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/module.h>
9 
10 #include <linux/sunrpc/svc_rdma.h>
11 
12 #include "xprt_rdma.h"
13 #include <trace/events/rpcrdma.h>
14 
15 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
16 
17 #undef SVCRDMA_BACKCHANNEL_DEBUG
18 
19 /**
20  * svc_rdma_handle_bc_reply - Process incoming backchannel reply
21  * @xprt: controlling backchannel transport
22  * @rdma_resp: pointer to incoming transport header
23  * @rcvbuf: XDR buffer into which to decode the reply
24  *
25  * Returns:
26  *	%0 if @rcvbuf is filled in, xprt_complete_rqst called,
27  *	%-EAGAIN if server should call ->recvfrom again.
28  */
29 int svc_rdma_handle_bc_reply(struct rpc_xprt *xprt, __be32 *rdma_resp,
30 			     struct xdr_buf *rcvbuf)
31 {
32 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
33 	struct kvec *dst, *src = &rcvbuf->head[0];
34 	struct rpc_rqst *req;
35 	unsigned long cwnd;
36 	u32 credits;
37 	size_t len;
38 	__be32 xid;
39 	__be32 *p;
40 	int ret;
41 
42 	p = (__be32 *)src->iov_base;
43 	len = src->iov_len;
44 	xid = *rdma_resp;
45 
46 #ifdef SVCRDMA_BACKCHANNEL_DEBUG
47 	pr_info("%s: xid=%08x, length=%zu\n",
48 		__func__, be32_to_cpu(xid), len);
49 	pr_info("%s: RPC/RDMA: %*ph\n",
50 		__func__, (int)RPCRDMA_HDRLEN_MIN, rdma_resp);
51 	pr_info("%s:      RPC: %*ph\n",
52 		__func__, (int)len, p);
53 #endif
54 
55 	ret = -EAGAIN;
56 	if (src->iov_len < 24)
57 		goto out_shortreply;
58 
59 	spin_lock(&xprt->recv_lock);
60 	req = xprt_lookup_rqst(xprt, xid);
61 	if (!req)
62 		goto out_notfound;
63 
64 	dst = &req->rq_private_buf.head[0];
65 	memcpy(&req->rq_private_buf, &req->rq_rcv_buf, sizeof(struct xdr_buf));
66 	if (dst->iov_len < len)
67 		goto out_unlock;
68 	memcpy(dst->iov_base, p, len);
69 
70 	credits = be32_to_cpup(rdma_resp + 2);
71 	if (credits == 0)
72 		credits = 1;	/* don't deadlock */
73 	else if (credits > r_xprt->rx_buf.rb_bc_max_requests)
74 		credits = r_xprt->rx_buf.rb_bc_max_requests;
75 
76 	spin_lock_bh(&xprt->transport_lock);
77 	cwnd = xprt->cwnd;
78 	xprt->cwnd = credits << RPC_CWNDSHIFT;
79 	if (xprt->cwnd > cwnd)
80 		xprt_release_rqst_cong(req->rq_task);
81 	spin_unlock_bh(&xprt->transport_lock);
82 
83 
84 	ret = 0;
85 	xprt_complete_rqst(req->rq_task, rcvbuf->len);
86 	rcvbuf->len = 0;
87 
88 out_unlock:
89 	spin_unlock(&xprt->recv_lock);
90 out:
91 	return ret;
92 
93 out_shortreply:
94 	dprintk("svcrdma: short bc reply: xprt=%p, len=%zu\n",
95 		xprt, src->iov_len);
96 	goto out;
97 
98 out_notfound:
99 	dprintk("svcrdma: unrecognized bc reply: xprt=%p, xid=%08x\n",
100 		xprt, be32_to_cpu(xid));
101 	goto out_unlock;
102 }
103 
104 /* Send a backwards direction RPC call.
105  *
106  * Caller holds the connection's mutex and has already marshaled
107  * the RPC/RDMA request.
108  *
109  * This is similar to svc_rdma_send_reply_msg, but takes a struct
110  * rpc_rqst instead, does not support chunks, and avoids blocking
111  * memory allocation.
112  *
113  * XXX: There is still an opportunity to block in svc_rdma_send()
114  * if there are no SQ entries to post the Send. This may occur if
115  * the adapter has a small maximum SQ depth.
116  */
117 static int svc_rdma_bc_sendto(struct svcxprt_rdma *rdma,
118 			      struct rpc_rqst *rqst,
119 			      struct svc_rdma_send_ctxt *ctxt)
120 {
121 	int ret;
122 
123 	ret = svc_rdma_map_reply_msg(rdma, ctxt, &rqst->rq_snd_buf, NULL);
124 	if (ret < 0)
125 		return -EIO;
126 
127 	/* Bump page refcnt so Send completion doesn't release
128 	 * the rq_buffer before all retransmits are complete.
129 	 */
130 	get_page(virt_to_page(rqst->rq_buffer));
131 	ctxt->sc_send_wr.opcode = IB_WR_SEND;
132 	return svc_rdma_send(rdma, &ctxt->sc_send_wr);
133 }
134 
135 /* Server-side transport endpoint wants a whole page for its send
136  * buffer. The client RPC code constructs the RPC header in this
137  * buffer before it invokes ->send_request.
138  */
139 static int
140 xprt_rdma_bc_allocate(struct rpc_task *task)
141 {
142 	struct rpc_rqst *rqst = task->tk_rqstp;
143 	size_t size = rqst->rq_callsize;
144 	struct page *page;
145 
146 	if (size > PAGE_SIZE) {
147 		WARN_ONCE(1, "svcrdma: large bc buffer request (size %zu)\n",
148 			  size);
149 		return -EINVAL;
150 	}
151 
152 	page = alloc_page(RPCRDMA_DEF_GFP);
153 	if (!page)
154 		return -ENOMEM;
155 	rqst->rq_buffer = page_address(page);
156 
157 	rqst->rq_rbuffer = kmalloc(rqst->rq_rcvsize, RPCRDMA_DEF_GFP);
158 	if (!rqst->rq_rbuffer) {
159 		put_page(page);
160 		return -ENOMEM;
161 	}
162 	return 0;
163 }
164 
165 static void
166 xprt_rdma_bc_free(struct rpc_task *task)
167 {
168 	struct rpc_rqst *rqst = task->tk_rqstp;
169 
170 	put_page(virt_to_page(rqst->rq_buffer));
171 	kfree(rqst->rq_rbuffer);
172 }
173 
174 static int
175 rpcrdma_bc_send_request(struct svcxprt_rdma *rdma, struct rpc_rqst *rqst)
176 {
177 	struct rpc_xprt *xprt = rqst->rq_xprt;
178 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
179 	struct svc_rdma_send_ctxt *ctxt;
180 	__be32 *p;
181 	int rc;
182 
183 	ctxt = svc_rdma_send_ctxt_get(rdma);
184 	if (!ctxt)
185 		goto drop_connection;
186 
187 	p = ctxt->sc_xprt_buf;
188 	*p++ = rqst->rq_xid;
189 	*p++ = rpcrdma_version;
190 	*p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_max_requests);
191 	*p++ = rdma_msg;
192 	*p++ = xdr_zero;
193 	*p++ = xdr_zero;
194 	*p   = xdr_zero;
195 	svc_rdma_sync_reply_hdr(rdma, ctxt, RPCRDMA_HDRLEN_MIN);
196 
197 #ifdef SVCRDMA_BACKCHANNEL_DEBUG
198 	pr_info("%s: %*ph\n", __func__, 64, rqst->rq_buffer);
199 #endif
200 
201 	rc = svc_rdma_bc_sendto(rdma, rqst, ctxt);
202 	if (rc) {
203 		svc_rdma_send_ctxt_put(rdma, ctxt);
204 		goto drop_connection;
205 	}
206 	return rc;
207 
208 drop_connection:
209 	dprintk("svcrdma: failed to send bc call\n");
210 	xprt_disconnect_done(xprt);
211 	return -ENOTCONN;
212 }
213 
214 /* Send an RPC call on the passive end of a transport
215  * connection.
216  */
217 static int
218 xprt_rdma_bc_send_request(struct rpc_task *task)
219 {
220 	struct rpc_rqst *rqst = task->tk_rqstp;
221 	struct svc_xprt *sxprt = rqst->rq_xprt->bc_xprt;
222 	struct svcxprt_rdma *rdma;
223 	int ret;
224 
225 	dprintk("svcrdma: sending bc call with xid: %08x\n",
226 		be32_to_cpu(rqst->rq_xid));
227 
228 	if (!mutex_trylock(&sxprt->xpt_mutex)) {
229 		rpc_sleep_on(&sxprt->xpt_bc_pending, task, NULL);
230 		if (!mutex_trylock(&sxprt->xpt_mutex))
231 			return -EAGAIN;
232 		rpc_wake_up_queued_task(&sxprt->xpt_bc_pending, task);
233 	}
234 
235 	ret = -ENOTCONN;
236 	rdma = container_of(sxprt, struct svcxprt_rdma, sc_xprt);
237 	if (!test_bit(XPT_DEAD, &sxprt->xpt_flags))
238 		ret = rpcrdma_bc_send_request(rdma, rqst);
239 
240 	mutex_unlock(&sxprt->xpt_mutex);
241 
242 	if (ret < 0)
243 		return ret;
244 	return 0;
245 }
246 
247 static void
248 xprt_rdma_bc_close(struct rpc_xprt *xprt)
249 {
250 	dprintk("svcrdma: %s: xprt %p\n", __func__, xprt);
251 }
252 
253 static void
254 xprt_rdma_bc_put(struct rpc_xprt *xprt)
255 {
256 	dprintk("svcrdma: %s: xprt %p\n", __func__, xprt);
257 
258 	xprt_free(xprt);
259 	module_put(THIS_MODULE);
260 }
261 
262 static const struct rpc_xprt_ops xprt_rdma_bc_procs = {
263 	.reserve_xprt		= xprt_reserve_xprt_cong,
264 	.release_xprt		= xprt_release_xprt_cong,
265 	.alloc_slot		= xprt_alloc_slot,
266 	.free_slot		= xprt_free_slot,
267 	.release_request	= xprt_release_rqst_cong,
268 	.buf_alloc		= xprt_rdma_bc_allocate,
269 	.buf_free		= xprt_rdma_bc_free,
270 	.send_request		= xprt_rdma_bc_send_request,
271 	.set_retrans_timeout	= xprt_set_retrans_timeout_def,
272 	.close			= xprt_rdma_bc_close,
273 	.destroy		= xprt_rdma_bc_put,
274 	.print_stats		= xprt_rdma_print_stats
275 };
276 
277 static const struct rpc_timeout xprt_rdma_bc_timeout = {
278 	.to_initval = 60 * HZ,
279 	.to_maxval = 60 * HZ,
280 };
281 
282 /* It shouldn't matter if the number of backchannel session slots
283  * doesn't match the number of RPC/RDMA credits. That just means
284  * one or the other will have extra slots that aren't used.
285  */
286 static struct rpc_xprt *
287 xprt_setup_rdma_bc(struct xprt_create *args)
288 {
289 	struct rpc_xprt *xprt;
290 	struct rpcrdma_xprt *new_xprt;
291 
292 	if (args->addrlen > sizeof(xprt->addr)) {
293 		dprintk("RPC:       %s: address too large\n", __func__);
294 		return ERR_PTR(-EBADF);
295 	}
296 
297 	xprt = xprt_alloc(args->net, sizeof(*new_xprt),
298 			  RPCRDMA_MAX_BC_REQUESTS,
299 			  RPCRDMA_MAX_BC_REQUESTS);
300 	if (!xprt) {
301 		dprintk("RPC:       %s: couldn't allocate rpc_xprt\n",
302 			__func__);
303 		return ERR_PTR(-ENOMEM);
304 	}
305 
306 	xprt->timeout = &xprt_rdma_bc_timeout;
307 	xprt_set_bound(xprt);
308 	xprt_set_connected(xprt);
309 	xprt->bind_timeout = RPCRDMA_BIND_TO;
310 	xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
311 	xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
312 
313 	xprt->prot = XPRT_TRANSPORT_BC_RDMA;
314 	xprt->tsh_size = 0;
315 	xprt->ops = &xprt_rdma_bc_procs;
316 
317 	memcpy(&xprt->addr, args->dstaddr, args->addrlen);
318 	xprt->addrlen = args->addrlen;
319 	xprt_rdma_format_addresses(xprt, (struct sockaddr *)&xprt->addr);
320 	xprt->resvport = 0;
321 
322 	xprt->max_payload = xprt_rdma_max_inline_read;
323 
324 	new_xprt = rpcx_to_rdmax(xprt);
325 	new_xprt->rx_buf.rb_bc_max_requests = xprt->max_reqs;
326 
327 	xprt_get(xprt);
328 	args->bc_xprt->xpt_bc_xprt = xprt;
329 	xprt->bc_xprt = args->bc_xprt;
330 
331 	if (!try_module_get(THIS_MODULE))
332 		goto out_fail;
333 
334 	/* Final put for backchannel xprt is in __svc_rdma_free */
335 	xprt_get(xprt);
336 	return xprt;
337 
338 out_fail:
339 	xprt_rdma_free_addresses(xprt);
340 	args->bc_xprt->xpt_bc_xprt = NULL;
341 	args->bc_xprt->xpt_bc_xps = NULL;
342 	xprt_put(xprt);
343 	xprt_free(xprt);
344 	return ERR_PTR(-EINVAL);
345 }
346 
347 struct xprt_class xprt_rdma_bc = {
348 	.list			= LIST_HEAD_INIT(xprt_rdma_bc.list),
349 	.name			= "rdma backchannel",
350 	.owner			= THIS_MODULE,
351 	.ident			= XPRT_TRANSPORT_BC_RDMA,
352 	.setup			= xprt_setup_rdma_bc,
353 };
354