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