xref: /openbmc/linux/net/sunrpc/xprtrdma/transport.c (revision 0661cb2a)
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 static struct xprt_class xprt_rdma;
77 
78 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
79 
80 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
81 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
82 static unsigned int min_inline_size = RPCRDMA_MIN_INLINE;
83 static unsigned int max_inline_size = RPCRDMA_MAX_INLINE;
84 static unsigned int max_padding = PAGE_SIZE;
85 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
86 static unsigned int max_memreg = RPCRDMA_LAST - 1;
87 static unsigned int dummy;
88 
89 static struct ctl_table_header *sunrpc_table_header;
90 
91 static struct ctl_table xr_tunables_table[] = {
92 	{
93 		.procname	= "rdma_slot_table_entries",
94 		.data		= &xprt_rdma_slot_table_entries,
95 		.maxlen		= sizeof(unsigned int),
96 		.mode		= 0644,
97 		.proc_handler	= proc_dointvec_minmax,
98 		.extra1		= &min_slot_table_size,
99 		.extra2		= &max_slot_table_size
100 	},
101 	{
102 		.procname	= "rdma_max_inline_read",
103 		.data		= &xprt_rdma_max_inline_read,
104 		.maxlen		= sizeof(unsigned int),
105 		.mode		= 0644,
106 		.proc_handler	= proc_dointvec_minmax,
107 		.extra1		= &min_inline_size,
108 		.extra2		= &max_inline_size,
109 	},
110 	{
111 		.procname	= "rdma_max_inline_write",
112 		.data		= &xprt_rdma_max_inline_write,
113 		.maxlen		= sizeof(unsigned int),
114 		.mode		= 0644,
115 		.proc_handler	= proc_dointvec_minmax,
116 		.extra1		= &min_inline_size,
117 		.extra2		= &max_inline_size,
118 	},
119 	{
120 		.procname	= "rdma_inline_write_padding",
121 		.data		= &dummy,
122 		.maxlen		= sizeof(unsigned int),
123 		.mode		= 0644,
124 		.proc_handler	= proc_dointvec_minmax,
125 		.extra1		= SYSCTL_ZERO,
126 		.extra2		= &max_padding,
127 	},
128 	{
129 		.procname	= "rdma_memreg_strategy",
130 		.data		= &xprt_rdma_memreg_strategy,
131 		.maxlen		= sizeof(unsigned int),
132 		.mode		= 0644,
133 		.proc_handler	= proc_dointvec_minmax,
134 		.extra1		= &min_memreg,
135 		.extra2		= &max_memreg,
136 	},
137 	{
138 		.procname	= "rdma_pad_optimize",
139 		.data		= &xprt_rdma_pad_optimize,
140 		.maxlen		= sizeof(unsigned int),
141 		.mode		= 0644,
142 		.proc_handler	= proc_dointvec,
143 	},
144 	{ },
145 };
146 
147 static struct ctl_table sunrpc_table[] = {
148 	{
149 		.procname	= "sunrpc",
150 		.mode		= 0555,
151 		.child		= xr_tunables_table
152 	},
153 	{ },
154 };
155 
156 #endif
157 
158 static const struct rpc_xprt_ops xprt_rdma_procs;
159 
160 static void
161 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
162 {
163 	struct sockaddr_in *sin = (struct sockaddr_in *)sap;
164 	char buf[20];
165 
166 	snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
167 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
168 
169 	xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
170 }
171 
172 static void
173 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
174 {
175 	struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
176 	char buf[40];
177 
178 	snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
179 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
180 
181 	xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
182 }
183 
184 void
185 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
186 {
187 	char buf[128];
188 
189 	switch (sap->sa_family) {
190 	case AF_INET:
191 		xprt_rdma_format_addresses4(xprt, sap);
192 		break;
193 	case AF_INET6:
194 		xprt_rdma_format_addresses6(xprt, sap);
195 		break;
196 	default:
197 		pr_err("rpcrdma: Unrecognized address family\n");
198 		return;
199 	}
200 
201 	(void)rpc_ntop(sap, buf, sizeof(buf));
202 	xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
203 
204 	snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
205 	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
206 
207 	snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
208 	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
209 
210 	xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
211 }
212 
213 void
214 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
215 {
216 	unsigned int i;
217 
218 	for (i = 0; i < RPC_DISPLAY_MAX; i++)
219 		switch (i) {
220 		case RPC_DISPLAY_PROTO:
221 		case RPC_DISPLAY_NETID:
222 			continue;
223 		default:
224 			kfree(xprt->address_strings[i]);
225 		}
226 }
227 
228 /**
229  * xprt_rdma_connect_worker - establish connection in the background
230  * @work: worker thread context
231  *
232  * Requester holds the xprt's send lock to prevent activity on this
233  * transport while a fresh connection is being established. RPC tasks
234  * sleep on the xprt's pending queue waiting for connect to complete.
235  */
236 static void
237 xprt_rdma_connect_worker(struct work_struct *work)
238 {
239 	struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
240 						   rx_connect_worker.work);
241 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
242 	int rc;
243 
244 	rc = rpcrdma_xprt_connect(r_xprt);
245 	xprt_clear_connecting(xprt);
246 	if (!rc) {
247 		xprt->connect_cookie++;
248 		xprt->stat.connect_count++;
249 		xprt->stat.connect_time += (long)jiffies -
250 					   xprt->stat.connect_start;
251 		xprt_set_connected(xprt);
252 		rc = -EAGAIN;
253 	} else {
254 		/* Force a call to xprt_rdma_close to clean up */
255 		spin_lock(&xprt->transport_lock);
256 		set_bit(XPRT_CLOSE_WAIT, &xprt->state);
257 		spin_unlock(&xprt->transport_lock);
258 	}
259 	xprt_wake_pending_tasks(xprt, rc);
260 }
261 
262 /**
263  * xprt_rdma_inject_disconnect - inject a connection fault
264  * @xprt: transport context
265  *
266  * If @xprt is connected, disconnect it to simulate spurious
267  * connection loss. Caller must hold @xprt's send lock to
268  * ensure that data structures and hardware resources are
269  * stable during the rdma_disconnect() call.
270  */
271 static void
272 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
273 {
274 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
275 
276 	trace_xprtrdma_op_inject_dsc(r_xprt);
277 	rdma_disconnect(r_xprt->rx_ep->re_id);
278 }
279 
280 /**
281  * xprt_rdma_destroy - Full tear down of transport
282  * @xprt: doomed transport context
283  *
284  * Caller guarantees there will be no more calls to us with
285  * this @xprt.
286  */
287 static void
288 xprt_rdma_destroy(struct rpc_xprt *xprt)
289 {
290 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
291 
292 	cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
293 
294 	rpcrdma_xprt_disconnect(r_xprt);
295 	rpcrdma_buffer_destroy(&r_xprt->rx_buf);
296 
297 	xprt_rdma_free_addresses(xprt);
298 	xprt_free(xprt);
299 
300 	module_put(THIS_MODULE);
301 }
302 
303 /* 60 second timeout, no retries */
304 static const struct rpc_timeout xprt_rdma_default_timeout = {
305 	.to_initval = 60 * HZ,
306 	.to_maxval = 60 * HZ,
307 };
308 
309 /**
310  * xprt_setup_rdma - Set up transport to use RDMA
311  *
312  * @args: rpc transport arguments
313  */
314 static struct rpc_xprt *
315 xprt_setup_rdma(struct xprt_create *args)
316 {
317 	struct rpc_xprt *xprt;
318 	struct rpcrdma_xprt *new_xprt;
319 	struct sockaddr *sap;
320 	int rc;
321 
322 	if (args->addrlen > sizeof(xprt->addr))
323 		return ERR_PTR(-EBADF);
324 
325 	if (!try_module_get(THIS_MODULE))
326 		return ERR_PTR(-EIO);
327 
328 	xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt), 0,
329 			  xprt_rdma_slot_table_entries);
330 	if (!xprt) {
331 		module_put(THIS_MODULE);
332 		return ERR_PTR(-ENOMEM);
333 	}
334 
335 	xprt->timeout = &xprt_rdma_default_timeout;
336 	xprt->connect_timeout = xprt->timeout->to_initval;
337 	xprt->max_reconnect_timeout = xprt->timeout->to_maxval;
338 	xprt->bind_timeout = RPCRDMA_BIND_TO;
339 	xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
340 	xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
341 
342 	xprt->resvport = 0;		/* privileged port not needed */
343 	xprt->ops = &xprt_rdma_procs;
344 
345 	/*
346 	 * Set up RDMA-specific connect data.
347 	 */
348 	sap = args->dstaddr;
349 
350 	/* Ensure xprt->addr holds valid server TCP (not RDMA)
351 	 * address, for any side protocols which peek at it */
352 	xprt->prot = IPPROTO_TCP;
353 	xprt->xprt_class = &xprt_rdma;
354 	xprt->addrlen = args->addrlen;
355 	memcpy(&xprt->addr, sap, xprt->addrlen);
356 
357 	if (rpc_get_port(sap))
358 		xprt_set_bound(xprt);
359 	xprt_rdma_format_addresses(xprt, sap);
360 
361 	new_xprt = rpcx_to_rdmax(xprt);
362 	rc = rpcrdma_buffer_create(new_xprt);
363 	if (rc) {
364 		xprt_rdma_free_addresses(xprt);
365 		xprt_free(xprt);
366 		module_put(THIS_MODULE);
367 		return ERR_PTR(rc);
368 	}
369 
370 	INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
371 			  xprt_rdma_connect_worker);
372 
373 	xprt->max_payload = RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT;
374 
375 	return xprt;
376 }
377 
378 /**
379  * xprt_rdma_close - close a transport connection
380  * @xprt: transport context
381  *
382  * Called during autoclose or device removal.
383  *
384  * Caller holds @xprt's send lock to prevent activity on this
385  * transport while the connection is torn down.
386  */
387 void xprt_rdma_close(struct rpc_xprt *xprt)
388 {
389 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
390 
391 	rpcrdma_xprt_disconnect(r_xprt);
392 
393 	xprt->reestablish_timeout = 0;
394 	++xprt->connect_cookie;
395 	xprt_disconnect_done(xprt);
396 }
397 
398 /**
399  * xprt_rdma_set_port - update server port with rpcbind result
400  * @xprt: controlling RPC transport
401  * @port: new port value
402  *
403  * Transport connect status is unchanged.
404  */
405 static void
406 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
407 {
408 	struct sockaddr *sap = (struct sockaddr *)&xprt->addr;
409 	char buf[8];
410 
411 	rpc_set_port(sap, port);
412 
413 	kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
414 	snprintf(buf, sizeof(buf), "%u", port);
415 	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
416 
417 	kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]);
418 	snprintf(buf, sizeof(buf), "%4hx", port);
419 	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
420 }
421 
422 /**
423  * xprt_rdma_timer - invoked when an RPC times out
424  * @xprt: controlling RPC transport
425  * @task: RPC task that timed out
426  *
427  * Invoked when the transport is still connected, but an RPC
428  * retransmit timeout occurs.
429  *
430  * Since RDMA connections don't have a keep-alive, forcibly
431  * disconnect and retry to connect. This drives full
432  * detection of the network path, and retransmissions of
433  * all pending RPCs.
434  */
435 static void
436 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
437 {
438 	xprt_force_disconnect(xprt);
439 }
440 
441 /**
442  * xprt_rdma_set_connect_timeout - set timeouts for establishing a connection
443  * @xprt: controlling transport instance
444  * @connect_timeout: reconnect timeout after client disconnects
445  * @reconnect_timeout: reconnect timeout after server disconnects
446  *
447  */
448 static void xprt_rdma_set_connect_timeout(struct rpc_xprt *xprt,
449 					  unsigned long connect_timeout,
450 					  unsigned long reconnect_timeout)
451 {
452 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
453 
454 	trace_xprtrdma_op_set_cto(r_xprt, connect_timeout, reconnect_timeout);
455 
456 	spin_lock(&xprt->transport_lock);
457 
458 	if (connect_timeout < xprt->connect_timeout) {
459 		struct rpc_timeout to;
460 		unsigned long initval;
461 
462 		to = *xprt->timeout;
463 		initval = connect_timeout;
464 		if (initval < RPCRDMA_INIT_REEST_TO << 1)
465 			initval = RPCRDMA_INIT_REEST_TO << 1;
466 		to.to_initval = initval;
467 		to.to_maxval = initval;
468 		r_xprt->rx_timeout = to;
469 		xprt->timeout = &r_xprt->rx_timeout;
470 		xprt->connect_timeout = connect_timeout;
471 	}
472 
473 	if (reconnect_timeout < xprt->max_reconnect_timeout)
474 		xprt->max_reconnect_timeout = reconnect_timeout;
475 
476 	spin_unlock(&xprt->transport_lock);
477 }
478 
479 /**
480  * xprt_rdma_connect - schedule an attempt to reconnect
481  * @xprt: transport state
482  * @task: RPC scheduler context (unused)
483  *
484  */
485 static void
486 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
487 {
488 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
489 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
490 	unsigned long delay;
491 
492 	delay = 0;
493 	if (ep && ep->re_connect_status != 0) {
494 		delay = xprt_reconnect_delay(xprt);
495 		xprt_reconnect_backoff(xprt, RPCRDMA_INIT_REEST_TO);
496 	}
497 	trace_xprtrdma_op_connect(r_xprt, delay);
498 	queue_delayed_work(xprtiod_workqueue, &r_xprt->rx_connect_worker,
499 			   delay);
500 }
501 
502 /**
503  * xprt_rdma_alloc_slot - allocate an rpc_rqst
504  * @xprt: controlling RPC transport
505  * @task: RPC task requesting a fresh rpc_rqst
506  *
507  * tk_status values:
508  *	%0 if task->tk_rqstp points to a fresh rpc_rqst
509  *	%-EAGAIN if no rpc_rqst is available; queued on backlog
510  */
511 static void
512 xprt_rdma_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
513 {
514 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
515 	struct rpcrdma_req *req;
516 
517 	req = rpcrdma_buffer_get(&r_xprt->rx_buf);
518 	if (!req)
519 		goto out_sleep;
520 	task->tk_rqstp = &req->rl_slot;
521 	task->tk_status = 0;
522 	return;
523 
524 out_sleep:
525 	task->tk_status = -EAGAIN;
526 	xprt_add_backlog(xprt, task);
527 }
528 
529 /**
530  * xprt_rdma_free_slot - release an rpc_rqst
531  * @xprt: controlling RPC transport
532  * @rqst: rpc_rqst to release
533  *
534  */
535 static void
536 xprt_rdma_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *rqst)
537 {
538 	struct rpcrdma_xprt *r_xprt =
539 		container_of(xprt, struct rpcrdma_xprt, rx_xprt);
540 
541 	rpcrdma_reply_put(&r_xprt->rx_buf, rpcr_to_rdmar(rqst));
542 	if (!xprt_wake_up_backlog(xprt, rqst)) {
543 		memset(rqst, 0, sizeof(*rqst));
544 		rpcrdma_buffer_put(&r_xprt->rx_buf, rpcr_to_rdmar(rqst));
545 	}
546 }
547 
548 static bool rpcrdma_check_regbuf(struct rpcrdma_xprt *r_xprt,
549 				 struct rpcrdma_regbuf *rb, size_t size,
550 				 gfp_t flags)
551 {
552 	if (unlikely(rdmab_length(rb) < size)) {
553 		if (!rpcrdma_regbuf_realloc(rb, size, flags))
554 			return false;
555 		r_xprt->rx_stats.hardway_register_count += size;
556 	}
557 	return true;
558 }
559 
560 /**
561  * xprt_rdma_allocate - allocate transport resources for an RPC
562  * @task: RPC task
563  *
564  * Return values:
565  *        0:	Success; rq_buffer points to RPC buffer to use
566  *   ENOMEM:	Out of memory, call again later
567  *      EIO:	A permanent error occurred, do not retry
568  */
569 static int
570 xprt_rdma_allocate(struct rpc_task *task)
571 {
572 	struct rpc_rqst *rqst = task->tk_rqstp;
573 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
574 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
575 	gfp_t flags;
576 
577 	flags = RPCRDMA_DEF_GFP;
578 	if (RPC_IS_SWAPPER(task))
579 		flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
580 
581 	if (!rpcrdma_check_regbuf(r_xprt, req->rl_sendbuf, rqst->rq_callsize,
582 				  flags))
583 		goto out_fail;
584 	if (!rpcrdma_check_regbuf(r_xprt, req->rl_recvbuf, rqst->rq_rcvsize,
585 				  flags))
586 		goto out_fail;
587 
588 	rqst->rq_buffer = rdmab_data(req->rl_sendbuf);
589 	rqst->rq_rbuffer = rdmab_data(req->rl_recvbuf);
590 	return 0;
591 
592 out_fail:
593 	return -ENOMEM;
594 }
595 
596 /**
597  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
598  * @task: RPC task
599  *
600  * Caller guarantees rqst->rq_buffer is non-NULL.
601  */
602 static void
603 xprt_rdma_free(struct rpc_task *task)
604 {
605 	struct rpc_rqst *rqst = task->tk_rqstp;
606 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
607 
608 	if (unlikely(!list_empty(&req->rl_registered))) {
609 		trace_xprtrdma_mrs_zap(task);
610 		frwr_unmap_sync(rpcx_to_rdmax(rqst->rq_xprt), req);
611 	}
612 
613 	/* XXX: If the RPC is completing because of a signal and
614 	 * not because a reply was received, we ought to ensure
615 	 * that the Send completion has fired, so that memory
616 	 * involved with the Send is not still visible to the NIC.
617 	 */
618 }
619 
620 /**
621  * xprt_rdma_send_request - marshal and send an RPC request
622  * @rqst: RPC message in rq_snd_buf
623  *
624  * Caller holds the transport's write lock.
625  *
626  * Returns:
627  *	%0 if the RPC message has been sent
628  *	%-ENOTCONN if the caller should reconnect and call again
629  *	%-EAGAIN if the caller should call again
630  *	%-ENOBUFS if the caller should call again after a delay
631  *	%-EMSGSIZE if encoding ran out of buffer space. The request
632  *		was not sent. Do not try to send this message again.
633  *	%-EIO if an I/O error occurred. The request was not sent.
634  *		Do not try to send this message again.
635  */
636 static int
637 xprt_rdma_send_request(struct rpc_rqst *rqst)
638 {
639 	struct rpc_xprt *xprt = rqst->rq_xprt;
640 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
641 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
642 	int rc = 0;
643 
644 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
645 	if (unlikely(!rqst->rq_buffer))
646 		return xprt_rdma_bc_send_reply(rqst);
647 #endif	/* CONFIG_SUNRPC_BACKCHANNEL */
648 
649 	if (!xprt_connected(xprt))
650 		return -ENOTCONN;
651 
652 	if (!xprt_request_get_cong(xprt, rqst))
653 		return -EBADSLT;
654 
655 	rc = rpcrdma_marshal_req(r_xprt, rqst);
656 	if (rc < 0)
657 		goto failed_marshal;
658 
659 	/* Must suppress retransmit to maintain credits */
660 	if (rqst->rq_connect_cookie == xprt->connect_cookie)
661 		goto drop_connection;
662 	rqst->rq_xtime = ktime_get();
663 
664 	if (rpcrdma_post_sends(r_xprt, req))
665 		goto drop_connection;
666 
667 	rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
668 
669 	/* An RPC with no reply will throw off credit accounting,
670 	 * so drop the connection to reset the credit grant.
671 	 */
672 	if (!rpc_reply_expected(rqst->rq_task))
673 		goto drop_connection;
674 	return 0;
675 
676 failed_marshal:
677 	if (rc != -ENOTCONN)
678 		return rc;
679 drop_connection:
680 	xprt_rdma_close(xprt);
681 	return -ENOTCONN;
682 }
683 
684 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
685 {
686 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
687 	long idle_time = 0;
688 
689 	if (xprt_connected(xprt))
690 		idle_time = (long)(jiffies - xprt->last_used) / HZ;
691 
692 	seq_puts(seq, "\txprt:\trdma ");
693 	seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
694 		   0,	/* need a local port? */
695 		   xprt->stat.bind_count,
696 		   xprt->stat.connect_count,
697 		   xprt->stat.connect_time / HZ,
698 		   idle_time,
699 		   xprt->stat.sends,
700 		   xprt->stat.recvs,
701 		   xprt->stat.bad_xids,
702 		   xprt->stat.req_u,
703 		   xprt->stat.bklog_u);
704 	seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
705 		   r_xprt->rx_stats.read_chunk_count,
706 		   r_xprt->rx_stats.write_chunk_count,
707 		   r_xprt->rx_stats.reply_chunk_count,
708 		   r_xprt->rx_stats.total_rdma_request,
709 		   r_xprt->rx_stats.total_rdma_reply,
710 		   r_xprt->rx_stats.pullup_copy_count,
711 		   r_xprt->rx_stats.fixup_copy_count,
712 		   r_xprt->rx_stats.hardway_register_count,
713 		   r_xprt->rx_stats.failed_marshal_count,
714 		   r_xprt->rx_stats.bad_reply_count,
715 		   r_xprt->rx_stats.nomsg_call_count);
716 	seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n",
717 		   r_xprt->rx_stats.mrs_recycled,
718 		   r_xprt->rx_stats.mrs_orphaned,
719 		   r_xprt->rx_stats.mrs_allocated,
720 		   r_xprt->rx_stats.local_inv_needed,
721 		   r_xprt->rx_stats.empty_sendctx_q,
722 		   r_xprt->rx_stats.reply_waits_for_send);
723 }
724 
725 static int
726 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
727 {
728 	return 0;
729 }
730 
731 static void
732 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
733 {
734 }
735 
736 /*
737  * Plumbing for rpc transport switch and kernel module
738  */
739 
740 static const struct rpc_xprt_ops xprt_rdma_procs = {
741 	.reserve_xprt		= xprt_reserve_xprt_cong,
742 	.release_xprt		= xprt_release_xprt_cong, /* sunrpc/xprt.c */
743 	.alloc_slot		= xprt_rdma_alloc_slot,
744 	.free_slot		= xprt_rdma_free_slot,
745 	.release_request	= xprt_release_rqst_cong,       /* ditto */
746 	.wait_for_reply_request	= xprt_wait_for_reply_request_def, /* ditto */
747 	.timer			= xprt_rdma_timer,
748 	.rpcbind		= rpcb_getport_async,	/* sunrpc/rpcb_clnt.c */
749 	.set_port		= xprt_rdma_set_port,
750 	.connect		= xprt_rdma_connect,
751 	.buf_alloc		= xprt_rdma_allocate,
752 	.buf_free		= xprt_rdma_free,
753 	.send_request		= xprt_rdma_send_request,
754 	.close			= xprt_rdma_close,
755 	.destroy		= xprt_rdma_destroy,
756 	.set_connect_timeout	= xprt_rdma_set_connect_timeout,
757 	.print_stats		= xprt_rdma_print_stats,
758 	.enable_swap		= xprt_rdma_enable_swap,
759 	.disable_swap		= xprt_rdma_disable_swap,
760 	.inject_disconnect	= xprt_rdma_inject_disconnect,
761 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
762 	.bc_setup		= xprt_rdma_bc_setup,
763 	.bc_maxpayload		= xprt_rdma_bc_maxpayload,
764 	.bc_num_slots		= xprt_rdma_bc_max_slots,
765 	.bc_free_rqst		= xprt_rdma_bc_free_rqst,
766 	.bc_destroy		= xprt_rdma_bc_destroy,
767 #endif
768 };
769 
770 static struct xprt_class xprt_rdma = {
771 	.list			= LIST_HEAD_INIT(xprt_rdma.list),
772 	.name			= "rdma",
773 	.owner			= THIS_MODULE,
774 	.ident			= XPRT_TRANSPORT_RDMA,
775 	.setup			= xprt_setup_rdma,
776 	.netid			= { "rdma", "rdma6", "" },
777 };
778 
779 void xprt_rdma_cleanup(void)
780 {
781 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
782 	if (sunrpc_table_header) {
783 		unregister_sysctl_table(sunrpc_table_header);
784 		sunrpc_table_header = NULL;
785 	}
786 #endif
787 
788 	xprt_unregister_transport(&xprt_rdma);
789 	xprt_unregister_transport(&xprt_rdma_bc);
790 }
791 
792 int xprt_rdma_init(void)
793 {
794 	int rc;
795 
796 	rc = xprt_register_transport(&xprt_rdma);
797 	if (rc)
798 		return rc;
799 
800 	rc = xprt_register_transport(&xprt_rdma_bc);
801 	if (rc) {
802 		xprt_unregister_transport(&xprt_rdma);
803 		return rc;
804 	}
805 
806 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
807 	if (!sunrpc_table_header)
808 		sunrpc_table_header = register_sysctl_table(sunrpc_table);
809 #endif
810 	return 0;
811 }
812