1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (c) 2015-2018 Oracle. All rights reserved.
4 *
5 * Support for backward direction RPCs on RPC/RDMA (server-side).
6 */
7
8 #include <linux/sunrpc/svc_rdma.h>
9
10 #include "xprt_rdma.h"
11 #include <trace/events/rpcrdma.h>
12
13 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
14
15 #undef SVCRDMA_BACKCHANNEL_DEBUG
16
17 /**
18 * svc_rdma_handle_bc_reply - Process incoming backchannel Reply
19 * @rqstp: resources for handling the Reply
20 * @rctxt: Received message
21 *
22 */
svc_rdma_handle_bc_reply(struct svc_rqst * rqstp,struct svc_rdma_recv_ctxt * rctxt)23 void svc_rdma_handle_bc_reply(struct svc_rqst *rqstp,
24 struct svc_rdma_recv_ctxt *rctxt)
25 {
26 struct svc_xprt *sxprt = rqstp->rq_xprt;
27 struct rpc_xprt *xprt = sxprt->xpt_bc_xprt;
28 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
29 struct xdr_buf *rcvbuf = &rqstp->rq_arg;
30 struct kvec *dst, *src = &rcvbuf->head[0];
31 __be32 *rdma_resp = rctxt->rc_recv_buf;
32 struct rpc_rqst *req;
33 u32 credits;
34 size_t len;
35 __be32 xid;
36 __be32 *p;
37
38 p = (__be32 *)src->iov_base;
39 len = src->iov_len;
40 xid = *rdma_resp;
41
42 #ifdef SVCRDMA_BACKCHANNEL_DEBUG
43 pr_info("%s: xid=%08x, length=%zu\n",
44 __func__, be32_to_cpu(xid), len);
45 pr_info("%s: RPC/RDMA: %*ph\n",
46 __func__, (int)RPCRDMA_HDRLEN_MIN, rdma_resp);
47 pr_info("%s: RPC: %*ph\n",
48 __func__, (int)len, p);
49 #endif
50
51 spin_lock(&xprt->queue_lock);
52 req = xprt_lookup_rqst(xprt, xid);
53 if (!req)
54 goto out_unlock;
55
56 dst = &req->rq_private_buf.head[0];
57 memcpy(&req->rq_private_buf, &req->rq_rcv_buf, sizeof(struct xdr_buf));
58 if (dst->iov_len < len)
59 goto out_unlock;
60 memcpy(dst->iov_base, p, len);
61 xprt_pin_rqst(req);
62 spin_unlock(&xprt->queue_lock);
63
64 credits = be32_to_cpup(rdma_resp + 2);
65 if (credits == 0)
66 credits = 1; /* don't deadlock */
67 else if (credits > r_xprt->rx_buf.rb_bc_max_requests)
68 credits = r_xprt->rx_buf.rb_bc_max_requests;
69
70 spin_lock(&xprt->transport_lock);
71 xprt->cwnd = credits << RPC_CWNDSHIFT;
72 spin_unlock(&xprt->transport_lock);
73
74 spin_lock(&xprt->queue_lock);
75 xprt_complete_rqst(req->rq_task, rcvbuf->len);
76 xprt_unpin_rqst(req);
77 rcvbuf->len = 0;
78
79 out_unlock:
80 spin_unlock(&xprt->queue_lock);
81 }
82
83 /* Send a backwards direction RPC call.
84 *
85 * Caller holds the connection's mutex and has already marshaled
86 * the RPC/RDMA request.
87 *
88 * This is similar to svc_rdma_send_reply_msg, but takes a struct
89 * rpc_rqst instead, does not support chunks, and avoids blocking
90 * memory allocation.
91 *
92 * XXX: There is still an opportunity to block in svc_rdma_send()
93 * if there are no SQ entries to post the Send. This may occur if
94 * the adapter has a small maximum SQ depth.
95 */
svc_rdma_bc_sendto(struct svcxprt_rdma * rdma,struct rpc_rqst * rqst,struct svc_rdma_send_ctxt * ctxt)96 static int svc_rdma_bc_sendto(struct svcxprt_rdma *rdma,
97 struct rpc_rqst *rqst,
98 struct svc_rdma_send_ctxt *ctxt)
99 {
100 int ret;
101
102 ret = svc_rdma_map_reply_msg(rdma, ctxt, &rqst->rq_snd_buf, NULL);
103 if (ret < 0)
104 return -EIO;
105
106 /* Bump page refcnt so Send completion doesn't release
107 * the rq_buffer before all retransmits are complete.
108 */
109 get_page(virt_to_page(rqst->rq_buffer));
110 ctxt->sc_send_wr.opcode = IB_WR_SEND;
111 return svc_rdma_send(rdma, &ctxt->sc_send_wr);
112 }
113
114 /* Server-side transport endpoint wants a whole page for its send
115 * buffer. The client RPC code constructs the RPC header in this
116 * buffer before it invokes ->send_request.
117 */
118 static int
xprt_rdma_bc_allocate(struct rpc_task * task)119 xprt_rdma_bc_allocate(struct rpc_task *task)
120 {
121 struct rpc_rqst *rqst = task->tk_rqstp;
122 size_t size = rqst->rq_callsize;
123 struct page *page;
124
125 if (size > PAGE_SIZE) {
126 WARN_ONCE(1, "svcrdma: large bc buffer request (size %zu)\n",
127 size);
128 return -EINVAL;
129 }
130
131 page = alloc_page(RPCRDMA_DEF_GFP);
132 if (!page)
133 return -ENOMEM;
134 rqst->rq_buffer = page_address(page);
135
136 rqst->rq_rbuffer = kmalloc(rqst->rq_rcvsize, RPCRDMA_DEF_GFP);
137 if (!rqst->rq_rbuffer) {
138 put_page(page);
139 return -ENOMEM;
140 }
141 return 0;
142 }
143
144 static void
xprt_rdma_bc_free(struct rpc_task * task)145 xprt_rdma_bc_free(struct rpc_task *task)
146 {
147 struct rpc_rqst *rqst = task->tk_rqstp;
148
149 put_page(virt_to_page(rqst->rq_buffer));
150 kfree(rqst->rq_rbuffer);
151 }
152
153 static int
rpcrdma_bc_send_request(struct svcxprt_rdma * rdma,struct rpc_rqst * rqst)154 rpcrdma_bc_send_request(struct svcxprt_rdma *rdma, struct rpc_rqst *rqst)
155 {
156 struct rpc_xprt *xprt = rqst->rq_xprt;
157 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
158 struct svc_rdma_send_ctxt *ctxt;
159 __be32 *p;
160 int rc;
161
162 ctxt = svc_rdma_send_ctxt_get(rdma);
163 if (!ctxt)
164 goto drop_connection;
165
166 p = ctxt->sc_xprt_buf;
167 *p++ = rqst->rq_xid;
168 *p++ = rpcrdma_version;
169 *p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_max_requests);
170 *p++ = rdma_msg;
171 *p++ = xdr_zero;
172 *p++ = xdr_zero;
173 *p = xdr_zero;
174 svc_rdma_sync_reply_hdr(rdma, ctxt, RPCRDMA_HDRLEN_MIN);
175
176 #ifdef SVCRDMA_BACKCHANNEL_DEBUG
177 pr_info("%s: %*ph\n", __func__, 64, rqst->rq_buffer);
178 #endif
179
180 rqst->rq_xtime = ktime_get();
181 rc = svc_rdma_bc_sendto(rdma, rqst, ctxt);
182 if (rc) {
183 svc_rdma_send_ctxt_put(rdma, ctxt);
184 goto drop_connection;
185 }
186 return 0;
187
188 drop_connection:
189 dprintk("svcrdma: failed to send bc call\n");
190 return -ENOTCONN;
191 }
192
193 /* Send an RPC call on the passive end of a transport
194 * connection.
195 */
196 static int
xprt_rdma_bc_send_request(struct rpc_rqst * rqst)197 xprt_rdma_bc_send_request(struct rpc_rqst *rqst)
198 {
199 struct svc_xprt *sxprt = rqst->rq_xprt->bc_xprt;
200 struct svcxprt_rdma *rdma;
201 int ret;
202
203 dprintk("svcrdma: sending bc call with xid: %08x\n",
204 be32_to_cpu(rqst->rq_xid));
205
206 mutex_lock(&sxprt->xpt_mutex);
207
208 ret = -ENOTCONN;
209 rdma = container_of(sxprt, struct svcxprt_rdma, sc_xprt);
210 if (!test_bit(XPT_DEAD, &sxprt->xpt_flags)) {
211 ret = rpcrdma_bc_send_request(rdma, rqst);
212 if (ret == -ENOTCONN)
213 svc_close_xprt(sxprt);
214 }
215
216 mutex_unlock(&sxprt->xpt_mutex);
217
218 if (ret < 0)
219 return ret;
220 return 0;
221 }
222
223 static void
xprt_rdma_bc_close(struct rpc_xprt * xprt)224 xprt_rdma_bc_close(struct rpc_xprt *xprt)
225 {
226 dprintk("svcrdma: %s: xprt %p\n", __func__, xprt);
227
228 xprt_disconnect_done(xprt);
229 xprt->cwnd = RPC_CWNDSHIFT;
230 }
231
232 static void
xprt_rdma_bc_put(struct rpc_xprt * xprt)233 xprt_rdma_bc_put(struct rpc_xprt *xprt)
234 {
235 dprintk("svcrdma: %s: xprt %p\n", __func__, xprt);
236
237 xprt_rdma_free_addresses(xprt);
238 xprt_free(xprt);
239 }
240
241 static const struct rpc_xprt_ops xprt_rdma_bc_procs = {
242 .reserve_xprt = xprt_reserve_xprt_cong,
243 .release_xprt = xprt_release_xprt_cong,
244 .alloc_slot = xprt_alloc_slot,
245 .free_slot = xprt_free_slot,
246 .release_request = xprt_release_rqst_cong,
247 .buf_alloc = xprt_rdma_bc_allocate,
248 .buf_free = xprt_rdma_bc_free,
249 .send_request = xprt_rdma_bc_send_request,
250 .wait_for_reply_request = xprt_wait_for_reply_request_def,
251 .close = xprt_rdma_bc_close,
252 .destroy = xprt_rdma_bc_put,
253 .print_stats = xprt_rdma_print_stats
254 };
255
256 static const struct rpc_timeout xprt_rdma_bc_timeout = {
257 .to_initval = 60 * HZ,
258 .to_maxval = 60 * HZ,
259 };
260
261 /* It shouldn't matter if the number of backchannel session slots
262 * doesn't match the number of RPC/RDMA credits. That just means
263 * one or the other will have extra slots that aren't used.
264 */
265 static struct rpc_xprt *
xprt_setup_rdma_bc(struct xprt_create * args)266 xprt_setup_rdma_bc(struct xprt_create *args)
267 {
268 struct rpc_xprt *xprt;
269 struct rpcrdma_xprt *new_xprt;
270
271 if (args->addrlen > sizeof(xprt->addr)) {
272 dprintk("RPC: %s: address too large\n", __func__);
273 return ERR_PTR(-EBADF);
274 }
275
276 xprt = xprt_alloc(args->net, sizeof(*new_xprt),
277 RPCRDMA_MAX_BC_REQUESTS,
278 RPCRDMA_MAX_BC_REQUESTS);
279 if (!xprt) {
280 dprintk("RPC: %s: couldn't allocate rpc_xprt\n",
281 __func__);
282 return ERR_PTR(-ENOMEM);
283 }
284
285 xprt->timeout = &xprt_rdma_bc_timeout;
286 xprt_set_bound(xprt);
287 xprt_set_connected(xprt);
288 xprt->bind_timeout = 0;
289 xprt->reestablish_timeout = 0;
290 xprt->idle_timeout = 0;
291
292 xprt->prot = XPRT_TRANSPORT_BC_RDMA;
293 xprt->ops = &xprt_rdma_bc_procs;
294
295 memcpy(&xprt->addr, args->dstaddr, args->addrlen);
296 xprt->addrlen = args->addrlen;
297 xprt_rdma_format_addresses(xprt, (struct sockaddr *)&xprt->addr);
298 xprt->resvport = 0;
299
300 xprt->max_payload = xprt_rdma_max_inline_read;
301
302 new_xprt = rpcx_to_rdmax(xprt);
303 new_xprt->rx_buf.rb_bc_max_requests = xprt->max_reqs;
304
305 xprt_get(xprt);
306 args->bc_xprt->xpt_bc_xprt = xprt;
307 xprt->bc_xprt = args->bc_xprt;
308
309 /* Final put for backchannel xprt is in __svc_rdma_free */
310 xprt_get(xprt);
311 return xprt;
312 }
313
314 struct xprt_class xprt_rdma_bc = {
315 .list = LIST_HEAD_INIT(xprt_rdma_bc.list),
316 .name = "rdma backchannel",
317 .owner = THIS_MODULE,
318 .ident = XPRT_TRANSPORT_BC_RDMA,
319 .setup = xprt_setup_rdma_bc,
320 };
321