63 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
64 __be32 *p;
65
66 rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
67 xdr_init_encode(&req->rl_stream, &req->rl_hdrbuf,
68 rdmab_data(req->rl_rdmabuf), rqst);
69
70 p = xdr_reserve_space(&req->rl_stream, 28);
71 if (unlikely(!p))
72 return -EIO;
73 *p++ = rqst->rq_xid;
74 *p++ = rpcrdma_version;
75 *p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests);
76 *p++ = rdma_msg;
77 *p++ = xdr_zero;
78 *p++ = xdr_zero;
79 *p = xdr_zero;
80
81 if (rpcrdma_prepare_send_sges(r_xprt, req, RPCRDMA_HDRLEN_MIN,
82 &rqst->rq_snd_buf, rpcrdma_noch_pullup))
83 return -EIO;
84
85 trace_xprtrdma_cb_reply(r_xprt, rqst);
86 return 0;
87 }
88
89 /**
90 * xprt_rdma_bc_send_reply - marshal and send a backchannel reply
91 * @rqst: RPC rqst with a backchannel RPC reply in rq_snd_buf
92 *
93 * Caller holds the transport's write lock.
94 *
95 * Returns:
96 * %0 if the RPC message has been sent
97 * %-ENOTCONN if the caller should reconnect and call again
98 * %-EIO if a permanent error occurred and the request was not
99 * sent. Do not try to send this message again.
100 */
xprt_rdma_bc_send_reply(struct rpc_rqst * rqst)101 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst)
102 {
103 struct rpc_xprt *xprt = rqst->rq_xprt;
104 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
105 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
106 int rc;
107
108 if (!xprt_connected(xprt))
109 return -ENOTCONN;
110
111 if (!xprt_request_get_cong(xprt, rqst))
112 return -EBADSLT;
113
114 rc = rpcrdma_bc_marshal_reply(rqst);
115 if (rc < 0)
116 goto failed_marshal;
117
118 if (frwr_send(r_xprt, req))
119 goto drop_connection;
120 return 0;
121
122 failed_marshal:
123 if (rc != -ENOTCONN)
124 return rc;
125 drop_connection:
126 xprt_rdma_close(xprt);
127 return -ENOTCONN;
128 }
129
130 /**
131 * xprt_rdma_bc_destroy - Release resources for handling backchannel requests
132 * @xprt: transport associated with these backchannel resources
133 * @reqs: number of incoming requests to destroy; ignored
134 */
xprt_rdma_bc_destroy(struct rpc_xprt * xprt,unsigned int reqs)135 void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs)
136 {
137 struct rpc_rqst *rqst, *tmp;
138
139 spin_lock(&xprt->bc_pa_lock);
140 list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) {
141 list_del(&rqst->rq_bc_pa_list);
142 spin_unlock(&xprt->bc_pa_lock);
143
144 rpcrdma_req_destroy(rpcr_to_rdmar(rqst));
145
146 spin_lock(&xprt->bc_pa_lock);
147 }
148 spin_unlock(&xprt->bc_pa_lock);
149 }
150
151 /**
152 * xprt_rdma_bc_free_rqst - Release a backchannel rqst
153 * @rqst: request to release
154 */
xprt_rdma_bc_free_rqst(struct rpc_rqst * rqst)155 void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst)
156 {
157 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
158 struct rpcrdma_rep *rep = req->rl_reply;
159 struct rpc_xprt *xprt = rqst->rq_xprt;
160 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
161
162 rpcrdma_rep_put(&r_xprt->rx_buf, rep);
163 req->rl_reply = NULL;
164
165 spin_lock(&xprt->bc_pa_lock);
166 list_add_tail(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
167 spin_unlock(&xprt->bc_pa_lock);
168 xprt_put(xprt);
169 }
170
rpcrdma_bc_rqst_get(struct rpcrdma_xprt * r_xprt)171 static struct rpc_rqst *rpcrdma_bc_rqst_get(struct rpcrdma_xprt *r_xprt)
172 {
173 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
174 struct rpcrdma_req *req;
175 struct rpc_rqst *rqst;
176 size_t size;
177
178 spin_lock(&xprt->bc_pa_lock);
179 rqst = list_first_entry_or_null(&xprt->bc_pa_list, struct rpc_rqst,
180 rq_bc_pa_list);
181 if (!rqst)
182 goto create_req;
183 list_del(&rqst->rq_bc_pa_list);
184 spin_unlock(&xprt->bc_pa_lock);
185 return rqst;
186
187 create_req:
188 spin_unlock(&xprt->bc_pa_lock);
189
190 /* Set a limit to prevent a remote from overrunning our resources.
191 */
192 if (xprt->bc_alloc_count >= RPCRDMA_BACKWARD_WRS)
193 return NULL;
194
195 size = min_t(size_t, r_xprt->rx_ep->re_inline_recv, PAGE_SIZE);
196 req = rpcrdma_req_create(r_xprt, size, GFP_KERNEL);
197 if (!req)
198 return NULL;
199 if (rpcrdma_req_setup(r_xprt, req)) {
200 rpcrdma_req_destroy(req);
201 return NULL;
202 }
203
204 xprt->bc_alloc_count++;
205 rqst = &req->rl_slot;
206 rqst->rq_xprt = xprt;
207 __set_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state);
208 xdr_buf_init(&rqst->rq_snd_buf, rdmab_data(req->rl_sendbuf), size);
209 return rqst;
210 }
211
212 /**
213 * rpcrdma_bc_receive_call - Handle a reverse-direction Call
214 * @r_xprt: transport receiving the call
215 * @rep: receive buffer containing the call
216 *
217 * Operational assumptions:
218 * o Backchannel credits are ignored, just as the NFS server
219 * forechannel currently does
220 * o The ULP manages a replay cache (eg, NFSv4.1 sessions).
221 * No replay detection is done at the transport level
222 */
rpcrdma_bc_receive_call(struct rpcrdma_xprt * r_xprt,struct rpcrdma_rep * rep)223 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt,
224 struct rpcrdma_rep *rep)
225 {
226 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
227 struct svc_serv *bc_serv;
228 struct rpcrdma_req *req;
229 struct rpc_rqst *rqst;
230 struct xdr_buf *buf;
231 size_t size;
232 __be32 *p;
233
234 p = xdr_inline_decode(&rep->rr_stream, 0);
235 size = xdr_stream_remaining(&rep->rr_stream);
236
237 #ifdef RPCRDMA_BACKCHANNEL_DEBUG
238 pr_info("RPC: %s: callback XID %08x, length=%u\n",
239 __func__, be32_to_cpup(p), size);
240 pr_info("RPC: %s: %*ph\n", __func__, size, p);
241 #endif
242
243 rqst = rpcrdma_bc_rqst_get(r_xprt);
244 if (!rqst)
245 goto out_overflow;
246
247 rqst->rq_reply_bytes_recvd = 0;
248 rqst->rq_xid = *p;
249
250 rqst->rq_private_buf.len = size;
251
252 buf = &rqst->rq_rcv_buf;
253 memset(buf, 0, sizeof(*buf));
254 buf->head[0].iov_base = p;
255 buf->head[0].iov_len = size;
256 buf->len = size;
257
258 /* The receive buffer has to be hooked to the rpcrdma_req
259 * so that it is not released while the req is pointing
260 * to its buffer, and so that it can be reposted after
261 * the Upper Layer is done decoding it.
262 */
263 req = rpcr_to_rdmar(rqst);
264 req->rl_reply = rep;
265 trace_xprtrdma_cb_call(r_xprt, rqst);
266
267 /* Queue rqst for ULP's callback service */
268 bc_serv = xprt->bc_serv;
269 xprt_get(xprt);
270 spin_lock(&bc_serv->sv_cb_lock);
271 list_add(&rqst->rq_bc_list, &bc_serv->sv_cb_list);
272 spin_unlock(&bc_serv->sv_cb_lock);
273
274 wake_up(&bc_serv->sv_cb_waitq);
275
276 r_xprt->rx_stats.bcall_count++;
277 return;
278
279 out_overflow:
280 pr_warn("RPC/RDMA backchannel overflow\n");
281 xprt_force_disconnect(xprt);
282 /* This receive buffer gets reposted automatically
283 * when the connection is re-established.
284 */
285 return;
286 }
287