xref: /openbmc/linux/net/sunrpc/xprtrdma/transport.c (revision cabce92d)
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 #endif
144 
145 static const struct rpc_xprt_ops xprt_rdma_procs;
146 
147 static void
148 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
149 {
150 	struct sockaddr_in *sin = (struct sockaddr_in *)sap;
151 	char buf[20];
152 
153 	snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
154 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
155 
156 	xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
157 }
158 
159 static void
160 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
161 {
162 	struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
163 	char buf[40];
164 
165 	snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
166 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
167 
168 	xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
169 }
170 
171 void
172 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
173 {
174 	char buf[128];
175 
176 	switch (sap->sa_family) {
177 	case AF_INET:
178 		xprt_rdma_format_addresses4(xprt, sap);
179 		break;
180 	case AF_INET6:
181 		xprt_rdma_format_addresses6(xprt, sap);
182 		break;
183 	default:
184 		pr_err("rpcrdma: Unrecognized address family\n");
185 		return;
186 	}
187 
188 	(void)rpc_ntop(sap, buf, sizeof(buf));
189 	xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
190 
191 	snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
192 	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
193 
194 	snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
195 	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
196 
197 	xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
198 }
199 
200 void
201 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
202 {
203 	unsigned int i;
204 
205 	for (i = 0; i < RPC_DISPLAY_MAX; i++)
206 		switch (i) {
207 		case RPC_DISPLAY_PROTO:
208 		case RPC_DISPLAY_NETID:
209 			continue;
210 		default:
211 			kfree(xprt->address_strings[i]);
212 		}
213 }
214 
215 /**
216  * xprt_rdma_connect_worker - establish connection in the background
217  * @work: worker thread context
218  *
219  * Requester holds the xprt's send lock to prevent activity on this
220  * transport while a fresh connection is being established. RPC tasks
221  * sleep on the xprt's pending queue waiting for connect to complete.
222  */
223 static void
224 xprt_rdma_connect_worker(struct work_struct *work)
225 {
226 	struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
227 						   rx_connect_worker.work);
228 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
229 	unsigned int pflags = current->flags;
230 	int rc;
231 
232 	if (atomic_read(&xprt->swapper))
233 		current->flags |= PF_MEMALLOC;
234 	rc = rpcrdma_xprt_connect(r_xprt);
235 	xprt_clear_connecting(xprt);
236 	if (!rc) {
237 		xprt->connect_cookie++;
238 		xprt->stat.connect_count++;
239 		xprt->stat.connect_time += (long)jiffies -
240 					   xprt->stat.connect_start;
241 		xprt_set_connected(xprt);
242 		rc = -EAGAIN;
243 	} else
244 		rpcrdma_xprt_disconnect(r_xprt);
245 	xprt_unlock_connect(xprt, r_xprt);
246 	xprt_wake_pending_tasks(xprt, rc);
247 	current_restore_flags(pflags, PF_MEMALLOC);
248 }
249 
250 /**
251  * xprt_rdma_inject_disconnect - inject a connection fault
252  * @xprt: transport context
253  *
254  * If @xprt is connected, disconnect it to simulate spurious
255  * connection loss. Caller must hold @xprt's send lock to
256  * ensure that data structures and hardware resources are
257  * stable during the rdma_disconnect() call.
258  */
259 static void
260 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
261 {
262 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
263 
264 	trace_xprtrdma_op_inject_dsc(r_xprt);
265 	rdma_disconnect(r_xprt->rx_ep->re_id);
266 }
267 
268 /**
269  * xprt_rdma_destroy - Full tear down of transport
270  * @xprt: doomed transport context
271  *
272  * Caller guarantees there will be no more calls to us with
273  * this @xprt.
274  */
275 static void
276 xprt_rdma_destroy(struct rpc_xprt *xprt)
277 {
278 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
279 
280 	cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
281 
282 	rpcrdma_xprt_disconnect(r_xprt);
283 	rpcrdma_buffer_destroy(&r_xprt->rx_buf);
284 
285 	xprt_rdma_free_addresses(xprt);
286 	xprt_free(xprt);
287 
288 	module_put(THIS_MODULE);
289 }
290 
291 /* 60 second timeout, no retries */
292 static const struct rpc_timeout xprt_rdma_default_timeout = {
293 	.to_initval = 60 * HZ,
294 	.to_maxval = 60 * HZ,
295 };
296 
297 /**
298  * xprt_setup_rdma - Set up transport to use RDMA
299  *
300  * @args: rpc transport arguments
301  */
302 static struct rpc_xprt *
303 xprt_setup_rdma(struct xprt_create *args)
304 {
305 	struct rpc_xprt *xprt;
306 	struct rpcrdma_xprt *new_xprt;
307 	struct sockaddr *sap;
308 	int rc;
309 
310 	if (args->addrlen > sizeof(xprt->addr))
311 		return ERR_PTR(-EBADF);
312 
313 	if (!try_module_get(THIS_MODULE))
314 		return ERR_PTR(-EIO);
315 
316 	xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt), 0,
317 			  xprt_rdma_slot_table_entries);
318 	if (!xprt) {
319 		module_put(THIS_MODULE);
320 		return ERR_PTR(-ENOMEM);
321 	}
322 
323 	xprt->timeout = &xprt_rdma_default_timeout;
324 	xprt->connect_timeout = xprt->timeout->to_initval;
325 	xprt->max_reconnect_timeout = xprt->timeout->to_maxval;
326 	xprt->bind_timeout = RPCRDMA_BIND_TO;
327 	xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
328 	xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
329 
330 	xprt->resvport = 0;		/* privileged port not needed */
331 	xprt->ops = &xprt_rdma_procs;
332 
333 	/*
334 	 * Set up RDMA-specific connect data.
335 	 */
336 	sap = args->dstaddr;
337 
338 	/* Ensure xprt->addr holds valid server TCP (not RDMA)
339 	 * address, for any side protocols which peek at it */
340 	xprt->prot = IPPROTO_TCP;
341 	xprt->xprt_class = &xprt_rdma;
342 	xprt->addrlen = args->addrlen;
343 	memcpy(&xprt->addr, sap, xprt->addrlen);
344 
345 	if (rpc_get_port(sap))
346 		xprt_set_bound(xprt);
347 	xprt_rdma_format_addresses(xprt, sap);
348 
349 	new_xprt = rpcx_to_rdmax(xprt);
350 	rc = rpcrdma_buffer_create(new_xprt);
351 	if (rc) {
352 		xprt_rdma_free_addresses(xprt);
353 		xprt_free(xprt);
354 		module_put(THIS_MODULE);
355 		return ERR_PTR(rc);
356 	}
357 
358 	INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
359 			  xprt_rdma_connect_worker);
360 
361 	xprt->max_payload = RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT;
362 
363 	return xprt;
364 }
365 
366 /**
367  * xprt_rdma_close - close a transport connection
368  * @xprt: transport context
369  *
370  * Called during autoclose or device removal.
371  *
372  * Caller holds @xprt's send lock to prevent activity on this
373  * transport while the connection is torn down.
374  */
375 void xprt_rdma_close(struct rpc_xprt *xprt)
376 {
377 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
378 
379 	rpcrdma_xprt_disconnect(r_xprt);
380 
381 	xprt->reestablish_timeout = 0;
382 	++xprt->connect_cookie;
383 	xprt_disconnect_done(xprt);
384 }
385 
386 /**
387  * xprt_rdma_set_port - update server port with rpcbind result
388  * @xprt: controlling RPC transport
389  * @port: new port value
390  *
391  * Transport connect status is unchanged.
392  */
393 static void
394 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
395 {
396 	struct sockaddr *sap = (struct sockaddr *)&xprt->addr;
397 	char buf[8];
398 
399 	rpc_set_port(sap, port);
400 
401 	kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
402 	snprintf(buf, sizeof(buf), "%u", port);
403 	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
404 
405 	kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]);
406 	snprintf(buf, sizeof(buf), "%4hx", port);
407 	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
408 }
409 
410 /**
411  * xprt_rdma_timer - invoked when an RPC times out
412  * @xprt: controlling RPC transport
413  * @task: RPC task that timed out
414  *
415  * Invoked when the transport is still connected, but an RPC
416  * retransmit timeout occurs.
417  *
418  * Since RDMA connections don't have a keep-alive, forcibly
419  * disconnect and retry to connect. This drives full
420  * detection of the network path, and retransmissions of
421  * all pending RPCs.
422  */
423 static void
424 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
425 {
426 	xprt_force_disconnect(xprt);
427 }
428 
429 /**
430  * xprt_rdma_set_connect_timeout - set timeouts for establishing a connection
431  * @xprt: controlling transport instance
432  * @connect_timeout: reconnect timeout after client disconnects
433  * @reconnect_timeout: reconnect timeout after server disconnects
434  *
435  */
436 static void xprt_rdma_set_connect_timeout(struct rpc_xprt *xprt,
437 					  unsigned long connect_timeout,
438 					  unsigned long reconnect_timeout)
439 {
440 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
441 
442 	trace_xprtrdma_op_set_cto(r_xprt, connect_timeout, reconnect_timeout);
443 
444 	spin_lock(&xprt->transport_lock);
445 
446 	if (connect_timeout < xprt->connect_timeout) {
447 		struct rpc_timeout to;
448 		unsigned long initval;
449 
450 		to = *xprt->timeout;
451 		initval = connect_timeout;
452 		if (initval < RPCRDMA_INIT_REEST_TO << 1)
453 			initval = RPCRDMA_INIT_REEST_TO << 1;
454 		to.to_initval = initval;
455 		to.to_maxval = initval;
456 		r_xprt->rx_timeout = to;
457 		xprt->timeout = &r_xprt->rx_timeout;
458 		xprt->connect_timeout = connect_timeout;
459 	}
460 
461 	if (reconnect_timeout < xprt->max_reconnect_timeout)
462 		xprt->max_reconnect_timeout = reconnect_timeout;
463 
464 	spin_unlock(&xprt->transport_lock);
465 }
466 
467 /**
468  * xprt_rdma_connect - schedule an attempt to reconnect
469  * @xprt: transport state
470  * @task: RPC scheduler context (unused)
471  *
472  */
473 static void
474 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
475 {
476 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
477 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
478 	unsigned long delay;
479 
480 	WARN_ON_ONCE(!xprt_lock_connect(xprt, task, r_xprt));
481 
482 	delay = 0;
483 	if (ep && ep->re_connect_status != 0) {
484 		delay = xprt_reconnect_delay(xprt);
485 		xprt_reconnect_backoff(xprt, RPCRDMA_INIT_REEST_TO);
486 	}
487 	trace_xprtrdma_op_connect(r_xprt, delay);
488 	queue_delayed_work(system_long_wq, &r_xprt->rx_connect_worker, delay);
489 }
490 
491 /**
492  * xprt_rdma_alloc_slot - allocate an rpc_rqst
493  * @xprt: controlling RPC transport
494  * @task: RPC task requesting a fresh rpc_rqst
495  *
496  * tk_status values:
497  *	%0 if task->tk_rqstp points to a fresh rpc_rqst
498  *	%-EAGAIN if no rpc_rqst is available; queued on backlog
499  */
500 static void
501 xprt_rdma_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
502 {
503 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
504 	struct rpcrdma_req *req;
505 
506 	req = rpcrdma_buffer_get(&r_xprt->rx_buf);
507 	if (!req)
508 		goto out_sleep;
509 	task->tk_rqstp = &req->rl_slot;
510 	task->tk_status = 0;
511 	return;
512 
513 out_sleep:
514 	task->tk_status = -ENOMEM;
515 	xprt_add_backlog(xprt, task);
516 }
517 
518 /**
519  * xprt_rdma_free_slot - release an rpc_rqst
520  * @xprt: controlling RPC transport
521  * @rqst: rpc_rqst to release
522  *
523  */
524 static void
525 xprt_rdma_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *rqst)
526 {
527 	struct rpcrdma_xprt *r_xprt =
528 		container_of(xprt, struct rpcrdma_xprt, rx_xprt);
529 
530 	rpcrdma_reply_put(&r_xprt->rx_buf, rpcr_to_rdmar(rqst));
531 	if (!xprt_wake_up_backlog(xprt, rqst)) {
532 		memset(rqst, 0, sizeof(*rqst));
533 		rpcrdma_buffer_put(&r_xprt->rx_buf, rpcr_to_rdmar(rqst));
534 	}
535 }
536 
537 static bool rpcrdma_check_regbuf(struct rpcrdma_xprt *r_xprt,
538 				 struct rpcrdma_regbuf *rb, size_t size,
539 				 gfp_t flags)
540 {
541 	if (unlikely(rdmab_length(rb) < size)) {
542 		if (!rpcrdma_regbuf_realloc(rb, size, flags))
543 			return false;
544 		r_xprt->rx_stats.hardway_register_count += size;
545 	}
546 	return true;
547 }
548 
549 /**
550  * xprt_rdma_allocate - allocate transport resources for an RPC
551  * @task: RPC task
552  *
553  * Return values:
554  *        0:	Success; rq_buffer points to RPC buffer to use
555  *   ENOMEM:	Out of memory, call again later
556  *      EIO:	A permanent error occurred, do not retry
557  */
558 static int
559 xprt_rdma_allocate(struct rpc_task *task)
560 {
561 	struct rpc_rqst *rqst = task->tk_rqstp;
562 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
563 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
564 	gfp_t flags = rpc_task_gfp_mask();
565 
566 	if (!rpcrdma_check_regbuf(r_xprt, req->rl_sendbuf, rqst->rq_callsize,
567 				  flags))
568 		goto out_fail;
569 	if (!rpcrdma_check_regbuf(r_xprt, req->rl_recvbuf, rqst->rq_rcvsize,
570 				  flags))
571 		goto out_fail;
572 
573 	rqst->rq_buffer = rdmab_data(req->rl_sendbuf);
574 	rqst->rq_rbuffer = rdmab_data(req->rl_recvbuf);
575 	return 0;
576 
577 out_fail:
578 	return -ENOMEM;
579 }
580 
581 /**
582  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
583  * @task: RPC task
584  *
585  * Caller guarantees rqst->rq_buffer is non-NULL.
586  */
587 static void
588 xprt_rdma_free(struct rpc_task *task)
589 {
590 	struct rpc_rqst *rqst = task->tk_rqstp;
591 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
592 
593 	if (unlikely(!list_empty(&req->rl_registered))) {
594 		trace_xprtrdma_mrs_zap(task);
595 		frwr_unmap_sync(rpcx_to_rdmax(rqst->rq_xprt), req);
596 	}
597 
598 	/* XXX: If the RPC is completing because of a signal and
599 	 * not because a reply was received, we ought to ensure
600 	 * that the Send completion has fired, so that memory
601 	 * involved with the Send is not still visible to the NIC.
602 	 */
603 }
604 
605 /**
606  * xprt_rdma_send_request - marshal and send an RPC request
607  * @rqst: RPC message in rq_snd_buf
608  *
609  * Caller holds the transport's write lock.
610  *
611  * Returns:
612  *	%0 if the RPC message has been sent
613  *	%-ENOTCONN if the caller should reconnect and call again
614  *	%-EAGAIN if the caller should call again
615  *	%-ENOBUFS if the caller should call again after a delay
616  *	%-EMSGSIZE if encoding ran out of buffer space. The request
617  *		was not sent. Do not try to send this message again.
618  *	%-EIO if an I/O error occurred. The request was not sent.
619  *		Do not try to send this message again.
620  */
621 static int
622 xprt_rdma_send_request(struct rpc_rqst *rqst)
623 {
624 	struct rpc_xprt *xprt = rqst->rq_xprt;
625 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
626 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
627 	int rc = 0;
628 
629 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
630 	if (unlikely(!rqst->rq_buffer))
631 		return xprt_rdma_bc_send_reply(rqst);
632 #endif	/* CONFIG_SUNRPC_BACKCHANNEL */
633 
634 	if (!xprt_connected(xprt))
635 		return -ENOTCONN;
636 
637 	if (!xprt_request_get_cong(xprt, rqst))
638 		return -EBADSLT;
639 
640 	rc = rpcrdma_marshal_req(r_xprt, rqst);
641 	if (rc < 0)
642 		goto failed_marshal;
643 
644 	/* Must suppress retransmit to maintain credits */
645 	if (rqst->rq_connect_cookie == xprt->connect_cookie)
646 		goto drop_connection;
647 	rqst->rq_xtime = ktime_get();
648 
649 	if (frwr_send(r_xprt, req))
650 		goto drop_connection;
651 
652 	rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
653 
654 	/* An RPC with no reply will throw off credit accounting,
655 	 * so drop the connection to reset the credit grant.
656 	 */
657 	if (!rpc_reply_expected(rqst->rq_task))
658 		goto drop_connection;
659 	return 0;
660 
661 failed_marshal:
662 	if (rc != -ENOTCONN)
663 		return rc;
664 drop_connection:
665 	xprt_rdma_close(xprt);
666 	return -ENOTCONN;
667 }
668 
669 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
670 {
671 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
672 	long idle_time = 0;
673 
674 	if (xprt_connected(xprt))
675 		idle_time = (long)(jiffies - xprt->last_used) / HZ;
676 
677 	seq_puts(seq, "\txprt:\trdma ");
678 	seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
679 		   0,	/* need a local port? */
680 		   xprt->stat.bind_count,
681 		   xprt->stat.connect_count,
682 		   xprt->stat.connect_time / HZ,
683 		   idle_time,
684 		   xprt->stat.sends,
685 		   xprt->stat.recvs,
686 		   xprt->stat.bad_xids,
687 		   xprt->stat.req_u,
688 		   xprt->stat.bklog_u);
689 	seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
690 		   r_xprt->rx_stats.read_chunk_count,
691 		   r_xprt->rx_stats.write_chunk_count,
692 		   r_xprt->rx_stats.reply_chunk_count,
693 		   r_xprt->rx_stats.total_rdma_request,
694 		   r_xprt->rx_stats.total_rdma_reply,
695 		   r_xprt->rx_stats.pullup_copy_count,
696 		   r_xprt->rx_stats.fixup_copy_count,
697 		   r_xprt->rx_stats.hardway_register_count,
698 		   r_xprt->rx_stats.failed_marshal_count,
699 		   r_xprt->rx_stats.bad_reply_count,
700 		   r_xprt->rx_stats.nomsg_call_count);
701 	seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n",
702 		   r_xprt->rx_stats.mrs_recycled,
703 		   r_xprt->rx_stats.mrs_orphaned,
704 		   r_xprt->rx_stats.mrs_allocated,
705 		   r_xprt->rx_stats.local_inv_needed,
706 		   r_xprt->rx_stats.empty_sendctx_q,
707 		   r_xprt->rx_stats.reply_waits_for_send);
708 }
709 
710 static int
711 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
712 {
713 	return 0;
714 }
715 
716 static void
717 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
718 {
719 }
720 
721 /*
722  * Plumbing for rpc transport switch and kernel module
723  */
724 
725 static const struct rpc_xprt_ops xprt_rdma_procs = {
726 	.reserve_xprt		= xprt_reserve_xprt_cong,
727 	.release_xprt		= xprt_release_xprt_cong, /* sunrpc/xprt.c */
728 	.alloc_slot		= xprt_rdma_alloc_slot,
729 	.free_slot		= xprt_rdma_free_slot,
730 	.release_request	= xprt_release_rqst_cong,       /* ditto */
731 	.wait_for_reply_request	= xprt_wait_for_reply_request_def, /* ditto */
732 	.timer			= xprt_rdma_timer,
733 	.rpcbind		= rpcb_getport_async,	/* sunrpc/rpcb_clnt.c */
734 	.set_port		= xprt_rdma_set_port,
735 	.connect		= xprt_rdma_connect,
736 	.buf_alloc		= xprt_rdma_allocate,
737 	.buf_free		= xprt_rdma_free,
738 	.send_request		= xprt_rdma_send_request,
739 	.close			= xprt_rdma_close,
740 	.destroy		= xprt_rdma_destroy,
741 	.set_connect_timeout	= xprt_rdma_set_connect_timeout,
742 	.print_stats		= xprt_rdma_print_stats,
743 	.enable_swap		= xprt_rdma_enable_swap,
744 	.disable_swap		= xprt_rdma_disable_swap,
745 	.inject_disconnect	= xprt_rdma_inject_disconnect,
746 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
747 	.bc_setup		= xprt_rdma_bc_setup,
748 	.bc_maxpayload		= xprt_rdma_bc_maxpayload,
749 	.bc_num_slots		= xprt_rdma_bc_max_slots,
750 	.bc_free_rqst		= xprt_rdma_bc_free_rqst,
751 	.bc_destroy		= xprt_rdma_bc_destroy,
752 #endif
753 };
754 
755 static struct xprt_class xprt_rdma = {
756 	.list			= LIST_HEAD_INIT(xprt_rdma.list),
757 	.name			= "rdma",
758 	.owner			= THIS_MODULE,
759 	.ident			= XPRT_TRANSPORT_RDMA,
760 	.setup			= xprt_setup_rdma,
761 	.netid			= { "rdma", "rdma6", "" },
762 };
763 
764 void xprt_rdma_cleanup(void)
765 {
766 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
767 	if (sunrpc_table_header) {
768 		unregister_sysctl_table(sunrpc_table_header);
769 		sunrpc_table_header = NULL;
770 	}
771 #endif
772 
773 	xprt_unregister_transport(&xprt_rdma);
774 	xprt_unregister_transport(&xprt_rdma_bc);
775 }
776 
777 int xprt_rdma_init(void)
778 {
779 	int rc;
780 
781 	rc = xprt_register_transport(&xprt_rdma);
782 	if (rc)
783 		return rc;
784 
785 	rc = xprt_register_transport(&xprt_rdma_bc);
786 	if (rc) {
787 		xprt_unregister_transport(&xprt_rdma);
788 		return rc;
789 	}
790 
791 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
792 	if (!sunrpc_table_header)
793 		sunrpc_table_header = register_sysctl("sunrpc", xr_tunables_table);
794 #endif
795 	return 0;
796 }
797