xref: /openbmc/linux/net/sunrpc/xprtrdma/transport.c (revision abcda807)
1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
3  * Copyright (c) 2014-2017 Oracle.  All rights reserved.
4  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the BSD-type
10  * license below:
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  *
16  *      Redistributions of source code must retain the above copyright
17  *      notice, this list of conditions and the following disclaimer.
18  *
19  *      Redistributions in binary form must reproduce the above
20  *      copyright notice, this list of conditions and the following
21  *      disclaimer in the documentation and/or other materials provided
22  *      with the distribution.
23  *
24  *      Neither the name of the Network Appliance, Inc. nor the names of
25  *      its contributors may be used to endorse or promote products
26  *      derived from this software without specific prior written
27  *      permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40  */
41 
42 /*
43  * transport.c
44  *
45  * This file contains the top-level implementation of an RPC RDMA
46  * transport.
47  *
48  * Naming convention: functions beginning with xprt_ are part of the
49  * transport switch. All others are RPC RDMA internal.
50  */
51 
52 #include <linux/module.h>
53 #include <linux/slab.h>
54 #include <linux/seq_file.h>
55 #include <linux/smp.h>
56 
57 #include <linux/sunrpc/addr.h>
58 #include <linux/sunrpc/svc_rdma.h>
59 
60 #include "xprt_rdma.h"
61 #include <trace/events/rpcrdma.h>
62 
63 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
64 # define RPCDBG_FACILITY	RPCDBG_TRANS
65 #endif
66 
67 /*
68  * tunables
69  */
70 
71 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
72 unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
73 unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
74 unsigned int xprt_rdma_memreg_strategy		= RPCRDMA_FRWR;
75 int xprt_rdma_pad_optimize;
76 
77 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
78 
79 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
80 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
81 static unsigned int min_inline_size = RPCRDMA_MIN_INLINE;
82 static unsigned int max_inline_size = RPCRDMA_MAX_INLINE;
83 static unsigned int max_padding = PAGE_SIZE;
84 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
85 static unsigned int max_memreg = RPCRDMA_LAST - 1;
86 static unsigned int dummy;
87 
88 static struct ctl_table_header *sunrpc_table_header;
89 
90 static struct ctl_table xr_tunables_table[] = {
91 	{
92 		.procname	= "rdma_slot_table_entries",
93 		.data		= &xprt_rdma_slot_table_entries,
94 		.maxlen		= sizeof(unsigned int),
95 		.mode		= 0644,
96 		.proc_handler	= proc_dointvec_minmax,
97 		.extra1		= &min_slot_table_size,
98 		.extra2		= &max_slot_table_size
99 	},
100 	{
101 		.procname	= "rdma_max_inline_read",
102 		.data		= &xprt_rdma_max_inline_read,
103 		.maxlen		= sizeof(unsigned int),
104 		.mode		= 0644,
105 		.proc_handler	= proc_dointvec_minmax,
106 		.extra1		= &min_inline_size,
107 		.extra2		= &max_inline_size,
108 	},
109 	{
110 		.procname	= "rdma_max_inline_write",
111 		.data		= &xprt_rdma_max_inline_write,
112 		.maxlen		= sizeof(unsigned int),
113 		.mode		= 0644,
114 		.proc_handler	= proc_dointvec_minmax,
115 		.extra1		= &min_inline_size,
116 		.extra2		= &max_inline_size,
117 	},
118 	{
119 		.procname	= "rdma_inline_write_padding",
120 		.data		= &dummy,
121 		.maxlen		= sizeof(unsigned int),
122 		.mode		= 0644,
123 		.proc_handler	= proc_dointvec_minmax,
124 		.extra1		= SYSCTL_ZERO,
125 		.extra2		= &max_padding,
126 	},
127 	{
128 		.procname	= "rdma_memreg_strategy",
129 		.data		= &xprt_rdma_memreg_strategy,
130 		.maxlen		= sizeof(unsigned int),
131 		.mode		= 0644,
132 		.proc_handler	= proc_dointvec_minmax,
133 		.extra1		= &min_memreg,
134 		.extra2		= &max_memreg,
135 	},
136 	{
137 		.procname	= "rdma_pad_optimize",
138 		.data		= &xprt_rdma_pad_optimize,
139 		.maxlen		= sizeof(unsigned int),
140 		.mode		= 0644,
141 		.proc_handler	= proc_dointvec,
142 	},
143 	{ },
144 };
145 
146 static struct ctl_table sunrpc_table[] = {
147 	{
148 		.procname	= "sunrpc",
149 		.mode		= 0555,
150 		.child		= xr_tunables_table
151 	},
152 	{ },
153 };
154 
155 #endif
156 
157 static const struct rpc_xprt_ops xprt_rdma_procs;
158 
159 static void
160 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
161 {
162 	struct sockaddr_in *sin = (struct sockaddr_in *)sap;
163 	char buf[20];
164 
165 	snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
166 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
167 
168 	xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
169 }
170 
171 static void
172 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
173 {
174 	struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
175 	char buf[40];
176 
177 	snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
178 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
179 
180 	xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
181 }
182 
183 void
184 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
185 {
186 	char buf[128];
187 
188 	switch (sap->sa_family) {
189 	case AF_INET:
190 		xprt_rdma_format_addresses4(xprt, sap);
191 		break;
192 	case AF_INET6:
193 		xprt_rdma_format_addresses6(xprt, sap);
194 		break;
195 	default:
196 		pr_err("rpcrdma: Unrecognized address family\n");
197 		return;
198 	}
199 
200 	(void)rpc_ntop(sap, buf, sizeof(buf));
201 	xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
202 
203 	snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
204 	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
205 
206 	snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
207 	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
208 
209 	xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
210 }
211 
212 void
213 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
214 {
215 	unsigned int i;
216 
217 	for (i = 0; i < RPC_DISPLAY_MAX; i++)
218 		switch (i) {
219 		case RPC_DISPLAY_PROTO:
220 		case RPC_DISPLAY_NETID:
221 			continue;
222 		default:
223 			kfree(xprt->address_strings[i]);
224 		}
225 }
226 
227 /**
228  * xprt_rdma_connect_worker - establish connection in the background
229  * @work: worker thread context
230  *
231  * Requester holds the xprt's send lock to prevent activity on this
232  * transport while a fresh connection is being established. RPC tasks
233  * sleep on the xprt's pending queue waiting for connect to complete.
234  */
235 static void
236 xprt_rdma_connect_worker(struct work_struct *work)
237 {
238 	struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
239 						   rx_connect_worker.work);
240 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
241 	int rc;
242 
243 	rc = rpcrdma_xprt_connect(r_xprt);
244 	xprt_clear_connecting(xprt);
245 	if (!rc) {
246 		xprt->connect_cookie++;
247 		xprt->stat.connect_count++;
248 		xprt->stat.connect_time += (long)jiffies -
249 					   xprt->stat.connect_start;
250 		xprt_set_connected(xprt);
251 		rc = -EAGAIN;
252 	} else {
253 		/* Force a call to xprt_rdma_close to clean up */
254 		spin_lock(&xprt->transport_lock);
255 		set_bit(XPRT_CLOSE_WAIT, &xprt->state);
256 		spin_unlock(&xprt->transport_lock);
257 	}
258 	xprt_wake_pending_tasks(xprt, rc);
259 }
260 
261 /**
262  * xprt_rdma_inject_disconnect - inject a connection fault
263  * @xprt: transport context
264  *
265  * If @xprt is connected, disconnect it to simulate spurious connection
266  * loss.
267  */
268 static void
269 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
270 {
271 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
272 
273 	trace_xprtrdma_op_inject_dsc(r_xprt);
274 	rdma_disconnect(r_xprt->rx_ep->re_id);
275 }
276 
277 /**
278  * xprt_rdma_destroy - Full tear down of transport
279  * @xprt: doomed transport context
280  *
281  * Caller guarantees there will be no more calls to us with
282  * this @xprt.
283  */
284 static void
285 xprt_rdma_destroy(struct rpc_xprt *xprt)
286 {
287 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
288 
289 	cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
290 
291 	rpcrdma_xprt_disconnect(r_xprt);
292 	rpcrdma_buffer_destroy(&r_xprt->rx_buf);
293 
294 	xprt_rdma_free_addresses(xprt);
295 	xprt_free(xprt);
296 
297 	module_put(THIS_MODULE);
298 }
299 
300 /* 60 second timeout, no retries */
301 static const struct rpc_timeout xprt_rdma_default_timeout = {
302 	.to_initval = 60 * HZ,
303 	.to_maxval = 60 * HZ,
304 };
305 
306 /**
307  * xprt_setup_rdma - Set up transport to use RDMA
308  *
309  * @args: rpc transport arguments
310  */
311 static struct rpc_xprt *
312 xprt_setup_rdma(struct xprt_create *args)
313 {
314 	struct rpc_xprt *xprt;
315 	struct rpcrdma_xprt *new_xprt;
316 	struct sockaddr *sap;
317 	int rc;
318 
319 	if (args->addrlen > sizeof(xprt->addr))
320 		return ERR_PTR(-EBADF);
321 
322 	if (!try_module_get(THIS_MODULE))
323 		return ERR_PTR(-EIO);
324 
325 	xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt), 0,
326 			  xprt_rdma_slot_table_entries);
327 	if (!xprt) {
328 		module_put(THIS_MODULE);
329 		return ERR_PTR(-ENOMEM);
330 	}
331 
332 	xprt->timeout = &xprt_rdma_default_timeout;
333 	xprt->connect_timeout = xprt->timeout->to_initval;
334 	xprt->max_reconnect_timeout = xprt->timeout->to_maxval;
335 	xprt->bind_timeout = RPCRDMA_BIND_TO;
336 	xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
337 	xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
338 
339 	xprt->resvport = 0;		/* privileged port not needed */
340 	xprt->ops = &xprt_rdma_procs;
341 
342 	/*
343 	 * Set up RDMA-specific connect data.
344 	 */
345 	sap = args->dstaddr;
346 
347 	/* Ensure xprt->addr holds valid server TCP (not RDMA)
348 	 * address, for any side protocols which peek at it */
349 	xprt->prot = IPPROTO_TCP;
350 	xprt->addrlen = args->addrlen;
351 	memcpy(&xprt->addr, sap, xprt->addrlen);
352 
353 	if (rpc_get_port(sap))
354 		xprt_set_bound(xprt);
355 	xprt_rdma_format_addresses(xprt, sap);
356 
357 	new_xprt = rpcx_to_rdmax(xprt);
358 	rc = rpcrdma_buffer_create(new_xprt);
359 	if (rc) {
360 		xprt_rdma_free_addresses(xprt);
361 		xprt_free(xprt);
362 		module_put(THIS_MODULE);
363 		return ERR_PTR(rc);
364 	}
365 
366 	INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
367 			  xprt_rdma_connect_worker);
368 
369 	xprt->max_payload = RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT;
370 
371 	return xprt;
372 }
373 
374 /**
375  * xprt_rdma_close - close a transport connection
376  * @xprt: transport context
377  *
378  * Called during autoclose or device removal.
379  *
380  * Caller holds @xprt's send lock to prevent activity on this
381  * transport while the connection is torn down.
382  */
383 void xprt_rdma_close(struct rpc_xprt *xprt)
384 {
385 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
386 
387 	rpcrdma_xprt_disconnect(r_xprt);
388 
389 	xprt->reestablish_timeout = 0;
390 	++xprt->connect_cookie;
391 	xprt_disconnect_done(xprt);
392 }
393 
394 /**
395  * xprt_rdma_set_port - update server port with rpcbind result
396  * @xprt: controlling RPC transport
397  * @port: new port value
398  *
399  * Transport connect status is unchanged.
400  */
401 static void
402 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
403 {
404 	struct sockaddr *sap = (struct sockaddr *)&xprt->addr;
405 	char buf[8];
406 
407 	rpc_set_port(sap, port);
408 
409 	kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
410 	snprintf(buf, sizeof(buf), "%u", port);
411 	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
412 
413 	kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]);
414 	snprintf(buf, sizeof(buf), "%4hx", port);
415 	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
416 }
417 
418 /**
419  * xprt_rdma_timer - invoked when an RPC times out
420  * @xprt: controlling RPC transport
421  * @task: RPC task that timed out
422  *
423  * Invoked when the transport is still connected, but an RPC
424  * retransmit timeout occurs.
425  *
426  * Since RDMA connections don't have a keep-alive, forcibly
427  * disconnect and retry to connect. This drives full
428  * detection of the network path, and retransmissions of
429  * all pending RPCs.
430  */
431 static void
432 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
433 {
434 	xprt_force_disconnect(xprt);
435 }
436 
437 /**
438  * xprt_rdma_set_connect_timeout - set timeouts for establishing a connection
439  * @xprt: controlling transport instance
440  * @connect_timeout: reconnect timeout after client disconnects
441  * @reconnect_timeout: reconnect timeout after server disconnects
442  *
443  */
444 static void xprt_rdma_set_connect_timeout(struct rpc_xprt *xprt,
445 					  unsigned long connect_timeout,
446 					  unsigned long reconnect_timeout)
447 {
448 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
449 
450 	trace_xprtrdma_op_set_cto(r_xprt, connect_timeout, reconnect_timeout);
451 
452 	spin_lock(&xprt->transport_lock);
453 
454 	if (connect_timeout < xprt->connect_timeout) {
455 		struct rpc_timeout to;
456 		unsigned long initval;
457 
458 		to = *xprt->timeout;
459 		initval = connect_timeout;
460 		if (initval < RPCRDMA_INIT_REEST_TO << 1)
461 			initval = RPCRDMA_INIT_REEST_TO << 1;
462 		to.to_initval = initval;
463 		to.to_maxval = initval;
464 		r_xprt->rx_timeout = to;
465 		xprt->timeout = &r_xprt->rx_timeout;
466 		xprt->connect_timeout = connect_timeout;
467 	}
468 
469 	if (reconnect_timeout < xprt->max_reconnect_timeout)
470 		xprt->max_reconnect_timeout = reconnect_timeout;
471 
472 	spin_unlock(&xprt->transport_lock);
473 }
474 
475 /**
476  * xprt_rdma_connect - schedule an attempt to reconnect
477  * @xprt: transport state
478  * @task: RPC scheduler context (unused)
479  *
480  */
481 static void
482 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
483 {
484 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
485 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
486 	unsigned long delay;
487 
488 	delay = 0;
489 	if (ep && ep->re_connect_status != 0) {
490 		delay = xprt_reconnect_delay(xprt);
491 		xprt_reconnect_backoff(xprt, RPCRDMA_INIT_REEST_TO);
492 	}
493 	trace_xprtrdma_op_connect(r_xprt, delay);
494 	queue_delayed_work(xprtiod_workqueue, &r_xprt->rx_connect_worker,
495 			   delay);
496 }
497 
498 /**
499  * xprt_rdma_alloc_slot - allocate an rpc_rqst
500  * @xprt: controlling RPC transport
501  * @task: RPC task requesting a fresh rpc_rqst
502  *
503  * tk_status values:
504  *	%0 if task->tk_rqstp points to a fresh rpc_rqst
505  *	%-EAGAIN if no rpc_rqst is available; queued on backlog
506  */
507 static void
508 xprt_rdma_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
509 {
510 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
511 	struct rpcrdma_req *req;
512 
513 	req = rpcrdma_buffer_get(&r_xprt->rx_buf);
514 	if (!req)
515 		goto out_sleep;
516 	task->tk_rqstp = &req->rl_slot;
517 	task->tk_status = 0;
518 	return;
519 
520 out_sleep:
521 	set_bit(XPRT_CONGESTED, &xprt->state);
522 	rpc_sleep_on(&xprt->backlog, task, NULL);
523 	task->tk_status = -EAGAIN;
524 }
525 
526 /**
527  * xprt_rdma_free_slot - release an rpc_rqst
528  * @xprt: controlling RPC transport
529  * @rqst: rpc_rqst to release
530  *
531  */
532 static void
533 xprt_rdma_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *rqst)
534 {
535 	struct rpcrdma_xprt *r_xprt =
536 		container_of(xprt, struct rpcrdma_xprt, rx_xprt);
537 
538 	memset(rqst, 0, sizeof(*rqst));
539 	rpcrdma_buffer_put(&r_xprt->rx_buf, rpcr_to_rdmar(rqst));
540 	if (unlikely(!rpc_wake_up_next(&xprt->backlog)))
541 		clear_bit(XPRT_CONGESTED, &xprt->state);
542 }
543 
544 static bool rpcrdma_check_regbuf(struct rpcrdma_xprt *r_xprt,
545 				 struct rpcrdma_regbuf *rb, size_t size,
546 				 gfp_t flags)
547 {
548 	if (unlikely(rdmab_length(rb) < size)) {
549 		if (!rpcrdma_regbuf_realloc(rb, size, flags))
550 			return false;
551 		r_xprt->rx_stats.hardway_register_count += size;
552 	}
553 	return true;
554 }
555 
556 /**
557  * xprt_rdma_allocate - allocate transport resources for an RPC
558  * @task: RPC task
559  *
560  * Return values:
561  *        0:	Success; rq_buffer points to RPC buffer to use
562  *   ENOMEM:	Out of memory, call again later
563  *      EIO:	A permanent error occurred, do not retry
564  */
565 static int
566 xprt_rdma_allocate(struct rpc_task *task)
567 {
568 	struct rpc_rqst *rqst = task->tk_rqstp;
569 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
570 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
571 	gfp_t flags;
572 
573 	flags = RPCRDMA_DEF_GFP;
574 	if (RPC_IS_SWAPPER(task))
575 		flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
576 
577 	if (!rpcrdma_check_regbuf(r_xprt, req->rl_sendbuf, rqst->rq_callsize,
578 				  flags))
579 		goto out_fail;
580 	if (!rpcrdma_check_regbuf(r_xprt, req->rl_recvbuf, rqst->rq_rcvsize,
581 				  flags))
582 		goto out_fail;
583 
584 	rqst->rq_buffer = rdmab_data(req->rl_sendbuf);
585 	rqst->rq_rbuffer = rdmab_data(req->rl_recvbuf);
586 	return 0;
587 
588 out_fail:
589 	return -ENOMEM;
590 }
591 
592 /**
593  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
594  * @task: RPC task
595  *
596  * Caller guarantees rqst->rq_buffer is non-NULL.
597  */
598 static void
599 xprt_rdma_free(struct rpc_task *task)
600 {
601 	struct rpc_rqst *rqst = task->tk_rqstp;
602 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
603 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
604 
605 	if (!list_empty(&req->rl_registered))
606 		frwr_unmap_sync(r_xprt, req);
607 
608 	/* XXX: If the RPC is completing because of a signal and
609 	 * not because a reply was received, we ought to ensure
610 	 * that the Send completion has fired, so that memory
611 	 * involved with the Send is not still visible to the NIC.
612 	 */
613 }
614 
615 /**
616  * xprt_rdma_send_request - marshal and send an RPC request
617  * @rqst: RPC message in rq_snd_buf
618  *
619  * Caller holds the transport's write lock.
620  *
621  * Returns:
622  *	%0 if the RPC message has been sent
623  *	%-ENOTCONN if the caller should reconnect and call again
624  *	%-EAGAIN if the caller should call again
625  *	%-ENOBUFS if the caller should call again after a delay
626  *	%-EMSGSIZE if encoding ran out of buffer space. The request
627  *		was not sent. Do not try to send this message again.
628  *	%-EIO if an I/O error occurred. The request was not sent.
629  *		Do not try to send this message again.
630  */
631 static int
632 xprt_rdma_send_request(struct rpc_rqst *rqst)
633 {
634 	struct rpc_xprt *xprt = rqst->rq_xprt;
635 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
636 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
637 	int rc = 0;
638 
639 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
640 	if (unlikely(!rqst->rq_buffer))
641 		return xprt_rdma_bc_send_reply(rqst);
642 #endif	/* CONFIG_SUNRPC_BACKCHANNEL */
643 
644 	if (!xprt_connected(xprt))
645 		return -ENOTCONN;
646 
647 	if (!xprt_request_get_cong(xprt, rqst))
648 		return -EBADSLT;
649 
650 	rc = rpcrdma_marshal_req(r_xprt, rqst);
651 	if (rc < 0)
652 		goto failed_marshal;
653 
654 	/* Must suppress retransmit to maintain credits */
655 	if (rqst->rq_connect_cookie == xprt->connect_cookie)
656 		goto drop_connection;
657 	rqst->rq_xtime = ktime_get();
658 
659 	if (rpcrdma_post_sends(r_xprt, req))
660 		goto drop_connection;
661 
662 	rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
663 
664 	/* An RPC with no reply will throw off credit accounting,
665 	 * so drop the connection to reset the credit grant.
666 	 */
667 	if (!rpc_reply_expected(rqst->rq_task))
668 		goto drop_connection;
669 	return 0;
670 
671 failed_marshal:
672 	if (rc != -ENOTCONN)
673 		return rc;
674 drop_connection:
675 	xprt_rdma_close(xprt);
676 	return -ENOTCONN;
677 }
678 
679 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
680 {
681 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
682 	long idle_time = 0;
683 
684 	if (xprt_connected(xprt))
685 		idle_time = (long)(jiffies - xprt->last_used) / HZ;
686 
687 	seq_puts(seq, "\txprt:\trdma ");
688 	seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
689 		   0,	/* need a local port? */
690 		   xprt->stat.bind_count,
691 		   xprt->stat.connect_count,
692 		   xprt->stat.connect_time / HZ,
693 		   idle_time,
694 		   xprt->stat.sends,
695 		   xprt->stat.recvs,
696 		   xprt->stat.bad_xids,
697 		   xprt->stat.req_u,
698 		   xprt->stat.bklog_u);
699 	seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
700 		   r_xprt->rx_stats.read_chunk_count,
701 		   r_xprt->rx_stats.write_chunk_count,
702 		   r_xprt->rx_stats.reply_chunk_count,
703 		   r_xprt->rx_stats.total_rdma_request,
704 		   r_xprt->rx_stats.total_rdma_reply,
705 		   r_xprt->rx_stats.pullup_copy_count,
706 		   r_xprt->rx_stats.fixup_copy_count,
707 		   r_xprt->rx_stats.hardway_register_count,
708 		   r_xprt->rx_stats.failed_marshal_count,
709 		   r_xprt->rx_stats.bad_reply_count,
710 		   r_xprt->rx_stats.nomsg_call_count);
711 	seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n",
712 		   r_xprt->rx_stats.mrs_recycled,
713 		   r_xprt->rx_stats.mrs_orphaned,
714 		   r_xprt->rx_stats.mrs_allocated,
715 		   r_xprt->rx_stats.local_inv_needed,
716 		   r_xprt->rx_stats.empty_sendctx_q,
717 		   r_xprt->rx_stats.reply_waits_for_send);
718 }
719 
720 static int
721 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
722 {
723 	return 0;
724 }
725 
726 static void
727 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
728 {
729 }
730 
731 /*
732  * Plumbing for rpc transport switch and kernel module
733  */
734 
735 static const struct rpc_xprt_ops xprt_rdma_procs = {
736 	.reserve_xprt		= xprt_reserve_xprt_cong,
737 	.release_xprt		= xprt_release_xprt_cong, /* sunrpc/xprt.c */
738 	.alloc_slot		= xprt_rdma_alloc_slot,
739 	.free_slot		= xprt_rdma_free_slot,
740 	.release_request	= xprt_release_rqst_cong,       /* ditto */
741 	.wait_for_reply_request	= xprt_wait_for_reply_request_def, /* ditto */
742 	.timer			= xprt_rdma_timer,
743 	.rpcbind		= rpcb_getport_async,	/* sunrpc/rpcb_clnt.c */
744 	.set_port		= xprt_rdma_set_port,
745 	.connect		= xprt_rdma_connect,
746 	.buf_alloc		= xprt_rdma_allocate,
747 	.buf_free		= xprt_rdma_free,
748 	.send_request		= xprt_rdma_send_request,
749 	.close			= xprt_rdma_close,
750 	.destroy		= xprt_rdma_destroy,
751 	.set_connect_timeout	= xprt_rdma_set_connect_timeout,
752 	.print_stats		= xprt_rdma_print_stats,
753 	.enable_swap		= xprt_rdma_enable_swap,
754 	.disable_swap		= xprt_rdma_disable_swap,
755 	.inject_disconnect	= xprt_rdma_inject_disconnect,
756 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
757 	.bc_setup		= xprt_rdma_bc_setup,
758 	.bc_maxpayload		= xprt_rdma_bc_maxpayload,
759 	.bc_num_slots		= xprt_rdma_bc_max_slots,
760 	.bc_free_rqst		= xprt_rdma_bc_free_rqst,
761 	.bc_destroy		= xprt_rdma_bc_destroy,
762 #endif
763 };
764 
765 static struct xprt_class xprt_rdma = {
766 	.list			= LIST_HEAD_INIT(xprt_rdma.list),
767 	.name			= "rdma",
768 	.owner			= THIS_MODULE,
769 	.ident			= XPRT_TRANSPORT_RDMA,
770 	.setup			= xprt_setup_rdma,
771 };
772 
773 void xprt_rdma_cleanup(void)
774 {
775 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
776 	if (sunrpc_table_header) {
777 		unregister_sysctl_table(sunrpc_table_header);
778 		sunrpc_table_header = NULL;
779 	}
780 #endif
781 
782 	xprt_unregister_transport(&xprt_rdma);
783 	xprt_unregister_transport(&xprt_rdma_bc);
784 }
785 
786 int xprt_rdma_init(void)
787 {
788 	int rc;
789 
790 	rc = xprt_register_transport(&xprt_rdma);
791 	if (rc)
792 		return rc;
793 
794 	rc = xprt_register_transport(&xprt_rdma_bc);
795 	if (rc) {
796 		xprt_unregister_transport(&xprt_rdma);
797 		return rc;
798 	}
799 
800 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
801 	if (!sunrpc_table_header)
802 		sunrpc_table_header = register_sysctl_table(sunrpc_table);
803 #endif
804 	return 0;
805 }
806