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  * @xprt: controlling backchannel transport
20  * @rdma_resp: pointer to incoming transport header
21  * @rcvbuf: XDR buffer into which to decode the reply
22  *
23  * Returns:
24  *	%0 if @rcvbuf is filled in, xprt_complete_rqst called,
25  *	%-EAGAIN if server should call ->recvfrom again.
26  */
27 int svc_rdma_handle_bc_reply(struct rpc_xprt *xprt, __be32 *rdma_resp,
28 			     struct xdr_buf *rcvbuf)
29 {
30 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
31 	struct kvec *dst, *src = &rcvbuf->head[0];
32 	struct rpc_rqst *req;
33 	u32 credits;
34 	size_t len;
35 	__be32 xid;
36 	__be32 *p;
37 	int ret;
38 
39 	p = (__be32 *)src->iov_base;
40 	len = src->iov_len;
41 	xid = *rdma_resp;
42 
43 #ifdef SVCRDMA_BACKCHANNEL_DEBUG
44 	pr_info("%s: xid=%08x, length=%zu\n",
45 		__func__, be32_to_cpu(xid), len);
46 	pr_info("%s: RPC/RDMA: %*ph\n",
47 		__func__, (int)RPCRDMA_HDRLEN_MIN, rdma_resp);
48 	pr_info("%s:      RPC: %*ph\n",
49 		__func__, (int)len, p);
50 #endif
51 
52 	ret = -EAGAIN;
53 	if (src->iov_len < 24)
54 		goto out_shortreply;
55 
56 	spin_lock(&xprt->queue_lock);
57 	req = xprt_lookup_rqst(xprt, xid);
58 	if (!req)
59 		goto out_notfound;
60 
61 	dst = &req->rq_private_buf.head[0];
62 	memcpy(&req->rq_private_buf, &req->rq_rcv_buf, sizeof(struct xdr_buf));
63 	if (dst->iov_len < len)
64 		goto out_unlock;
65 	memcpy(dst->iov_base, p, len);
66 	xprt_pin_rqst(req);
67 	spin_unlock(&xprt->queue_lock);
68 
69 	credits = be32_to_cpup(rdma_resp + 2);
70 	if (credits == 0)
71 		credits = 1;	/* don't deadlock */
72 	else if (credits > r_xprt->rx_buf.rb_bc_max_requests)
73 		credits = r_xprt->rx_buf.rb_bc_max_requests;
74 
75 	spin_lock(&xprt->transport_lock);
76 	xprt->cwnd = credits << RPC_CWNDSHIFT;
77 	spin_unlock(&xprt->transport_lock);
78 
79 	spin_lock(&xprt->queue_lock);
80 	ret = 0;
81 	xprt_complete_rqst(req->rq_task, rcvbuf->len);
82 	xprt_unpin_rqst(req);
83 	rcvbuf->len = 0;
84 
85 out_unlock:
86 	spin_unlock(&xprt->queue_lock);
87 out:
88 	return ret;
89 
90 out_shortreply:
91 	dprintk("svcrdma: short bc reply: xprt=%p, len=%zu\n",
92 		xprt, src->iov_len);
93 	goto out;
94 
95 out_notfound:
96 	dprintk("svcrdma: unrecognized bc reply: xprt=%p, xid=%08x\n",
97 		xprt, be32_to_cpu(xid));
98 	goto out_unlock;
99 }
100 
101 /* Send a backwards direction RPC call.
102  *
103  * Caller holds the connection's mutex and has already marshaled
104  * the RPC/RDMA request.
105  *
106  * This is similar to svc_rdma_send_reply_msg, but takes a struct
107  * rpc_rqst instead, does not support chunks, and avoids blocking
108  * memory allocation.
109  *
110  * XXX: There is still an opportunity to block in svc_rdma_send()
111  * if there are no SQ entries to post the Send. This may occur if
112  * the adapter has a small maximum SQ depth.
113  */
114 static int svc_rdma_bc_sendto(struct svcxprt_rdma *rdma,
115 			      struct rpc_rqst *rqst,
116 			      struct svc_rdma_send_ctxt *ctxt)
117 {
118 	int ret;
119 
120 	ret = svc_rdma_map_reply_msg(rdma, ctxt, &rqst->rq_snd_buf, NULL);
121 	if (ret < 0)
122 		return -EIO;
123 
124 	/* Bump page refcnt so Send completion doesn't release
125 	 * the rq_buffer before all retransmits are complete.
126 	 */
127 	get_page(virt_to_page(rqst->rq_buffer));
128 	ctxt->sc_send_wr.opcode = IB_WR_SEND;
129 	return svc_rdma_send(rdma, &ctxt->sc_send_wr);
130 }
131 
132 /* Server-side transport endpoint wants a whole page for its send
133  * buffer. The client RPC code constructs the RPC header in this
134  * buffer before it invokes ->send_request.
135  */
136 static int
137 xprt_rdma_bc_allocate(struct rpc_task *task)
138 {
139 	struct rpc_rqst *rqst = task->tk_rqstp;
140 	size_t size = rqst->rq_callsize;
141 	struct page *page;
142 
143 	if (size > PAGE_SIZE) {
144 		WARN_ONCE(1, "svcrdma: large bc buffer request (size %zu)\n",
145 			  size);
146 		return -EINVAL;
147 	}
148 
149 	page = alloc_page(RPCRDMA_DEF_GFP);
150 	if (!page)
151 		return -ENOMEM;
152 	rqst->rq_buffer = page_address(page);
153 
154 	rqst->rq_rbuffer = kmalloc(rqst->rq_rcvsize, RPCRDMA_DEF_GFP);
155 	if (!rqst->rq_rbuffer) {
156 		put_page(page);
157 		return -ENOMEM;
158 	}
159 	return 0;
160 }
161 
162 static void
163 xprt_rdma_bc_free(struct rpc_task *task)
164 {
165 	struct rpc_rqst *rqst = task->tk_rqstp;
166 
167 	put_page(virt_to_page(rqst->rq_buffer));
168 	kfree(rqst->rq_rbuffer);
169 }
170 
171 static int
172 rpcrdma_bc_send_request(struct svcxprt_rdma *rdma, struct rpc_rqst *rqst)
173 {
174 	struct rpc_xprt *xprt = rqst->rq_xprt;
175 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
176 	struct svc_rdma_send_ctxt *ctxt;
177 	__be32 *p;
178 	int rc;
179 
180 	ctxt = svc_rdma_send_ctxt_get(rdma);
181 	if (!ctxt)
182 		goto drop_connection;
183 
184 	p = ctxt->sc_xprt_buf;
185 	*p++ = rqst->rq_xid;
186 	*p++ = rpcrdma_version;
187 	*p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_max_requests);
188 	*p++ = rdma_msg;
189 	*p++ = xdr_zero;
190 	*p++ = xdr_zero;
191 	*p   = xdr_zero;
192 	svc_rdma_sync_reply_hdr(rdma, ctxt, RPCRDMA_HDRLEN_MIN);
193 
194 #ifdef SVCRDMA_BACKCHANNEL_DEBUG
195 	pr_info("%s: %*ph\n", __func__, 64, rqst->rq_buffer);
196 #endif
197 
198 	rqst->rq_xtime = ktime_get();
199 	rc = svc_rdma_bc_sendto(rdma, rqst, ctxt);
200 	if (rc) {
201 		svc_rdma_send_ctxt_put(rdma, ctxt);
202 		goto drop_connection;
203 	}
204 	return 0;
205 
206 drop_connection:
207 	dprintk("svcrdma: failed to send bc call\n");
208 	return -ENOTCONN;
209 }
210 
211 /* Send an RPC call on the passive end of a transport
212  * connection.
213  */
214 static int
215 xprt_rdma_bc_send_request(struct rpc_rqst *rqst)
216 {
217 	struct svc_xprt *sxprt = rqst->rq_xprt->bc_xprt;
218 	struct svcxprt_rdma *rdma;
219 	int ret;
220 
221 	dprintk("svcrdma: sending bc call with xid: %08x\n",
222 		be32_to_cpu(rqst->rq_xid));
223 
224 	mutex_lock(&sxprt->xpt_mutex);
225 
226 	ret = -ENOTCONN;
227 	rdma = container_of(sxprt, struct svcxprt_rdma, sc_xprt);
228 	if (!test_bit(XPT_DEAD, &sxprt->xpt_flags)) {
229 		ret = rpcrdma_bc_send_request(rdma, rqst);
230 		if (ret == -ENOTCONN)
231 			svc_close_xprt(sxprt);
232 	}
233 
234 	mutex_unlock(&sxprt->xpt_mutex);
235 
236 	if (ret < 0)
237 		return ret;
238 	return 0;
239 }
240 
241 static void
242 xprt_rdma_bc_close(struct rpc_xprt *xprt)
243 {
244 	dprintk("svcrdma: %s: xprt %p\n", __func__, xprt);
245 	xprt->cwnd = RPC_CWNDSHIFT;
246 }
247 
248 static void
249 xprt_rdma_bc_put(struct rpc_xprt *xprt)
250 {
251 	dprintk("svcrdma: %s: xprt %p\n", __func__, xprt);
252 
253 	xprt_free(xprt);
254 }
255 
256 static const struct rpc_xprt_ops xprt_rdma_bc_procs = {
257 	.reserve_xprt		= xprt_reserve_xprt_cong,
258 	.release_xprt		= xprt_release_xprt_cong,
259 	.alloc_slot		= xprt_alloc_slot,
260 	.free_slot		= xprt_free_slot,
261 	.release_request	= xprt_release_rqst_cong,
262 	.buf_alloc		= xprt_rdma_bc_allocate,
263 	.buf_free		= xprt_rdma_bc_free,
264 	.send_request		= xprt_rdma_bc_send_request,
265 	.wait_for_reply_request	= xprt_wait_for_reply_request_def,
266 	.close			= xprt_rdma_bc_close,
267 	.destroy		= xprt_rdma_bc_put,
268 	.print_stats		= xprt_rdma_print_stats
269 };
270 
271 static const struct rpc_timeout xprt_rdma_bc_timeout = {
272 	.to_initval = 60 * HZ,
273 	.to_maxval = 60 * HZ,
274 };
275 
276 /* It shouldn't matter if the number of backchannel session slots
277  * doesn't match the number of RPC/RDMA credits. That just means
278  * one or the other will have extra slots that aren't used.
279  */
280 static struct rpc_xprt *
281 xprt_setup_rdma_bc(struct xprt_create *args)
282 {
283 	struct rpc_xprt *xprt;
284 	struct rpcrdma_xprt *new_xprt;
285 
286 	if (args->addrlen > sizeof(xprt->addr)) {
287 		dprintk("RPC:       %s: address too large\n", __func__);
288 		return ERR_PTR(-EBADF);
289 	}
290 
291 	xprt = xprt_alloc(args->net, sizeof(*new_xprt),
292 			  RPCRDMA_MAX_BC_REQUESTS,
293 			  RPCRDMA_MAX_BC_REQUESTS);
294 	if (!xprt) {
295 		dprintk("RPC:       %s: couldn't allocate rpc_xprt\n",
296 			__func__);
297 		return ERR_PTR(-ENOMEM);
298 	}
299 
300 	xprt->timeout = &xprt_rdma_bc_timeout;
301 	xprt_set_bound(xprt);
302 	xprt_set_connected(xprt);
303 	xprt->bind_timeout = RPCRDMA_BIND_TO;
304 	xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
305 	xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
306 
307 	xprt->prot = XPRT_TRANSPORT_BC_RDMA;
308 	xprt->ops = &xprt_rdma_bc_procs;
309 
310 	memcpy(&xprt->addr, args->dstaddr, args->addrlen);
311 	xprt->addrlen = args->addrlen;
312 	xprt_rdma_format_addresses(xprt, (struct sockaddr *)&xprt->addr);
313 	xprt->resvport = 0;
314 
315 	xprt->max_payload = xprt_rdma_max_inline_read;
316 
317 	new_xprt = rpcx_to_rdmax(xprt);
318 	new_xprt->rx_buf.rb_bc_max_requests = xprt->max_reqs;
319 
320 	xprt_get(xprt);
321 	args->bc_xprt->xpt_bc_xprt = xprt;
322 	xprt->bc_xprt = args->bc_xprt;
323 
324 	/* Final put for backchannel xprt is in __svc_rdma_free */
325 	xprt_get(xprt);
326 	return xprt;
327 }
328 
329 struct xprt_class xprt_rdma_bc = {
330 	.list			= LIST_HEAD_INIT(xprt_rdma_bc.list),
331 	.name			= "rdma backchannel",
332 	.owner			= THIS_MODULE,
333 	.ident			= XPRT_TRANSPORT_BC_RDMA,
334 	.setup			= xprt_setup_rdma_bc,
335 };
336