GNU Linux-libre 4.19.286-gnu1
[releases.git] / net / sunrpc / xprtrdma / backchannel.c
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/module.h>
9 #include <linux/sunrpc/xprt.h>
10 #include <linux/sunrpc/svc.h>
11 #include <linux/sunrpc/svc_xprt.h>
12 #include <linux/sunrpc/svc_rdma.h>
13
14 #include "xprt_rdma.h"
15 #include <trace/events/rpcrdma.h>
16
17 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
18 # define RPCDBG_FACILITY        RPCDBG_TRANS
19 #endif
20
21 #undef RPCRDMA_BACKCHANNEL_DEBUG
22
23 static void rpcrdma_bc_free_rqst(struct rpcrdma_xprt *r_xprt,
24                                  struct rpc_rqst *rqst)
25 {
26         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
27         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
28
29         spin_lock(&buf->rb_reqslock);
30         list_del(&req->rl_all);
31         spin_unlock(&buf->rb_reqslock);
32
33         rpcrdma_destroy_req(req);
34 }
35
36 static int rpcrdma_bc_setup_reqs(struct rpcrdma_xprt *r_xprt,
37                                  unsigned int count)
38 {
39         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
40         struct rpc_rqst *rqst;
41         unsigned int i;
42
43         for (i = 0; i < (count << 1); i++) {
44                 struct rpcrdma_regbuf *rb;
45                 struct rpcrdma_req *req;
46                 size_t size;
47
48                 req = rpcrdma_create_req(r_xprt);
49                 if (IS_ERR(req))
50                         return PTR_ERR(req);
51                 rqst = &req->rl_slot;
52
53                 rqst->rq_xprt = xprt;
54                 INIT_LIST_HEAD(&rqst->rq_list);
55                 INIT_LIST_HEAD(&rqst->rq_bc_list);
56                 __set_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state);
57                 spin_lock_bh(&xprt->bc_pa_lock);
58                 list_add(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
59                 spin_unlock_bh(&xprt->bc_pa_lock);
60
61                 size = r_xprt->rx_data.inline_rsize;
62                 rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, GFP_KERNEL);
63                 if (IS_ERR(rb))
64                         goto out_fail;
65                 req->rl_sendbuf = rb;
66                 xdr_buf_init(&rqst->rq_snd_buf, rb->rg_base,
67                              min_t(size_t, size, PAGE_SIZE));
68         }
69         return 0;
70
71 out_fail:
72         rpcrdma_bc_free_rqst(r_xprt, rqst);
73         return -ENOMEM;
74 }
75
76 /**
77  * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests
78  * @xprt: transport associated with these backchannel resources
79  * @reqs: number of concurrent incoming requests to expect
80  *
81  * Returns 0 on success; otherwise a negative errno
82  */
83 int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs)
84 {
85         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
86         int rc;
87
88         /* The backchannel reply path returns each rpc_rqst to the
89          * bc_pa_list _after_ the reply is sent. If the server is
90          * faster than the client, it can send another backward
91          * direction request before the rpc_rqst is returned to the
92          * list. The client rejects the request in this case.
93          *
94          * Twice as many rpc_rqsts are prepared to ensure there is
95          * always an rpc_rqst available as soon as a reply is sent.
96          */
97         if (reqs > RPCRDMA_BACKWARD_WRS >> 1)
98                 goto out_err;
99
100         rc = rpcrdma_bc_setup_reqs(r_xprt, reqs);
101         if (rc)
102                 goto out_free;
103
104         r_xprt->rx_buf.rb_bc_srv_max_requests = reqs;
105         request_module("svcrdma");
106         trace_xprtrdma_cb_setup(r_xprt, reqs);
107         return 0;
108
109 out_free:
110         xprt_rdma_bc_destroy(xprt, reqs);
111
112 out_err:
113         pr_err("RPC:       %s: setup backchannel transport failed\n", __func__);
114         return -ENOMEM;
115 }
116
117 /**
118  * xprt_rdma_bc_up - Create transport endpoint for backchannel service
119  * @serv: server endpoint
120  * @net: network namespace
121  *
122  * The "xprt" is an implied argument: it supplies the name of the
123  * backchannel transport class.
124  *
125  * Returns zero on success, negative errno on failure
126  */
127 int xprt_rdma_bc_up(struct svc_serv *serv, struct net *net)
128 {
129         int ret;
130
131         ret = svc_create_xprt(serv, "rdma-bc", net, PF_INET, 0, 0);
132         if (ret < 0)
133                 return ret;
134         return 0;
135 }
136
137 /**
138  * xprt_rdma_bc_maxpayload - Return maximum backchannel message size
139  * @xprt: transport
140  *
141  * Returns maximum size, in bytes, of a backchannel message
142  */
143 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *xprt)
144 {
145         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
146         struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
147         size_t maxmsg;
148
149         maxmsg = min_t(unsigned int, cdata->inline_rsize, cdata->inline_wsize);
150         maxmsg = min_t(unsigned int, maxmsg, PAGE_SIZE);
151         return maxmsg - RPCRDMA_HDRLEN_MIN;
152 }
153
154 static int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst)
155 {
156         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
157         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
158         __be32 *p;
159
160         rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
161         xdr_init_encode(&req->rl_stream, &req->rl_hdrbuf,
162                         req->rl_rdmabuf->rg_base);
163
164         p = xdr_reserve_space(&req->rl_stream, 28);
165         if (unlikely(!p))
166                 return -EIO;
167         *p++ = rqst->rq_xid;
168         *p++ = rpcrdma_version;
169         *p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests);
170         *p++ = rdma_msg;
171         *p++ = xdr_zero;
172         *p++ = xdr_zero;
173         *p = xdr_zero;
174
175         if (rpcrdma_prepare_send_sges(r_xprt, req, RPCRDMA_HDRLEN_MIN,
176                                       &rqst->rq_snd_buf, rpcrdma_noch))
177                 return -EIO;
178
179         trace_xprtrdma_cb_reply(rqst);
180         return 0;
181 }
182
183 /**
184  * xprt_rdma_bc_send_reply - marshal and send a backchannel reply
185  * @rqst: RPC rqst with a backchannel RPC reply in rq_snd_buf
186  *
187  * Caller holds the transport's write lock.
188  *
189  * Returns:
190  *      %0 if the RPC message has been sent
191  *      %-ENOTCONN if the caller should reconnect and call again
192  *      %-EIO if a permanent error occurred and the request was not
193  *              sent. Do not try to send this message again.
194  */
195 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst)
196 {
197         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
198         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
199         int rc;
200
201         if (!xprt_connected(rqst->rq_xprt))
202                 goto drop_connection;
203
204         rc = rpcrdma_bc_marshal_reply(rqst);
205         if (rc < 0)
206                 goto failed_marshal;
207
208         rpcrdma_post_recvs(r_xprt, true);
209         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
210                 goto drop_connection;
211         return 0;
212
213 failed_marshal:
214         if (rc != -ENOTCONN)
215                 return rc;
216 drop_connection:
217         xprt_disconnect_done(rqst->rq_xprt);
218         return -ENOTCONN;
219 }
220
221 /**
222  * xprt_rdma_bc_destroy - Release resources for handling backchannel requests
223  * @xprt: transport associated with these backchannel resources
224  * @reqs: number of incoming requests to destroy; ignored
225  */
226 void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs)
227 {
228         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
229         struct rpc_rqst *rqst, *tmp;
230
231         spin_lock_bh(&xprt->bc_pa_lock);
232         list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) {
233                 list_del(&rqst->rq_bc_pa_list);
234                 spin_unlock_bh(&xprt->bc_pa_lock);
235
236                 rpcrdma_bc_free_rqst(r_xprt, rqst);
237
238                 spin_lock_bh(&xprt->bc_pa_lock);
239         }
240         spin_unlock_bh(&xprt->bc_pa_lock);
241 }
242
243 /**
244  * xprt_rdma_bc_free_rqst - Release a backchannel rqst
245  * @rqst: request to release
246  */
247 void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst)
248 {
249         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
250         struct rpc_xprt *xprt = rqst->rq_xprt;
251
252         dprintk("RPC:       %s: freeing rqst %p (req %p)\n",
253                 __func__, rqst, req);
254
255         rpcrdma_recv_buffer_put(req->rl_reply);
256         req->rl_reply = NULL;
257
258         spin_lock_bh(&xprt->bc_pa_lock);
259         list_add_tail(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
260         spin_unlock_bh(&xprt->bc_pa_lock);
261 }
262
263 /**
264  * rpcrdma_bc_receive_call - Handle a backward direction call
265  * @r_xprt: transport receiving the call
266  * @rep: receive buffer containing the call
267  *
268  * Operational assumptions:
269  *    o Backchannel credits are ignored, just as the NFS server
270  *      forechannel currently does
271  *    o The ULP manages a replay cache (eg, NFSv4.1 sessions).
272  *      No replay detection is done at the transport level
273  */
274 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt,
275                              struct rpcrdma_rep *rep)
276 {
277         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
278         struct svc_serv *bc_serv;
279         struct rpcrdma_req *req;
280         struct rpc_rqst *rqst;
281         struct xdr_buf *buf;
282         size_t size;
283         __be32 *p;
284
285         p = xdr_inline_decode(&rep->rr_stream, 0);
286         size = xdr_stream_remaining(&rep->rr_stream);
287
288 #ifdef RPCRDMA_BACKCHANNEL_DEBUG
289         pr_info("RPC:       %s: callback XID %08x, length=%u\n",
290                 __func__, be32_to_cpup(p), size);
291         pr_info("RPC:       %s: %*ph\n", __func__, size, p);
292 #endif
293
294         /* Grab a free bc rqst */
295         spin_lock(&xprt->bc_pa_lock);
296         if (list_empty(&xprt->bc_pa_list)) {
297                 spin_unlock(&xprt->bc_pa_lock);
298                 goto out_overflow;
299         }
300         rqst = list_first_entry(&xprt->bc_pa_list,
301                                 struct rpc_rqst, rq_bc_pa_list);
302         list_del(&rqst->rq_bc_pa_list);
303         spin_unlock(&xprt->bc_pa_lock);
304
305         /* Prepare rqst */
306         rqst->rq_reply_bytes_recvd = 0;
307         rqst->rq_bytes_sent = 0;
308         rqst->rq_xid = *p;
309
310         rqst->rq_private_buf.len = size;
311
312         buf = &rqst->rq_rcv_buf;
313         memset(buf, 0, sizeof(*buf));
314         buf->head[0].iov_base = p;
315         buf->head[0].iov_len = size;
316         buf->len = size;
317
318         /* The receive buffer has to be hooked to the rpcrdma_req
319          * so that it is not released while the req is pointing
320          * to its buffer, and so that it can be reposted after
321          * the Upper Layer is done decoding it.
322          */
323         req = rpcr_to_rdmar(rqst);
324         req->rl_reply = rep;
325         trace_xprtrdma_cb_call(rqst);
326
327         /* Queue rqst for ULP's callback service */
328         bc_serv = xprt->bc_serv;
329         spin_lock(&bc_serv->sv_cb_lock);
330         list_add(&rqst->rq_bc_list, &bc_serv->sv_cb_list);
331         spin_unlock(&bc_serv->sv_cb_lock);
332
333         wake_up(&bc_serv->sv_cb_waitq);
334
335         r_xprt->rx_stats.bcall_count++;
336         return;
337
338 out_overflow:
339         pr_warn("RPC/RDMA backchannel overflow\n");
340         xprt_disconnect_done(xprt);
341         /* This receive buffer gets reposted automatically
342          * when the connection is re-established.
343          */
344         return;
345 }