GNU Linux-libre 4.14.266-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
13 #include "xprt_rdma.h"
14
15 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
16 # define RPCDBG_FACILITY        RPCDBG_TRANS
17 #endif
18
19 #undef RPCRDMA_BACKCHANNEL_DEBUG
20
21 static void rpcrdma_bc_free_rqst(struct rpcrdma_xprt *r_xprt,
22                                  struct rpc_rqst *rqst)
23 {
24         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
25         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
26
27         spin_lock(&buf->rb_reqslock);
28         list_del(&req->rl_all);
29         spin_unlock(&buf->rb_reqslock);
30
31         rpcrdma_destroy_req(req);
32
33         kfree(rqst);
34 }
35
36 static int rpcrdma_bc_setup_rqst(struct rpcrdma_xprt *r_xprt,
37                                  struct rpc_rqst *rqst)
38 {
39         struct rpcrdma_regbuf *rb;
40         struct rpcrdma_req *req;
41         size_t size;
42
43         req = rpcrdma_create_req(r_xprt);
44         if (IS_ERR(req))
45                 return PTR_ERR(req);
46         req->rl_backchannel = true;
47
48         rb = rpcrdma_alloc_regbuf(RPCRDMA_HDRBUF_SIZE,
49                                   DMA_TO_DEVICE, GFP_KERNEL);
50         if (IS_ERR(rb))
51                 goto out_fail;
52         req->rl_rdmabuf = rb;
53         xdr_buf_init(&req->rl_hdrbuf, rb->rg_base, rdmab_length(rb));
54
55         size = r_xprt->rx_data.inline_rsize;
56         rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, GFP_KERNEL);
57         if (IS_ERR(rb))
58                 goto out_fail;
59         req->rl_sendbuf = rb;
60         xdr_buf_init(&rqst->rq_snd_buf, rb->rg_base,
61                      min_t(size_t, size, PAGE_SIZE));
62         rpcrdma_set_xprtdata(rqst, req);
63         return 0;
64
65 out_fail:
66         rpcrdma_bc_free_rqst(r_xprt, rqst);
67         return -ENOMEM;
68 }
69
70 /* Allocate and add receive buffers to the rpcrdma_buffer's
71  * existing list of rep's. These are released when the
72  * transport is destroyed.
73  */
74 static int rpcrdma_bc_setup_reps(struct rpcrdma_xprt *r_xprt,
75                                  unsigned int count)
76 {
77         int rc = 0;
78
79         while (count--) {
80                 rc = rpcrdma_create_rep(r_xprt);
81                 if (rc)
82                         break;
83         }
84         return rc;
85 }
86
87 /**
88  * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests
89  * @xprt: transport associated with these backchannel resources
90  * @reqs: number of concurrent incoming requests to expect
91  *
92  * Returns 0 on success; otherwise a negative errno
93  */
94 int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs)
95 {
96         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
97         struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
98         struct rpc_rqst *rqst;
99         unsigned int i;
100         int rc;
101
102         /* The backchannel reply path returns each rpc_rqst to the
103          * bc_pa_list _after_ the reply is sent. If the server is
104          * faster than the client, it can send another backward
105          * direction request before the rpc_rqst is returned to the
106          * list. The client rejects the request in this case.
107          *
108          * Twice as many rpc_rqsts are prepared to ensure there is
109          * always an rpc_rqst available as soon as a reply is sent.
110          */
111         if (reqs > RPCRDMA_BACKWARD_WRS >> 1)
112                 goto out_err;
113
114         for (i = 0; i < (reqs << 1); i++) {
115                 rqst = kzalloc(sizeof(*rqst), GFP_KERNEL);
116                 if (!rqst)
117                         goto out_free;
118
119                 dprintk("RPC:       %s: new rqst %p\n", __func__, rqst);
120
121                 rqst->rq_xprt = &r_xprt->rx_xprt;
122                 INIT_LIST_HEAD(&rqst->rq_list);
123                 INIT_LIST_HEAD(&rqst->rq_bc_list);
124
125                 if (rpcrdma_bc_setup_rqst(r_xprt, rqst))
126                         goto out_free;
127
128                 spin_lock_bh(&xprt->bc_pa_lock);
129                 list_add(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
130                 spin_unlock_bh(&xprt->bc_pa_lock);
131         }
132
133         rc = rpcrdma_bc_setup_reps(r_xprt, reqs);
134         if (rc)
135                 goto out_free;
136
137         rc = rpcrdma_ep_post_extra_recv(r_xprt, reqs);
138         if (rc)
139                 goto out_free;
140
141         buffer->rb_bc_srv_max_requests = reqs;
142         request_module("svcrdma");
143
144         return 0;
145
146 out_free:
147         xprt_rdma_bc_destroy(xprt, reqs);
148
149 out_err:
150         pr_err("RPC:       %s: setup backchannel transport failed\n", __func__);
151         return -ENOMEM;
152 }
153
154 /**
155  * xprt_rdma_bc_up - Create transport endpoint for backchannel service
156  * @serv: server endpoint
157  * @net: network namespace
158  *
159  * The "xprt" is an implied argument: it supplies the name of the
160  * backchannel transport class.
161  *
162  * Returns zero on success, negative errno on failure
163  */
164 int xprt_rdma_bc_up(struct svc_serv *serv, struct net *net)
165 {
166         int ret;
167
168         ret = svc_create_xprt(serv, "rdma-bc", net, PF_INET, 0, 0);
169         if (ret < 0)
170                 return ret;
171         return 0;
172 }
173
174 /**
175  * xprt_rdma_bc_maxpayload - Return maximum backchannel message size
176  * @xprt: transport
177  *
178  * Returns maximum size, in bytes, of a backchannel message
179  */
180 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *xprt)
181 {
182         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
183         struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
184         size_t maxmsg;
185
186         maxmsg = min_t(unsigned int, cdata->inline_rsize, cdata->inline_wsize);
187         maxmsg = min_t(unsigned int, maxmsg, PAGE_SIZE);
188         return maxmsg - RPCRDMA_HDRLEN_MIN;
189 }
190
191 /**
192  * rpcrdma_bc_marshal_reply - Send backwards direction reply
193  * @rqst: buffer containing RPC reply data
194  *
195  * Returns zero on success.
196  */
197 int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst)
198 {
199         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
200         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
201         __be32 *p;
202
203         rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
204         xdr_init_encode(&req->rl_stream, &req->rl_hdrbuf,
205                         req->rl_rdmabuf->rg_base);
206
207         p = xdr_reserve_space(&req->rl_stream, 28);
208         if (unlikely(!p))
209                 return -EIO;
210         *p++ = rqst->rq_xid;
211         *p++ = rpcrdma_version;
212         *p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests);
213         *p++ = rdma_msg;
214         *p++ = xdr_zero;
215         *p++ = xdr_zero;
216         *p = xdr_zero;
217
218         if (!rpcrdma_prepare_send_sges(&r_xprt->rx_ia, req, RPCRDMA_HDRLEN_MIN,
219                                        &rqst->rq_snd_buf, rpcrdma_noch))
220                 return -EIO;
221         return 0;
222 }
223
224 /**
225  * xprt_rdma_bc_destroy - Release resources for handling backchannel requests
226  * @xprt: transport associated with these backchannel resources
227  * @reqs: number of incoming requests to destroy; ignored
228  */
229 void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs)
230 {
231         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
232         struct rpc_rqst *rqst, *tmp;
233
234         spin_lock_bh(&xprt->bc_pa_lock);
235         list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) {
236                 list_del(&rqst->rq_bc_pa_list);
237                 spin_unlock_bh(&xprt->bc_pa_lock);
238
239                 rpcrdma_bc_free_rqst(r_xprt, rqst);
240
241                 spin_lock_bh(&xprt->bc_pa_lock);
242         }
243         spin_unlock_bh(&xprt->bc_pa_lock);
244 }
245
246 /**
247  * xprt_rdma_bc_free_rqst - Release a backchannel rqst
248  * @rqst: request to release
249  */
250 void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst)
251 {
252         struct rpc_xprt *xprt = rqst->rq_xprt;
253
254         dprintk("RPC:       %s: freeing rqst %p (req %p)\n",
255                 __func__, rqst, rpcr_to_rdmar(rqst));
256
257         smp_mb__before_atomic();
258         WARN_ON_ONCE(!test_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state));
259         clear_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state);
260         smp_mb__after_atomic();
261
262         spin_lock_bh(&xprt->bc_pa_lock);
263         list_add_tail(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
264         spin_unlock_bh(&xprt->bc_pa_lock);
265 }
266
267 /**
268  * rpcrdma_bc_receive_call - Handle a backward direction call
269  * @xprt: transport receiving the call
270  * @rep: receive buffer containing the call
271  *
272  * Operational assumptions:
273  *    o Backchannel credits are ignored, just as the NFS server
274  *      forechannel currently does
275  *    o The ULP manages a replay cache (eg, NFSv4.1 sessions).
276  *      No replay detection is done at the transport level
277  */
278 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt,
279                              struct rpcrdma_rep *rep)
280 {
281         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
282         struct svc_serv *bc_serv;
283         struct rpcrdma_req *req;
284         struct rpc_rqst *rqst;
285         struct xdr_buf *buf;
286         size_t size;
287         __be32 *p;
288
289         p = xdr_inline_decode(&rep->rr_stream, 0);
290         size = xdr_stream_remaining(&rep->rr_stream);
291
292 #ifdef RPCRDMA_BACKCHANNEL_DEBUG
293         pr_info("RPC:       %s: callback XID %08x, length=%u\n",
294                 __func__, be32_to_cpup(p), size);
295         pr_info("RPC:       %s: %*ph\n", __func__, size, p);
296 #endif
297
298         /* Grab a free bc rqst */
299         spin_lock(&xprt->bc_pa_lock);
300         if (list_empty(&xprt->bc_pa_list)) {
301                 spin_unlock(&xprt->bc_pa_lock);
302                 goto out_overflow;
303         }
304         rqst = list_first_entry(&xprt->bc_pa_list,
305                                 struct rpc_rqst, rq_bc_pa_list);
306         list_del(&rqst->rq_bc_pa_list);
307         spin_unlock(&xprt->bc_pa_lock);
308         dprintk("RPC:       %s: using rqst %p\n", __func__, rqst);
309
310         /* Prepare rqst */
311         rqst->rq_reply_bytes_recvd = 0;
312         rqst->rq_bytes_sent = 0;
313         rqst->rq_xid = *p;
314
315         rqst->rq_private_buf.len = size;
316         set_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state);
317
318         buf = &rqst->rq_rcv_buf;
319         memset(buf, 0, sizeof(*buf));
320         buf->head[0].iov_base = p;
321         buf->head[0].iov_len = size;
322         buf->len = size;
323
324         /* The receive buffer has to be hooked to the rpcrdma_req
325          * so that it is not released while the req is pointing
326          * to its buffer, and so that it can be reposted after
327          * the Upper Layer is done decoding it.
328          */
329         req = rpcr_to_rdmar(rqst);
330         dprintk("RPC:       %s: attaching rep %p to req %p\n",
331                 __func__, rep, req);
332         req->rl_reply = rep;
333
334         /* Defeat the retransmit detection logic in send_request */
335         req->rl_connect_cookie = 0;
336
337         /* Queue rqst for ULP's callback service */
338         bc_serv = xprt->bc_serv;
339         spin_lock(&bc_serv->sv_cb_lock);
340         list_add(&rqst->rq_bc_list, &bc_serv->sv_cb_list);
341         spin_unlock(&bc_serv->sv_cb_lock);
342
343         wake_up(&bc_serv->sv_cb_waitq);
344
345         r_xprt->rx_stats.bcall_count++;
346         return;
347
348 out_overflow:
349         pr_warn("RPC/RDMA backchannel overflow\n");
350         xprt_disconnect_done(xprt);
351         /* This receive buffer gets reposted automatically
352          * when the connection is re-established.
353          */
354         return;
355 }