1 /*
2 * Copyright (c) 2015 Oracle. All rights reserved.
3 *
4 * Support for backward direction RPCs on RPC/RDMA.
5 */
6
7 #include <linux/module.h>
8 #include <linux/sunrpc/xprt.h>
9 #include <linux/sunrpc/svc.h>
10 #include <linux/sunrpc/svc_xprt.h>
11
12 #include "xprt_rdma.h"
13
14 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
15 # define RPCDBG_FACILITY RPCDBG_TRANS
16 #endif
17
18 #define RPCRDMA_BACKCHANNEL_DEBUG
19
rpcrdma_bc_free_rqst(struct rpcrdma_xprt * r_xprt,struct rpc_rqst * rqst)20 static void rpcrdma_bc_free_rqst(struct rpcrdma_xprt *r_xprt,
21 struct rpc_rqst *rqst)
22 {
23 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
24 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
25
26 spin_lock(&buf->rb_reqslock);
27 list_del(&req->rl_all);
28 spin_unlock(&buf->rb_reqslock);
29
30 rpcrdma_destroy_req(&r_xprt->rx_ia, req);
31
32 kfree(rqst);
33 }
34
rpcrdma_bc_setup_rqst(struct rpcrdma_xprt * r_xprt,struct rpc_rqst * rqst)35 static int rpcrdma_bc_setup_rqst(struct rpcrdma_xprt *r_xprt,
36 struct rpc_rqst *rqst)
37 {
38 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
39 struct rpcrdma_regbuf *rb;
40 struct rpcrdma_req *req;
41 struct xdr_buf *buf;
42 size_t size;
43
44 req = rpcrdma_create_req(r_xprt);
45 if (IS_ERR(req))
46 return PTR_ERR(req);
47 req->rl_backchannel = true;
48
49 size = RPCRDMA_INLINE_WRITE_THRESHOLD(rqst);
50 rb = rpcrdma_alloc_regbuf(ia, size, GFP_KERNEL);
51 if (IS_ERR(rb))
52 goto out_fail;
53 req->rl_rdmabuf = rb;
54
55 size += RPCRDMA_INLINE_READ_THRESHOLD(rqst);
56 rb = rpcrdma_alloc_regbuf(ia, size, GFP_KERNEL);
57 if (IS_ERR(rb))
58 goto out_fail;
59 rb->rg_owner = req;
60 req->rl_sendbuf = rb;
61 /* so that rpcr_to_rdmar works when receiving a request */
62 rqst->rq_buffer = (void *)req->rl_sendbuf->rg_base;
63
64 buf = &rqst->rq_snd_buf;
65 buf->head[0].iov_base = rqst->rq_buffer;
66 buf->head[0].iov_len = 0;
67 buf->tail[0].iov_base = NULL;
68 buf->tail[0].iov_len = 0;
69 buf->page_len = 0;
70 buf->len = 0;
71 buf->buflen = size;
72
73 return 0;
74
75 out_fail:
76 rpcrdma_bc_free_rqst(r_xprt, rqst);
77 return -ENOMEM;
78 }
79
80 /* Allocate and add receive buffers to the rpcrdma_buffer's
81 * existing list of rep's. These are released when the
82 * transport is destroyed.
83 */
rpcrdma_bc_setup_reps(struct rpcrdma_xprt * r_xprt,unsigned int count)84 static int rpcrdma_bc_setup_reps(struct rpcrdma_xprt *r_xprt,
85 unsigned int count)
86 {
87 int rc = 0;
88
89 while (count--) {
90 rc = rpcrdma_create_rep(r_xprt);
91 if (rc)
92 break;
93 }
94 return rc;
95 }
96
97 /**
98 * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests
99 * @xprt: transport associated with these backchannel resources
100 * @reqs: number of concurrent incoming requests to expect
101 *
102 * Returns 0 on success; otherwise a negative errno
103 */
xprt_rdma_bc_setup(struct rpc_xprt * xprt,unsigned int reqs)104 int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs)
105 {
106 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
107 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
108 struct rpc_rqst *rqst;
109 unsigned int i;
110 int rc;
111
112 /* The backchannel reply path returns each rpc_rqst to the
113 * bc_pa_list _after_ the reply is sent. If the server is
114 * faster than the client, it can send another backward
115 * direction request before the rpc_rqst is returned to the
116 * list. The client rejects the request in this case.
117 *
118 * Twice as many rpc_rqsts are prepared to ensure there is
119 * always an rpc_rqst available as soon as a reply is sent.
120 */
121 if (reqs > RPCRDMA_BACKWARD_WRS >> 1)
122 goto out_err;
123
124 for (i = 0; i < (reqs << 1); i++) {
125 rqst = kzalloc(sizeof(*rqst), GFP_KERNEL);
126 if (!rqst) {
127 pr_err("RPC: %s: Failed to create bc rpc_rqst\n",
128 __func__);
129 goto out_free;
130 }
131
132 rqst->rq_xprt = &r_xprt->rx_xprt;
133 INIT_LIST_HEAD(&rqst->rq_list);
134 INIT_LIST_HEAD(&rqst->rq_bc_list);
135
136 if (rpcrdma_bc_setup_rqst(r_xprt, rqst))
137 goto out_free;
138
139 spin_lock_bh(&xprt->bc_pa_lock);
140 list_add(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
141 spin_unlock_bh(&xprt->bc_pa_lock);
142 }
143
144 rc = rpcrdma_bc_setup_reps(r_xprt, reqs);
145 if (rc)
146 goto out_free;
147
148 rc = rpcrdma_ep_post_extra_recv(r_xprt, reqs);
149 if (rc)
150 goto out_free;
151
152 buffer->rb_bc_srv_max_requests = reqs;
153 request_module("svcrdma");
154
155 return 0;
156
157 out_free:
158 xprt_rdma_bc_destroy(xprt, reqs);
159
160 out_err:
161 pr_err("RPC: %s: setup backchannel transport failed\n", __func__);
162 return -ENOMEM;
163 }
164
165 /**
166 * xprt_rdma_bc_up - Create transport endpoint for backchannel service
167 * @serv: server endpoint
168 * @net: network namespace
169 *
170 * The "xprt" is an implied argument: it supplies the name of the
171 * backchannel transport class.
172 *
173 * Returns zero on success, negative errno on failure
174 */
xprt_rdma_bc_up(struct svc_serv * serv,struct net * net)175 int xprt_rdma_bc_up(struct svc_serv *serv, struct net *net)
176 {
177 int ret;
178
179 ret = svc_create_xprt(serv, "rdma-bc", net, PF_INET, 0, 0);
180 if (ret < 0)
181 return ret;
182 return 0;
183 }
184
185 /**
186 * rpcrdma_bc_marshal_reply - Send backwards direction reply
187 * @rqst: buffer containing RPC reply data
188 *
189 * Returns zero on success.
190 */
rpcrdma_bc_marshal_reply(struct rpc_rqst * rqst)191 int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst)
192 {
193 struct rpc_xprt *xprt = rqst->rq_xprt;
194 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
195 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
196 struct rpcrdma_msg *headerp;
197 size_t rpclen;
198
199 headerp = rdmab_to_msg(req->rl_rdmabuf);
200 headerp->rm_xid = rqst->rq_xid;
201 headerp->rm_vers = rpcrdma_version;
202 headerp->rm_credit =
203 cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests);
204 headerp->rm_type = rdma_msg;
205 headerp->rm_body.rm_chunks[0] = xdr_zero;
206 headerp->rm_body.rm_chunks[1] = xdr_zero;
207 headerp->rm_body.rm_chunks[2] = xdr_zero;
208
209 rpclen = rqst->rq_svec[0].iov_len;
210
211 pr_info("RPC: %s: rpclen %zd headerp 0x%p lkey 0x%x\n",
212 __func__, rpclen, headerp, rdmab_lkey(req->rl_rdmabuf));
213 pr_info("RPC: %s: RPC/RDMA: %*ph\n",
214 __func__, (int)RPCRDMA_HDRLEN_MIN, headerp);
215 pr_info("RPC: %s: RPC: %*ph\n",
216 __func__, (int)rpclen, rqst->rq_svec[0].iov_base);
217
218 req->rl_send_iov[0].addr = rdmab_addr(req->rl_rdmabuf);
219 req->rl_send_iov[0].length = RPCRDMA_HDRLEN_MIN;
220 req->rl_send_iov[0].lkey = rdmab_lkey(req->rl_rdmabuf);
221
222 req->rl_send_iov[1].addr = rdmab_addr(req->rl_sendbuf);
223 req->rl_send_iov[1].length = rpclen;
224 req->rl_send_iov[1].lkey = rdmab_lkey(req->rl_sendbuf);
225
226 req->rl_niovs = 2;
227 return 0;
228 }
229
230 /**
231 * xprt_rdma_bc_destroy - Release resources for handling backchannel requests
232 * @xprt: transport associated with these backchannel resources
233 * @reqs: number of incoming requests to destroy; ignored
234 */
xprt_rdma_bc_destroy(struct rpc_xprt * xprt,unsigned int reqs)235 void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs)
236 {
237 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
238 struct rpc_rqst *rqst, *tmp;
239
240 spin_lock_bh(&xprt->bc_pa_lock);
241 list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) {
242 list_del(&rqst->rq_bc_pa_list);
243 spin_unlock_bh(&xprt->bc_pa_lock);
244
245 rpcrdma_bc_free_rqst(r_xprt, rqst);
246
247 spin_lock_bh(&xprt->bc_pa_lock);
248 }
249 spin_unlock_bh(&xprt->bc_pa_lock);
250 }
251
252 /**
253 * xprt_rdma_bc_free_rqst - Release a backchannel rqst
254 * @rqst: request to release
255 */
xprt_rdma_bc_free_rqst(struct rpc_rqst * rqst)256 void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst)
257 {
258 struct rpc_xprt *xprt = rqst->rq_xprt;
259
260 smp_mb__before_atomic();
261 WARN_ON_ONCE(!test_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state));
262 clear_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state);
263 smp_mb__after_atomic();
264
265 spin_lock_bh(&xprt->bc_pa_lock);
266 list_add_tail(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
267 spin_unlock_bh(&xprt->bc_pa_lock);
268 }
269
270 /**
271 * rpcrdma_bc_receive_call - Handle a backward direction call
272 * @xprt: transport receiving the call
273 * @rep: receive buffer containing the call
274 *
275 * Called in the RPC reply handler, which runs in a tasklet.
276 * Be quick about it.
277 *
278 * Operational assumptions:
279 * o Backchannel credits are ignored, just as the NFS server
280 * forechannel currently does
281 * o The ULP manages a replay cache (eg, NFSv4.1 sessions).
282 * No replay detection is done at the transport level
283 */
rpcrdma_bc_receive_call(struct rpcrdma_xprt * r_xprt,struct rpcrdma_rep * rep)284 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt,
285 struct rpcrdma_rep *rep)
286 {
287 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
288 struct rpcrdma_msg *headerp;
289 struct svc_serv *bc_serv;
290 struct rpcrdma_req *req;
291 struct rpc_rqst *rqst;
292 struct xdr_buf *buf;
293 size_t size;
294 __be32 *p;
295
296 headerp = rdmab_to_msg(rep->rr_rdmabuf);
297 #ifdef RPCRDMA_BACKCHANNEL_DEBUG
298 pr_info("RPC: %s: callback XID %08x, length=%u\n",
299 __func__, be32_to_cpu(headerp->rm_xid), rep->rr_len);
300 pr_info("RPC: %s: %*ph\n", __func__, rep->rr_len, headerp);
301 #endif
302
303 /* Sanity check:
304 * Need at least enough bytes for RPC/RDMA header, as code
305 * here references the header fields by array offset. Also,
306 * backward calls are always inline, so ensure there
307 * are some bytes beyond the RPC/RDMA header.
308 */
309 if (rep->rr_len < RPCRDMA_HDRLEN_MIN + 24)
310 goto out_short;
311 p = (__be32 *)((unsigned char *)headerp + RPCRDMA_HDRLEN_MIN);
312 size = rep->rr_len - RPCRDMA_HDRLEN_MIN;
313
314 /* Grab a free bc rqst */
315 spin_lock(&xprt->bc_pa_lock);
316 if (list_empty(&xprt->bc_pa_list)) {
317 spin_unlock(&xprt->bc_pa_lock);
318 goto out_overflow;
319 }
320 rqst = list_first_entry(&xprt->bc_pa_list,
321 struct rpc_rqst, rq_bc_pa_list);
322 list_del(&rqst->rq_bc_pa_list);
323 spin_unlock(&xprt->bc_pa_lock);
324 #ifdef RPCRDMA_BACKCHANNEL_DEBUG
325 pr_info("RPC: %s: using rqst %p\n", __func__, rqst);
326 #endif
327
328 /* Prepare rqst */
329 rqst->rq_reply_bytes_recvd = 0;
330 rqst->rq_bytes_sent = 0;
331 rqst->rq_xid = headerp->rm_xid;
332
333 rqst->rq_private_buf.len = size;
334 set_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state);
335
336 buf = &rqst->rq_rcv_buf;
337 memset(buf, 0, sizeof(*buf));
338 buf->head[0].iov_base = p;
339 buf->head[0].iov_len = size;
340 buf->len = size;
341
342 /* The receive buffer has to be hooked to the rpcrdma_req
343 * so that it can be reposted after the server is done
344 * parsing it but just before sending the backward
345 * direction reply.
346 */
347 req = rpcr_to_rdmar(rqst);
348 #ifdef RPCRDMA_BACKCHANNEL_DEBUG
349 pr_info("RPC: %s: attaching rep %p to req %p\n",
350 __func__, rep, req);
351 #endif
352 req->rl_reply = rep;
353
354 /* Defeat the retransmit detection logic in send_request */
355 req->rl_connect_cookie = 0;
356
357 /* Queue rqst for ULP's callback service */
358 bc_serv = xprt->bc_serv;
359 spin_lock(&bc_serv->sv_cb_lock);
360 list_add(&rqst->rq_bc_list, &bc_serv->sv_cb_list);
361 spin_unlock(&bc_serv->sv_cb_lock);
362
363 wake_up(&bc_serv->sv_cb_waitq);
364
365 r_xprt->rx_stats.bcall_count++;
366 return;
367
368 out_overflow:
369 pr_warn("RPC/RDMA backchannel overflow\n");
370 xprt_disconnect_done(xprt);
371 /* This receive buffer gets reposted automatically
372 * when the connection is re-established.
373 */
374 return;
375
376 out_short:
377 pr_warn("RPC/RDMA short backward direction call\n");
378
379 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep))
380 xprt_disconnect_done(xprt);
381 else
382 pr_warn("RPC: %s: reposting rep %p\n",
383 __func__, rep);
384 }
385