xref: /openbmc/linux/net/sunrpc/xprtrdma/transport.c (revision b625fe69)
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 /*
64  * tunables
65  */
66 
67 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
68 unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
69 unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
70 unsigned int xprt_rdma_memreg_strategy		= RPCRDMA_FRWR;
71 int xprt_rdma_pad_optimize;
72 static struct xprt_class xprt_rdma;
73 
74 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
75 
76 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
77 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
78 static unsigned int min_inline_size = RPCRDMA_MIN_INLINE;
79 static unsigned int max_inline_size = RPCRDMA_MAX_INLINE;
80 static unsigned int max_padding = PAGE_SIZE;
81 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
82 static unsigned int max_memreg = RPCRDMA_LAST - 1;
83 static unsigned int dummy;
84 
85 static struct ctl_table_header *sunrpc_table_header;
86 
87 static struct ctl_table xr_tunables_table[] = {
88 	{
89 		.procname	= "rdma_slot_table_entries",
90 		.data		= &xprt_rdma_slot_table_entries,
91 		.maxlen		= sizeof(unsigned int),
92 		.mode		= 0644,
93 		.proc_handler	= proc_dointvec_minmax,
94 		.extra1		= &min_slot_table_size,
95 		.extra2		= &max_slot_table_size
96 	},
97 	{
98 		.procname	= "rdma_max_inline_read",
99 		.data		= &xprt_rdma_max_inline_read,
100 		.maxlen		= sizeof(unsigned int),
101 		.mode		= 0644,
102 		.proc_handler	= proc_dointvec_minmax,
103 		.extra1		= &min_inline_size,
104 		.extra2		= &max_inline_size,
105 	},
106 	{
107 		.procname	= "rdma_max_inline_write",
108 		.data		= &xprt_rdma_max_inline_write,
109 		.maxlen		= sizeof(unsigned int),
110 		.mode		= 0644,
111 		.proc_handler	= proc_dointvec_minmax,
112 		.extra1		= &min_inline_size,
113 		.extra2		= &max_inline_size,
114 	},
115 	{
116 		.procname	= "rdma_inline_write_padding",
117 		.data		= &dummy,
118 		.maxlen		= sizeof(unsigned int),
119 		.mode		= 0644,
120 		.proc_handler	= proc_dointvec_minmax,
121 		.extra1		= SYSCTL_ZERO,
122 		.extra2		= &max_padding,
123 	},
124 	{
125 		.procname	= "rdma_memreg_strategy",
126 		.data		= &xprt_rdma_memreg_strategy,
127 		.maxlen		= sizeof(unsigned int),
128 		.mode		= 0644,
129 		.proc_handler	= proc_dointvec_minmax,
130 		.extra1		= &min_memreg,
131 		.extra2		= &max_memreg,
132 	},
133 	{
134 		.procname	= "rdma_pad_optimize",
135 		.data		= &xprt_rdma_pad_optimize,
136 		.maxlen		= sizeof(unsigned int),
137 		.mode		= 0644,
138 		.proc_handler	= proc_dointvec,
139 	},
140 	{ },
141 };
142 
143 static struct ctl_table sunrpc_table[] = {
144 	{
145 		.procname	= "sunrpc",
146 		.mode		= 0555,
147 		.child		= xr_tunables_table
148 	},
149 	{ },
150 };
151 
152 #endif
153 
154 static const struct rpc_xprt_ops xprt_rdma_procs;
155 
156 static void
157 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
158 {
159 	struct sockaddr_in *sin = (struct sockaddr_in *)sap;
160 	char buf[20];
161 
162 	snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
163 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
164 
165 	xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
166 }
167 
168 static void
169 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
170 {
171 	struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
172 	char buf[40];
173 
174 	snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
175 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
176 
177 	xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
178 }
179 
180 void
181 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
182 {
183 	char buf[128];
184 
185 	switch (sap->sa_family) {
186 	case AF_INET:
187 		xprt_rdma_format_addresses4(xprt, sap);
188 		break;
189 	case AF_INET6:
190 		xprt_rdma_format_addresses6(xprt, sap);
191 		break;
192 	default:
193 		pr_err("rpcrdma: Unrecognized address family\n");
194 		return;
195 	}
196 
197 	(void)rpc_ntop(sap, buf, sizeof(buf));
198 	xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
199 
200 	snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
201 	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
202 
203 	snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
204 	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
205 
206 	xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
207 }
208 
209 void
210 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
211 {
212 	unsigned int i;
213 
214 	for (i = 0; i < RPC_DISPLAY_MAX; i++)
215 		switch (i) {
216 		case RPC_DISPLAY_PROTO:
217 		case RPC_DISPLAY_NETID:
218 			continue;
219 		default:
220 			kfree(xprt->address_strings[i]);
221 		}
222 }
223 
224 /**
225  * xprt_rdma_connect_worker - establish connection in the background
226  * @work: worker thread context
227  *
228  * Requester holds the xprt's send lock to prevent activity on this
229  * transport while a fresh connection is being established. RPC tasks
230  * sleep on the xprt's pending queue waiting for connect to complete.
231  */
232 static void
233 xprt_rdma_connect_worker(struct work_struct *work)
234 {
235 	struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
236 						   rx_connect_worker.work);
237 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
238 	int rc;
239 
240 	rc = rpcrdma_xprt_connect(r_xprt);
241 	xprt_clear_connecting(xprt);
242 	if (!rc) {
243 		xprt->connect_cookie++;
244 		xprt->stat.connect_count++;
245 		xprt->stat.connect_time += (long)jiffies -
246 					   xprt->stat.connect_start;
247 		xprt_set_connected(xprt);
248 		rc = -EAGAIN;
249 	} else
250 		rpcrdma_xprt_disconnect(r_xprt);
251 	xprt_unlock_connect(xprt, r_xprt);
252 	xprt_wake_pending_tasks(xprt, rc);
253 }
254 
255 /**
256  * xprt_rdma_inject_disconnect - inject a connection fault
257  * @xprt: transport context
258  *
259  * If @xprt is connected, disconnect it to simulate spurious
260  * connection loss. Caller must hold @xprt's send lock to
261  * ensure that data structures and hardware resources are
262  * stable during the rdma_disconnect() call.
263  */
264 static void
265 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
266 {
267 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
268 
269 	trace_xprtrdma_op_inject_dsc(r_xprt);
270 	rdma_disconnect(r_xprt->rx_ep->re_id);
271 }
272 
273 /**
274  * xprt_rdma_destroy - Full tear down of transport
275  * @xprt: doomed transport context
276  *
277  * Caller guarantees there will be no more calls to us with
278  * this @xprt.
279  */
280 static void
281 xprt_rdma_destroy(struct rpc_xprt *xprt)
282 {
283 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
284 
285 	cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
286 
287 	rpcrdma_xprt_disconnect(r_xprt);
288 	rpcrdma_buffer_destroy(&r_xprt->rx_buf);
289 
290 	xprt_rdma_free_addresses(xprt);
291 	xprt_free(xprt);
292 
293 	module_put(THIS_MODULE);
294 }
295 
296 /* 60 second timeout, no retries */
297 static const struct rpc_timeout xprt_rdma_default_timeout = {
298 	.to_initval = 60 * HZ,
299 	.to_maxval = 60 * HZ,
300 };
301 
302 /**
303  * xprt_setup_rdma - Set up transport to use RDMA
304  *
305  * @args: rpc transport arguments
306  */
307 static struct rpc_xprt *
308 xprt_setup_rdma(struct xprt_create *args)
309 {
310 	struct rpc_xprt *xprt;
311 	struct rpcrdma_xprt *new_xprt;
312 	struct sockaddr *sap;
313 	int rc;
314 
315 	if (args->addrlen > sizeof(xprt->addr))
316 		return ERR_PTR(-EBADF);
317 
318 	if (!try_module_get(THIS_MODULE))
319 		return ERR_PTR(-EIO);
320 
321 	xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt), 0,
322 			  xprt_rdma_slot_table_entries);
323 	if (!xprt) {
324 		module_put(THIS_MODULE);
325 		return ERR_PTR(-ENOMEM);
326 	}
327 
328 	xprt->timeout = &xprt_rdma_default_timeout;
329 	xprt->connect_timeout = xprt->timeout->to_initval;
330 	xprt->max_reconnect_timeout = xprt->timeout->to_maxval;
331 	xprt->bind_timeout = RPCRDMA_BIND_TO;
332 	xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
333 	xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
334 
335 	xprt->resvport = 0;		/* privileged port not needed */
336 	xprt->ops = &xprt_rdma_procs;
337 
338 	/*
339 	 * Set up RDMA-specific connect data.
340 	 */
341 	sap = args->dstaddr;
342 
343 	/* Ensure xprt->addr holds valid server TCP (not RDMA)
344 	 * address, for any side protocols which peek at it */
345 	xprt->prot = IPPROTO_TCP;
346 	xprt->xprt_class = &xprt_rdma;
347 	xprt->addrlen = args->addrlen;
348 	memcpy(&xprt->addr, sap, xprt->addrlen);
349 
350 	if (rpc_get_port(sap))
351 		xprt_set_bound(xprt);
352 	xprt_rdma_format_addresses(xprt, sap);
353 
354 	new_xprt = rpcx_to_rdmax(xprt);
355 	rc = rpcrdma_buffer_create(new_xprt);
356 	if (rc) {
357 		xprt_rdma_free_addresses(xprt);
358 		xprt_free(xprt);
359 		module_put(THIS_MODULE);
360 		return ERR_PTR(rc);
361 	}
362 
363 	INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
364 			  xprt_rdma_connect_worker);
365 
366 	xprt->max_payload = RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT;
367 
368 	return xprt;
369 }
370 
371 /**
372  * xprt_rdma_close - close a transport connection
373  * @xprt: transport context
374  *
375  * Called during autoclose or device removal.
376  *
377  * Caller holds @xprt's send lock to prevent activity on this
378  * transport while the connection is torn down.
379  */
380 void xprt_rdma_close(struct rpc_xprt *xprt)
381 {
382 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
383 
384 	rpcrdma_xprt_disconnect(r_xprt);
385 
386 	xprt->reestablish_timeout = 0;
387 	++xprt->connect_cookie;
388 	xprt_disconnect_done(xprt);
389 }
390 
391 /**
392  * xprt_rdma_set_port - update server port with rpcbind result
393  * @xprt: controlling RPC transport
394  * @port: new port value
395  *
396  * Transport connect status is unchanged.
397  */
398 static void
399 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
400 {
401 	struct sockaddr *sap = (struct sockaddr *)&xprt->addr;
402 	char buf[8];
403 
404 	rpc_set_port(sap, port);
405 
406 	kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
407 	snprintf(buf, sizeof(buf), "%u", port);
408 	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
409 
410 	kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]);
411 	snprintf(buf, sizeof(buf), "%4hx", port);
412 	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
413 }
414 
415 /**
416  * xprt_rdma_timer - invoked when an RPC times out
417  * @xprt: controlling RPC transport
418  * @task: RPC task that timed out
419  *
420  * Invoked when the transport is still connected, but an RPC
421  * retransmit timeout occurs.
422  *
423  * Since RDMA connections don't have a keep-alive, forcibly
424  * disconnect and retry to connect. This drives full
425  * detection of the network path, and retransmissions of
426  * all pending RPCs.
427  */
428 static void
429 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
430 {
431 	xprt_force_disconnect(xprt);
432 }
433 
434 /**
435  * xprt_rdma_set_connect_timeout - set timeouts for establishing a connection
436  * @xprt: controlling transport instance
437  * @connect_timeout: reconnect timeout after client disconnects
438  * @reconnect_timeout: reconnect timeout after server disconnects
439  *
440  */
441 static void xprt_rdma_set_connect_timeout(struct rpc_xprt *xprt,
442 					  unsigned long connect_timeout,
443 					  unsigned long reconnect_timeout)
444 {
445 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
446 
447 	trace_xprtrdma_op_set_cto(r_xprt, connect_timeout, reconnect_timeout);
448 
449 	spin_lock(&xprt->transport_lock);
450 
451 	if (connect_timeout < xprt->connect_timeout) {
452 		struct rpc_timeout to;
453 		unsigned long initval;
454 
455 		to = *xprt->timeout;
456 		initval = connect_timeout;
457 		if (initval < RPCRDMA_INIT_REEST_TO << 1)
458 			initval = RPCRDMA_INIT_REEST_TO << 1;
459 		to.to_initval = initval;
460 		to.to_maxval = initval;
461 		r_xprt->rx_timeout = to;
462 		xprt->timeout = &r_xprt->rx_timeout;
463 		xprt->connect_timeout = connect_timeout;
464 	}
465 
466 	if (reconnect_timeout < xprt->max_reconnect_timeout)
467 		xprt->max_reconnect_timeout = reconnect_timeout;
468 
469 	spin_unlock(&xprt->transport_lock);
470 }
471 
472 /**
473  * xprt_rdma_connect - schedule an attempt to reconnect
474  * @xprt: transport state
475  * @task: RPC scheduler context (unused)
476  *
477  */
478 static void
479 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
480 {
481 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
482 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
483 	unsigned long delay;
484 
485 	WARN_ON_ONCE(!xprt_lock_connect(xprt, task, r_xprt));
486 
487 	delay = 0;
488 	if (ep && ep->re_connect_status != 0) {
489 		delay = xprt_reconnect_delay(xprt);
490 		xprt_reconnect_backoff(xprt, RPCRDMA_INIT_REEST_TO);
491 	}
492 	trace_xprtrdma_op_connect(r_xprt, delay);
493 	queue_delayed_work(xprtiod_workqueue, &r_xprt->rx_connect_worker,
494 			   delay);
495 }
496 
497 /**
498  * xprt_rdma_alloc_slot - allocate an rpc_rqst
499  * @xprt: controlling RPC transport
500  * @task: RPC task requesting a fresh rpc_rqst
501  *
502  * tk_status values:
503  *	%0 if task->tk_rqstp points to a fresh rpc_rqst
504  *	%-EAGAIN if no rpc_rqst is available; queued on backlog
505  */
506 static void
507 xprt_rdma_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
508 {
509 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
510 	struct rpcrdma_req *req;
511 
512 	req = rpcrdma_buffer_get(&r_xprt->rx_buf);
513 	if (!req)
514 		goto out_sleep;
515 	task->tk_rqstp = &req->rl_slot;
516 	task->tk_status = 0;
517 	return;
518 
519 out_sleep:
520 	task->tk_status = -EAGAIN;
521 	xprt_add_backlog(xprt, task);
522 }
523 
524 /**
525  * xprt_rdma_free_slot - release an rpc_rqst
526  * @xprt: controlling RPC transport
527  * @rqst: rpc_rqst to release
528  *
529  */
530 static void
531 xprt_rdma_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *rqst)
532 {
533 	struct rpcrdma_xprt *r_xprt =
534 		container_of(xprt, struct rpcrdma_xprt, rx_xprt);
535 
536 	rpcrdma_reply_put(&r_xprt->rx_buf, rpcr_to_rdmar(rqst));
537 	if (!xprt_wake_up_backlog(xprt, rqst)) {
538 		memset(rqst, 0, sizeof(*rqst));
539 		rpcrdma_buffer_put(&r_xprt->rx_buf, rpcr_to_rdmar(rqst));
540 	}
541 }
542 
543 static bool rpcrdma_check_regbuf(struct rpcrdma_xprt *r_xprt,
544 				 struct rpcrdma_regbuf *rb, size_t size,
545 				 gfp_t flags)
546 {
547 	if (unlikely(rdmab_length(rb) < size)) {
548 		if (!rpcrdma_regbuf_realloc(rb, size, flags))
549 			return false;
550 		r_xprt->rx_stats.hardway_register_count += size;
551 	}
552 	return true;
553 }
554 
555 /**
556  * xprt_rdma_allocate - allocate transport resources for an RPC
557  * @task: RPC task
558  *
559  * Return values:
560  *        0:	Success; rq_buffer points to RPC buffer to use
561  *   ENOMEM:	Out of memory, call again later
562  *      EIO:	A permanent error occurred, do not retry
563  */
564 static int
565 xprt_rdma_allocate(struct rpc_task *task)
566 {
567 	struct rpc_rqst *rqst = task->tk_rqstp;
568 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
569 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
570 	gfp_t flags;
571 
572 	flags = RPCRDMA_DEF_GFP;
573 	if (RPC_IS_SWAPPER(task))
574 		flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
575 
576 	if (!rpcrdma_check_regbuf(r_xprt, req->rl_sendbuf, rqst->rq_callsize,
577 				  flags))
578 		goto out_fail;
579 	if (!rpcrdma_check_regbuf(r_xprt, req->rl_recvbuf, rqst->rq_rcvsize,
580 				  flags))
581 		goto out_fail;
582 
583 	rqst->rq_buffer = rdmab_data(req->rl_sendbuf);
584 	rqst->rq_rbuffer = rdmab_data(req->rl_recvbuf);
585 	return 0;
586 
587 out_fail:
588 	return -ENOMEM;
589 }
590 
591 /**
592  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
593  * @task: RPC task
594  *
595  * Caller guarantees rqst->rq_buffer is non-NULL.
596  */
597 static void
598 xprt_rdma_free(struct rpc_task *task)
599 {
600 	struct rpc_rqst *rqst = task->tk_rqstp;
601 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
602 
603 	if (unlikely(!list_empty(&req->rl_registered))) {
604 		trace_xprtrdma_mrs_zap(task);
605 		frwr_unmap_sync(rpcx_to_rdmax(rqst->rq_xprt), req);
606 	}
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 (frwr_send(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 	.netid			= { "rdma", "rdma6", "" },
772 };
773 
774 void xprt_rdma_cleanup(void)
775 {
776 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
777 	if (sunrpc_table_header) {
778 		unregister_sysctl_table(sunrpc_table_header);
779 		sunrpc_table_header = NULL;
780 	}
781 #endif
782 
783 	xprt_unregister_transport(&xprt_rdma);
784 	xprt_unregister_transport(&xprt_rdma_bc);
785 }
786 
787 int xprt_rdma_init(void)
788 {
789 	int rc;
790 
791 	rc = xprt_register_transport(&xprt_rdma);
792 	if (rc)
793 		return rc;
794 
795 	rc = xprt_register_transport(&xprt_rdma_bc);
796 	if (rc) {
797 		xprt_unregister_transport(&xprt_rdma);
798 		return rc;
799 	}
800 
801 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
802 	if (!sunrpc_table_header)
803 		sunrpc_table_header = register_sysctl_table(sunrpc_table);
804 #endif
805 	return 0;
806 }
807