xref: /openbmc/linux/net/sunrpc/xprtrdma/transport.c (revision 2209fda3)
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 static 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 zero;
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		= &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_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
245 	xprt_clear_connecting(xprt);
246 	if (r_xprt->rx_ep.rep_connected > 0) {
247 		if (!xprt_test_and_set_connected(xprt)) {
248 			xprt->stat.connect_count++;
249 			xprt->stat.connect_time += (long)jiffies -
250 						   xprt->stat.connect_start;
251 			xprt_wake_pending_tasks(xprt, -EAGAIN);
252 		}
253 	} else {
254 		if (xprt_test_and_clear_connected(xprt))
255 			xprt_wake_pending_tasks(xprt, rc);
256 	}
257 }
258 
259 /**
260  * xprt_rdma_inject_disconnect - inject a connection fault
261  * @xprt: transport context
262  *
263  * If @xprt is connected, disconnect it to simulate spurious connection
264  * loss.
265  */
266 static void
267 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
268 {
269 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
270 
271 	trace_xprtrdma_inject_dsc(r_xprt);
272 	rdma_disconnect(r_xprt->rx_ia.ri_id);
273 }
274 
275 /**
276  * xprt_rdma_destroy - Full tear down of transport
277  * @xprt: doomed transport context
278  *
279  * Caller guarantees there will be no more calls to us with
280  * this @xprt.
281  */
282 static void
283 xprt_rdma_destroy(struct rpc_xprt *xprt)
284 {
285 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
286 
287 	trace_xprtrdma_destroy(r_xprt);
288 
289 	cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
290 
291 	rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
292 	rpcrdma_buffer_destroy(&r_xprt->rx_buf);
293 	rpcrdma_ia_close(&r_xprt->rx_ia);
294 
295 	xprt_rdma_free_addresses(xprt);
296 	xprt_free(xprt);
297 
298 	module_put(THIS_MODULE);
299 }
300 
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 rpcrdma_create_data_internal cdata;
315 	struct rpc_xprt *xprt;
316 	struct rpcrdma_xprt *new_xprt;
317 	struct rpcrdma_ep *new_ep;
318 	struct sockaddr *sap;
319 	int rc;
320 
321 	if (args->addrlen > sizeof(xprt->addr)) {
322 		dprintk("RPC:       %s: address too large\n", __func__);
323 		return ERR_PTR(-EBADF);
324 	}
325 
326 	xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt), 0, 0);
327 	if (xprt == NULL) {
328 		dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
329 			__func__);
330 		return ERR_PTR(-ENOMEM);
331 	}
332 
333 	/* 60 second timeout, no retries */
334 	xprt->timeout = &xprt_rdma_default_timeout;
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->tsh_size = 0;		/* RPC-RDMA handles framing */
341 	xprt->ops = &xprt_rdma_procs;
342 
343 	/*
344 	 * Set up RDMA-specific connect data.
345 	 */
346 	sap = args->dstaddr;
347 
348 	/* Ensure xprt->addr holds valid server TCP (not RDMA)
349 	 * address, for any side protocols which peek at it */
350 	xprt->prot = IPPROTO_TCP;
351 	xprt->addrlen = args->addrlen;
352 	memcpy(&xprt->addr, sap, xprt->addrlen);
353 
354 	if (rpc_get_port(sap))
355 		xprt_set_bound(xprt);
356 	xprt_rdma_format_addresses(xprt, sap);
357 
358 	cdata.max_requests = xprt_rdma_slot_table_entries;
359 
360 	cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
361 	cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
362 
363 	cdata.inline_wsize = xprt_rdma_max_inline_write;
364 	if (cdata.inline_wsize > cdata.wsize)
365 		cdata.inline_wsize = cdata.wsize;
366 
367 	cdata.inline_rsize = xprt_rdma_max_inline_read;
368 	if (cdata.inline_rsize > cdata.rsize)
369 		cdata.inline_rsize = cdata.rsize;
370 
371 	/*
372 	 * Create new transport instance, which includes initialized
373 	 *  o ia
374 	 *  o endpoint
375 	 *  o buffers
376 	 */
377 
378 	new_xprt = rpcx_to_rdmax(xprt);
379 
380 	rc = rpcrdma_ia_open(new_xprt);
381 	if (rc)
382 		goto out1;
383 
384 	/*
385 	 * initialize and create ep
386 	 */
387 	new_xprt->rx_data = cdata;
388 	new_ep = &new_xprt->rx_ep;
389 
390 	rc = rpcrdma_ep_create(&new_xprt->rx_ep,
391 				&new_xprt->rx_ia, &new_xprt->rx_data);
392 	if (rc)
393 		goto out2;
394 
395 	rc = rpcrdma_buffer_create(new_xprt);
396 	if (rc)
397 		goto out3;
398 
399 	INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
400 			  xprt_rdma_connect_worker);
401 
402 	xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
403 	if (xprt->max_payload == 0)
404 		goto out4;
405 	xprt->max_payload <<= PAGE_SHIFT;
406 	dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
407 		__func__, xprt->max_payload);
408 
409 	if (!try_module_get(THIS_MODULE))
410 		goto out4;
411 
412 	dprintk("RPC:       %s: %s:%s\n", __func__,
413 		xprt->address_strings[RPC_DISPLAY_ADDR],
414 		xprt->address_strings[RPC_DISPLAY_PORT]);
415 	trace_xprtrdma_create(new_xprt);
416 	return xprt;
417 
418 out4:
419 	rpcrdma_buffer_destroy(&new_xprt->rx_buf);
420 	rc = -ENODEV;
421 out3:
422 	rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
423 out2:
424 	rpcrdma_ia_close(&new_xprt->rx_ia);
425 out1:
426 	trace_xprtrdma_destroy(new_xprt);
427 	xprt_rdma_free_addresses(xprt);
428 	xprt_free(xprt);
429 	return ERR_PTR(rc);
430 }
431 
432 /**
433  * xprt_rdma_close - close a transport connection
434  * @xprt: transport context
435  *
436  * Called during transport shutdown, reconnect, or device removal.
437  * Caller holds @xprt's send lock to prevent activity on this
438  * transport while the connection is torn down.
439  */
440 static void
441 xprt_rdma_close(struct rpc_xprt *xprt)
442 {
443 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
444 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
445 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
446 
447 	dprintk("RPC:       %s: closing xprt %p\n", __func__, xprt);
448 
449 	if (test_and_clear_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags)) {
450 		xprt_clear_connected(xprt);
451 		rpcrdma_ia_remove(ia);
452 		return;
453 	}
454 	if (ep->rep_connected == -ENODEV)
455 		return;
456 	if (ep->rep_connected > 0)
457 		xprt->reestablish_timeout = 0;
458 	xprt_disconnect_done(xprt);
459 	rpcrdma_ep_disconnect(ep, ia);
460 
461 	/* Prepare @xprt for the next connection by reinitializing
462 	 * its credit grant to one (see RFC 8166, Section 3.3.3).
463 	 */
464 	r_xprt->rx_buf.rb_credits = 1;
465 	xprt->cwnd = RPC_CWNDSHIFT;
466 }
467 
468 /**
469  * xprt_rdma_set_port - update server port with rpcbind result
470  * @xprt: controlling RPC transport
471  * @port: new port value
472  *
473  * Transport connect status is unchanged.
474  */
475 static void
476 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
477 {
478 	struct sockaddr *sap = (struct sockaddr *)&xprt->addr;
479 	char buf[8];
480 
481 	dprintk("RPC:       %s: setting port for xprt %p (%s:%s) to %u\n",
482 		__func__, xprt,
483 		xprt->address_strings[RPC_DISPLAY_ADDR],
484 		xprt->address_strings[RPC_DISPLAY_PORT],
485 		port);
486 
487 	rpc_set_port(sap, port);
488 
489 	kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
490 	snprintf(buf, sizeof(buf), "%u", port);
491 	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
492 
493 	kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]);
494 	snprintf(buf, sizeof(buf), "%4hx", port);
495 	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
496 }
497 
498 /**
499  * xprt_rdma_timer - invoked when an RPC times out
500  * @xprt: controlling RPC transport
501  * @task: RPC task that timed out
502  *
503  * Invoked when the transport is still connected, but an RPC
504  * retransmit timeout occurs.
505  *
506  * Since RDMA connections don't have a keep-alive, forcibly
507  * disconnect and retry to connect. This drives full
508  * detection of the network path, and retransmissions of
509  * all pending RPCs.
510  */
511 static void
512 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
513 {
514 	xprt_force_disconnect(xprt);
515 }
516 
517 /**
518  * xprt_rdma_connect - try to establish a transport connection
519  * @xprt: transport state
520  * @task: RPC scheduler context
521  *
522  */
523 static void
524 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
525 {
526 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
527 
528 	if (r_xprt->rx_ep.rep_connected != 0) {
529 		/* Reconnect */
530 		schedule_delayed_work(&r_xprt->rx_connect_worker,
531 				      xprt->reestablish_timeout);
532 		xprt->reestablish_timeout <<= 1;
533 		if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
534 			xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
535 		else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
536 			xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
537 	} else {
538 		schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
539 		if (!RPC_IS_ASYNC(task))
540 			flush_delayed_work(&r_xprt->rx_connect_worker);
541 	}
542 }
543 
544 /**
545  * xprt_rdma_alloc_slot - allocate an rpc_rqst
546  * @xprt: controlling RPC transport
547  * @task: RPC task requesting a fresh rpc_rqst
548  *
549  * tk_status values:
550  *	%0 if task->tk_rqstp points to a fresh rpc_rqst
551  *	%-EAGAIN if no rpc_rqst is available; queued on backlog
552  */
553 static void
554 xprt_rdma_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
555 {
556 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
557 	struct rpcrdma_req *req;
558 
559 	req = rpcrdma_buffer_get(&r_xprt->rx_buf);
560 	if (!req)
561 		goto out_sleep;
562 	task->tk_rqstp = &req->rl_slot;
563 	task->tk_status = 0;
564 	return;
565 
566 out_sleep:
567 	rpc_sleep_on(&xprt->backlog, task, NULL);
568 	task->tk_status = -EAGAIN;
569 }
570 
571 /**
572  * xprt_rdma_free_slot - release an rpc_rqst
573  * @xprt: controlling RPC transport
574  * @rqst: rpc_rqst to release
575  *
576  */
577 static void
578 xprt_rdma_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *rqst)
579 {
580 	memset(rqst, 0, sizeof(*rqst));
581 	rpcrdma_buffer_put(rpcr_to_rdmar(rqst));
582 	rpc_wake_up_next(&xprt->backlog);
583 }
584 
585 static bool
586 rpcrdma_get_sendbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
587 		    size_t size, gfp_t flags)
588 {
589 	struct rpcrdma_regbuf *rb;
590 
591 	if (req->rl_sendbuf && rdmab_length(req->rl_sendbuf) >= size)
592 		return true;
593 
594 	rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
595 	if (IS_ERR(rb))
596 		return false;
597 
598 	rpcrdma_free_regbuf(req->rl_sendbuf);
599 	r_xprt->rx_stats.hardway_register_count += size;
600 	req->rl_sendbuf = rb;
601 	return true;
602 }
603 
604 /* The rq_rcv_buf is used only if a Reply chunk is necessary.
605  * The decision to use a Reply chunk is made later in
606  * rpcrdma_marshal_req. This buffer is registered at that time.
607  *
608  * Otherwise, the associated RPC Reply arrives in a separate
609  * Receive buffer, arbitrarily chosen by the HCA. The buffer
610  * allocated here for the RPC Reply is not utilized in that
611  * case. See rpcrdma_inline_fixup.
612  *
613  * A regbuf is used here to remember the buffer size.
614  */
615 static bool
616 rpcrdma_get_recvbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
617 		    size_t size, gfp_t flags)
618 {
619 	struct rpcrdma_regbuf *rb;
620 
621 	if (req->rl_recvbuf && rdmab_length(req->rl_recvbuf) >= size)
622 		return true;
623 
624 	rb = rpcrdma_alloc_regbuf(size, DMA_NONE, flags);
625 	if (IS_ERR(rb))
626 		return false;
627 
628 	rpcrdma_free_regbuf(req->rl_recvbuf);
629 	r_xprt->rx_stats.hardway_register_count += size;
630 	req->rl_recvbuf = rb;
631 	return true;
632 }
633 
634 /**
635  * xprt_rdma_allocate - allocate transport resources for an RPC
636  * @task: RPC task
637  *
638  * Return values:
639  *        0:	Success; rq_buffer points to RPC buffer to use
640  *   ENOMEM:	Out of memory, call again later
641  *      EIO:	A permanent error occurred, do not retry
642  */
643 static int
644 xprt_rdma_allocate(struct rpc_task *task)
645 {
646 	struct rpc_rqst *rqst = task->tk_rqstp;
647 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
648 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
649 	gfp_t flags;
650 
651 	flags = RPCRDMA_DEF_GFP;
652 	if (RPC_IS_SWAPPER(task))
653 		flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
654 
655 	if (!rpcrdma_get_sendbuf(r_xprt, req, rqst->rq_callsize, flags))
656 		goto out_fail;
657 	if (!rpcrdma_get_recvbuf(r_xprt, req, rqst->rq_rcvsize, flags))
658 		goto out_fail;
659 
660 	rqst->rq_buffer = req->rl_sendbuf->rg_base;
661 	rqst->rq_rbuffer = req->rl_recvbuf->rg_base;
662 	trace_xprtrdma_allocate(task, req);
663 	return 0;
664 
665 out_fail:
666 	trace_xprtrdma_allocate(task, NULL);
667 	return -ENOMEM;
668 }
669 
670 /**
671  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
672  * @task: RPC task
673  *
674  * Caller guarantees rqst->rq_buffer is non-NULL.
675  */
676 static void
677 xprt_rdma_free(struct rpc_task *task)
678 {
679 	struct rpc_rqst *rqst = task->tk_rqstp;
680 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
681 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
682 
683 	if (test_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags))
684 		rpcrdma_release_rqst(r_xprt, req);
685 	trace_xprtrdma_rpc_done(task, req);
686 }
687 
688 /**
689  * xprt_rdma_send_request - marshal and send an RPC request
690  * @rqst: RPC message in rq_snd_buf
691  *
692  * Caller holds the transport's write lock.
693  *
694  * Returns:
695  *	%0 if the RPC message has been sent
696  *	%-ENOTCONN if the caller should reconnect and call again
697  *	%-EAGAIN if the caller should call again
698  *	%-ENOBUFS if the caller should call again after a delay
699  *	%-EIO if a permanent error occurred and the request was not
700  *		sent. Do not try to send this message again.
701  */
702 static int
703 xprt_rdma_send_request(struct rpc_rqst *rqst)
704 {
705 	struct rpc_xprt *xprt = rqst->rq_xprt;
706 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
707 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
708 	int rc = 0;
709 
710 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
711 	if (unlikely(!rqst->rq_buffer))
712 		return xprt_rdma_bc_send_reply(rqst);
713 #endif	/* CONFIG_SUNRPC_BACKCHANNEL */
714 
715 	if (!xprt_connected(xprt))
716 		goto drop_connection;
717 
718 	if (!xprt_request_get_cong(xprt, rqst))
719 		return -EBADSLT;
720 
721 	rc = rpcrdma_marshal_req(r_xprt, rqst);
722 	if (rc < 0)
723 		goto failed_marshal;
724 
725 	/* Must suppress retransmit to maintain credits */
726 	if (rqst->rq_connect_cookie == xprt->connect_cookie)
727 		goto drop_connection;
728 	rqst->rq_xtime = ktime_get();
729 
730 	__set_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags);
731 	if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
732 		goto drop_connection;
733 
734 	rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
735 	rqst->rq_bytes_sent = 0;
736 
737 	/* An RPC with no reply will throw off credit accounting,
738 	 * so drop the connection to reset the credit grant.
739 	 */
740 	if (!rpc_reply_expected(rqst->rq_task))
741 		goto drop_connection;
742 	return 0;
743 
744 failed_marshal:
745 	if (rc != -ENOTCONN)
746 		return rc;
747 drop_connection:
748 	xprt_disconnect_done(xprt);
749 	return -ENOTCONN;	/* implies disconnect */
750 }
751 
752 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
753 {
754 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
755 	long idle_time = 0;
756 
757 	if (xprt_connected(xprt))
758 		idle_time = (long)(jiffies - xprt->last_used) / HZ;
759 
760 	seq_puts(seq, "\txprt:\trdma ");
761 	seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
762 		   0,	/* need a local port? */
763 		   xprt->stat.bind_count,
764 		   xprt->stat.connect_count,
765 		   xprt->stat.connect_time / HZ,
766 		   idle_time,
767 		   xprt->stat.sends,
768 		   xprt->stat.recvs,
769 		   xprt->stat.bad_xids,
770 		   xprt->stat.req_u,
771 		   xprt->stat.bklog_u);
772 	seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
773 		   r_xprt->rx_stats.read_chunk_count,
774 		   r_xprt->rx_stats.write_chunk_count,
775 		   r_xprt->rx_stats.reply_chunk_count,
776 		   r_xprt->rx_stats.total_rdma_request,
777 		   r_xprt->rx_stats.total_rdma_reply,
778 		   r_xprt->rx_stats.pullup_copy_count,
779 		   r_xprt->rx_stats.fixup_copy_count,
780 		   r_xprt->rx_stats.hardway_register_count,
781 		   r_xprt->rx_stats.failed_marshal_count,
782 		   r_xprt->rx_stats.bad_reply_count,
783 		   r_xprt->rx_stats.nomsg_call_count);
784 	seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n",
785 		   r_xprt->rx_stats.mrs_recycled,
786 		   r_xprt->rx_stats.mrs_orphaned,
787 		   r_xprt->rx_stats.mrs_allocated,
788 		   r_xprt->rx_stats.local_inv_needed,
789 		   r_xprt->rx_stats.empty_sendctx_q,
790 		   r_xprt->rx_stats.reply_waits_for_send);
791 }
792 
793 static int
794 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
795 {
796 	return 0;
797 }
798 
799 static void
800 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
801 {
802 }
803 
804 /*
805  * Plumbing for rpc transport switch and kernel module
806  */
807 
808 static const struct rpc_xprt_ops xprt_rdma_procs = {
809 	.reserve_xprt		= xprt_reserve_xprt_cong,
810 	.release_xprt		= xprt_release_xprt_cong, /* sunrpc/xprt.c */
811 	.alloc_slot		= xprt_rdma_alloc_slot,
812 	.free_slot		= xprt_rdma_free_slot,
813 	.release_request	= xprt_release_rqst_cong,       /* ditto */
814 	.set_retrans_timeout	= xprt_set_retrans_timeout_def, /* ditto */
815 	.timer			= xprt_rdma_timer,
816 	.rpcbind		= rpcb_getport_async,	/* sunrpc/rpcb_clnt.c */
817 	.set_port		= xprt_rdma_set_port,
818 	.connect		= xprt_rdma_connect,
819 	.buf_alloc		= xprt_rdma_allocate,
820 	.buf_free		= xprt_rdma_free,
821 	.send_request		= xprt_rdma_send_request,
822 	.close			= xprt_rdma_close,
823 	.destroy		= xprt_rdma_destroy,
824 	.print_stats		= xprt_rdma_print_stats,
825 	.enable_swap		= xprt_rdma_enable_swap,
826 	.disable_swap		= xprt_rdma_disable_swap,
827 	.inject_disconnect	= xprt_rdma_inject_disconnect,
828 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
829 	.bc_setup		= xprt_rdma_bc_setup,
830 	.bc_up			= xprt_rdma_bc_up,
831 	.bc_maxpayload		= xprt_rdma_bc_maxpayload,
832 	.bc_free_rqst		= xprt_rdma_bc_free_rqst,
833 	.bc_destroy		= xprt_rdma_bc_destroy,
834 #endif
835 };
836 
837 static struct xprt_class xprt_rdma = {
838 	.list			= LIST_HEAD_INIT(xprt_rdma.list),
839 	.name			= "rdma",
840 	.owner			= THIS_MODULE,
841 	.ident			= XPRT_TRANSPORT_RDMA,
842 	.setup			= xprt_setup_rdma,
843 };
844 
845 void xprt_rdma_cleanup(void)
846 {
847 	int rc;
848 
849 	dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
850 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
851 	if (sunrpc_table_header) {
852 		unregister_sysctl_table(sunrpc_table_header);
853 		sunrpc_table_header = NULL;
854 	}
855 #endif
856 	rc = xprt_unregister_transport(&xprt_rdma);
857 	if (rc)
858 		dprintk("RPC:       %s: xprt_unregister returned %i\n",
859 			__func__, rc);
860 
861 	rpcrdma_destroy_wq();
862 
863 	rc = xprt_unregister_transport(&xprt_rdma_bc);
864 	if (rc)
865 		dprintk("RPC:       %s: xprt_unregister(bc) returned %i\n",
866 			__func__, rc);
867 }
868 
869 int xprt_rdma_init(void)
870 {
871 	int rc;
872 
873 	rc = rpcrdma_alloc_wq();
874 	if (rc)
875 		return rc;
876 
877 	rc = xprt_register_transport(&xprt_rdma);
878 	if (rc) {
879 		rpcrdma_destroy_wq();
880 		return rc;
881 	}
882 
883 	rc = xprt_register_transport(&xprt_rdma_bc);
884 	if (rc) {
885 		xprt_unregister_transport(&xprt_rdma);
886 		rpcrdma_destroy_wq();
887 		return rc;
888 	}
889 
890 	dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
891 
892 	dprintk("Defaults:\n");
893 	dprintk("\tSlots %d\n"
894 		"\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
895 		xprt_rdma_slot_table_entries,
896 		xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
897 	dprintk("\tPadding 0\n\tMemreg %d\n", xprt_rdma_memreg_strategy);
898 
899 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
900 	if (!sunrpc_table_header)
901 		sunrpc_table_header = register_sysctl_table(sunrpc_table);
902 #endif
903 	return 0;
904 }
905