xref: /linux/net/sunrpc/xprtrdma/transport.c (revision c0e297dc61f8d4453e07afbea1fa8d0e67cd4a34)
1 /*
2  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the BSD-type
8  * license below:
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  *
14  *      Redistributions of source code must retain the above copyright
15  *      notice, this list of conditions and the following disclaimer.
16  *
17  *      Redistributions in binary form must reproduce the above
18  *      copyright notice, this list of conditions and the following
19  *      disclaimer in the documentation and/or other materials provided
20  *      with the distribution.
21  *
22  *      Neither the name of the Network Appliance, Inc. nor the names of
23  *      its contributors may be used to endorse or promote products
24  *      derived from this software without specific prior written
25  *      permission.
26  *
27  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
28  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
29  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
30  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
31  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
32  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
33  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
34  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
35  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
36  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
37  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
38  */
39 
40 /*
41  * transport.c
42  *
43  * This file contains the top-level implementation of an RPC RDMA
44  * transport.
45  *
46  * Naming convention: functions beginning with xprt_ are part of the
47  * transport switch. All others are RPC RDMA internal.
48  */
49 
50 #include <linux/module.h>
51 #include <linux/slab.h>
52 #include <linux/seq_file.h>
53 #include <linux/sunrpc/addr.h>
54 
55 #include "xprt_rdma.h"
56 
57 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
58 # define RPCDBG_FACILITY	RPCDBG_TRANS
59 #endif
60 
61 /*
62  * tunables
63  */
64 
65 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
66 static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
67 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
68 static unsigned int xprt_rdma_inline_write_padding;
69 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
70 		int xprt_rdma_pad_optimize = 1;
71 
72 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
73 
74 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
75 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
76 static unsigned int zero;
77 static unsigned int max_padding = PAGE_SIZE;
78 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
79 static unsigned int max_memreg = RPCRDMA_LAST - 1;
80 
81 static struct ctl_table_header *sunrpc_table_header;
82 
83 static struct ctl_table xr_tunables_table[] = {
84 	{
85 		.procname	= "rdma_slot_table_entries",
86 		.data		= &xprt_rdma_slot_table_entries,
87 		.maxlen		= sizeof(unsigned int),
88 		.mode		= 0644,
89 		.proc_handler	= proc_dointvec_minmax,
90 		.extra1		= &min_slot_table_size,
91 		.extra2		= &max_slot_table_size
92 	},
93 	{
94 		.procname	= "rdma_max_inline_read",
95 		.data		= &xprt_rdma_max_inline_read,
96 		.maxlen		= sizeof(unsigned int),
97 		.mode		= 0644,
98 		.proc_handler	= proc_dointvec,
99 	},
100 	{
101 		.procname	= "rdma_max_inline_write",
102 		.data		= &xprt_rdma_max_inline_write,
103 		.maxlen		= sizeof(unsigned int),
104 		.mode		= 0644,
105 		.proc_handler	= proc_dointvec,
106 	},
107 	{
108 		.procname	= "rdma_inline_write_padding",
109 		.data		= &xprt_rdma_inline_write_padding,
110 		.maxlen		= sizeof(unsigned int),
111 		.mode		= 0644,
112 		.proc_handler	= proc_dointvec_minmax,
113 		.extra1		= &zero,
114 		.extra2		= &max_padding,
115 	},
116 	{
117 		.procname	= "rdma_memreg_strategy",
118 		.data		= &xprt_rdma_memreg_strategy,
119 		.maxlen		= sizeof(unsigned int),
120 		.mode		= 0644,
121 		.proc_handler	= proc_dointvec_minmax,
122 		.extra1		= &min_memreg,
123 		.extra2		= &max_memreg,
124 	},
125 	{
126 		.procname	= "rdma_pad_optimize",
127 		.data		= &xprt_rdma_pad_optimize,
128 		.maxlen		= sizeof(unsigned int),
129 		.mode		= 0644,
130 		.proc_handler	= proc_dointvec,
131 	},
132 	{ },
133 };
134 
135 static struct ctl_table sunrpc_table[] = {
136 	{
137 		.procname	= "sunrpc",
138 		.mode		= 0555,
139 		.child		= xr_tunables_table
140 	},
141 	{ },
142 };
143 
144 #endif
145 
146 #define RPCRDMA_BIND_TO		(60U * HZ)
147 #define RPCRDMA_INIT_REEST_TO	(5U * HZ)
148 #define RPCRDMA_MAX_REEST_TO	(30U * HZ)
149 #define RPCRDMA_IDLE_DISC_TO	(5U * 60 * HZ)
150 
151 static struct rpc_xprt_ops xprt_rdma_procs;	/* forward reference */
152 
153 static void
154 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
155 {
156 	struct sockaddr_in *sin = (struct sockaddr_in *)sap;
157 	char buf[20];
158 
159 	snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
160 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
161 
162 	xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
163 }
164 
165 static void
166 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
167 {
168 	struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
169 	char buf[40];
170 
171 	snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
172 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
173 
174 	xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
175 }
176 
177 static void
178 xprt_rdma_format_addresses(struct rpc_xprt *xprt)
179 {
180 	struct sockaddr *sap = (struct sockaddr *)
181 					&rpcx_to_rdmad(xprt).addr;
182 	char buf[128];
183 
184 	switch (sap->sa_family) {
185 	case AF_INET:
186 		xprt_rdma_format_addresses4(xprt, sap);
187 		break;
188 	case AF_INET6:
189 		xprt_rdma_format_addresses6(xprt, sap);
190 		break;
191 	default:
192 		pr_err("rpcrdma: Unrecognized address family\n");
193 		return;
194 	}
195 
196 	(void)rpc_ntop(sap, buf, sizeof(buf));
197 	xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
198 
199 	snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
200 	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
201 
202 	snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
203 	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
204 
205 	xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
206 }
207 
208 static void
209 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
210 {
211 	unsigned int i;
212 
213 	for (i = 0; i < RPC_DISPLAY_MAX; i++)
214 		switch (i) {
215 		case RPC_DISPLAY_PROTO:
216 		case RPC_DISPLAY_NETID:
217 			continue;
218 		default:
219 			kfree(xprt->address_strings[i]);
220 		}
221 }
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 	int rc = 0;
230 
231 	xprt_clear_connected(xprt);
232 
233 	dprintk("RPC:       %s: %sconnect\n", __func__,
234 			r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
235 	rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
236 	if (rc)
237 		xprt_wake_pending_tasks(xprt, rc);
238 
239 	dprintk("RPC:       %s: exit\n", __func__);
240 	xprt_clear_connecting(xprt);
241 }
242 
243 static void
244 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
245 {
246 	struct rpcrdma_xprt *r_xprt = container_of(xprt, struct rpcrdma_xprt,
247 						   rx_xprt);
248 
249 	pr_info("rpcrdma: injecting transport disconnect on xprt=%p\n", xprt);
250 	rdma_disconnect(r_xprt->rx_ia.ri_id);
251 }
252 
253 /*
254  * xprt_rdma_destroy
255  *
256  * Destroy the xprt.
257  * Free all memory associated with the object, including its own.
258  * NOTE: none of the *destroy methods free memory for their top-level
259  * objects, even though they may have allocated it (they do free
260  * private memory). It's up to the caller to handle it. In this
261  * case (RDMA transport), all structure memory is inlined with the
262  * struct rpcrdma_xprt.
263  */
264 static void
265 xprt_rdma_destroy(struct rpc_xprt *xprt)
266 {
267 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
268 
269 	dprintk("RPC:       %s: called\n", __func__);
270 
271 	cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
272 
273 	xprt_clear_connected(xprt);
274 
275 	rpcrdma_buffer_destroy(&r_xprt->rx_buf);
276 	rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
277 	rpcrdma_ia_close(&r_xprt->rx_ia);
278 
279 	xprt_rdma_free_addresses(xprt);
280 
281 	xprt_free(xprt);
282 
283 	dprintk("RPC:       %s: returning\n", __func__);
284 
285 	module_put(THIS_MODULE);
286 }
287 
288 static const struct rpc_timeout xprt_rdma_default_timeout = {
289 	.to_initval = 60 * HZ,
290 	.to_maxval = 60 * HZ,
291 };
292 
293 /**
294  * xprt_setup_rdma - Set up transport to use RDMA
295  *
296  * @args: rpc transport arguments
297  */
298 static struct rpc_xprt *
299 xprt_setup_rdma(struct xprt_create *args)
300 {
301 	struct rpcrdma_create_data_internal cdata;
302 	struct rpc_xprt *xprt;
303 	struct rpcrdma_xprt *new_xprt;
304 	struct rpcrdma_ep *new_ep;
305 	struct sockaddr_in *sin;
306 	int rc;
307 
308 	if (args->addrlen > sizeof(xprt->addr)) {
309 		dprintk("RPC:       %s: address too large\n", __func__);
310 		return ERR_PTR(-EBADF);
311 	}
312 
313 	xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
314 			xprt_rdma_slot_table_entries,
315 			xprt_rdma_slot_table_entries);
316 	if (xprt == NULL) {
317 		dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
318 			__func__);
319 		return ERR_PTR(-ENOMEM);
320 	}
321 
322 	/* 60 second timeout, no retries */
323 	xprt->timeout = &xprt_rdma_default_timeout;
324 	xprt->bind_timeout = RPCRDMA_BIND_TO;
325 	xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
326 	xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
327 
328 	xprt->resvport = 0;		/* privileged port not needed */
329 	xprt->tsh_size = 0;		/* RPC-RDMA handles framing */
330 	xprt->ops = &xprt_rdma_procs;
331 
332 	/*
333 	 * Set up RDMA-specific connect data.
334 	 */
335 
336 	/* Put server RDMA address in local cdata */
337 	memcpy(&cdata.addr, args->dstaddr, args->addrlen);
338 
339 	/* Ensure xprt->addr holds valid server TCP (not RDMA)
340 	 * address, for any side protocols which peek at it */
341 	xprt->prot = IPPROTO_TCP;
342 	xprt->addrlen = args->addrlen;
343 	memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
344 
345 	sin = (struct sockaddr_in *)&cdata.addr;
346 	if (ntohs(sin->sin_port) != 0)
347 		xprt_set_bound(xprt);
348 
349 	dprintk("RPC:       %s: %pI4:%u\n",
350 		__func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port));
351 
352 	/* Set max requests */
353 	cdata.max_requests = xprt->max_reqs;
354 
355 	/* Set some length limits */
356 	cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
357 	cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
358 
359 	cdata.inline_wsize = xprt_rdma_max_inline_write;
360 	if (cdata.inline_wsize > cdata.wsize)
361 		cdata.inline_wsize = cdata.wsize;
362 
363 	cdata.inline_rsize = xprt_rdma_max_inline_read;
364 	if (cdata.inline_rsize > cdata.rsize)
365 		cdata.inline_rsize = cdata.rsize;
366 
367 	cdata.padding = xprt_rdma_inline_write_padding;
368 
369 	/*
370 	 * Create new transport instance, which includes initialized
371 	 *  o ia
372 	 *  o endpoint
373 	 *  o buffers
374 	 */
375 
376 	new_xprt = rpcx_to_rdmax(xprt);
377 
378 	rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
379 				xprt_rdma_memreg_strategy);
380 	if (rc)
381 		goto out1;
382 
383 	/*
384 	 * initialize and create ep
385 	 */
386 	new_xprt->rx_data = cdata;
387 	new_ep = &new_xprt->rx_ep;
388 	new_ep->rep_remote_addr = cdata.addr;
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 	/*
396 	 * Allocate pre-registered send and receive buffers for headers and
397 	 * any inline data. Also specify any padding which will be provided
398 	 * from a preregistered zero buffer.
399 	 */
400 	rc = rpcrdma_buffer_create(new_xprt);
401 	if (rc)
402 		goto out3;
403 
404 	/*
405 	 * Register a callback for connection events. This is necessary because
406 	 * connection loss notification is async. We also catch connection loss
407 	 * when reaping receives.
408 	 */
409 	INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
410 			  xprt_rdma_connect_worker);
411 
412 	xprt_rdma_format_addresses(xprt);
413 	xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
414 	if (xprt->max_payload == 0)
415 		goto out4;
416 	xprt->max_payload <<= PAGE_SHIFT;
417 	dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
418 		__func__, xprt->max_payload);
419 
420 	if (!try_module_get(THIS_MODULE))
421 		goto out4;
422 
423 	return xprt;
424 
425 out4:
426 	xprt_rdma_free_addresses(xprt);
427 	rc = -EINVAL;
428 out3:
429 	rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
430 out2:
431 	rpcrdma_ia_close(&new_xprt->rx_ia);
432 out1:
433 	xprt_free(xprt);
434 	return ERR_PTR(rc);
435 }
436 
437 /*
438  * Close a connection, during shutdown or timeout/reconnect
439  */
440 static void
441 xprt_rdma_close(struct rpc_xprt *xprt)
442 {
443 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
444 
445 	dprintk("RPC:       %s: closing\n", __func__);
446 	if (r_xprt->rx_ep.rep_connected > 0)
447 		xprt->reestablish_timeout = 0;
448 	xprt_disconnect_done(xprt);
449 	rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
450 }
451 
452 static void
453 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
454 {
455 	struct sockaddr_in *sap;
456 
457 	sap = (struct sockaddr_in *)&xprt->addr;
458 	sap->sin_port = htons(port);
459 	sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
460 	sap->sin_port = htons(port);
461 	dprintk("RPC:       %s: %u\n", __func__, port);
462 }
463 
464 static void
465 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
466 {
467 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
468 
469 	if (r_xprt->rx_ep.rep_connected != 0) {
470 		/* Reconnect */
471 		schedule_delayed_work(&r_xprt->rx_connect_worker,
472 				      xprt->reestablish_timeout);
473 		xprt->reestablish_timeout <<= 1;
474 		if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
475 			xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
476 		else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
477 			xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
478 	} else {
479 		schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
480 		if (!RPC_IS_ASYNC(task))
481 			flush_delayed_work(&r_xprt->rx_connect_worker);
482 	}
483 }
484 
485 /*
486  * The RDMA allocate/free functions need the task structure as a place
487  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
488  * sequence.
489  *
490  * The RPC layer allocates both send and receive buffers in the same call
491  * (rq_send_buf and rq_rcv_buf are both part of a single contiguous buffer).
492  * We may register rq_rcv_buf when using reply chunks.
493  */
494 static void *
495 xprt_rdma_allocate(struct rpc_task *task, size_t size)
496 {
497 	struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
498 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
499 	struct rpcrdma_regbuf *rb;
500 	struct rpcrdma_req *req;
501 	size_t min_size;
502 	gfp_t flags;
503 
504 	req = rpcrdma_buffer_get(&r_xprt->rx_buf);
505 	if (req == NULL)
506 		return NULL;
507 
508 	flags = GFP_NOIO | __GFP_NOWARN;
509 	if (RPC_IS_SWAPPER(task))
510 		flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
511 
512 	if (req->rl_rdmabuf == NULL)
513 		goto out_rdmabuf;
514 	if (req->rl_sendbuf == NULL)
515 		goto out_sendbuf;
516 	if (size > req->rl_sendbuf->rg_size)
517 		goto out_sendbuf;
518 
519 out:
520 	dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
521 	req->rl_connect_cookie = 0;	/* our reserved value */
522 	return req->rl_sendbuf->rg_base;
523 
524 out_rdmabuf:
525 	min_size = RPCRDMA_INLINE_WRITE_THRESHOLD(task->tk_rqstp);
526 	rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, min_size, flags);
527 	if (IS_ERR(rb))
528 		goto out_fail;
529 	req->rl_rdmabuf = rb;
530 
531 out_sendbuf:
532 	/* XDR encoding and RPC/RDMA marshaling of this request has not
533 	 * yet occurred. Thus a lower bound is needed to prevent buffer
534 	 * overrun during marshaling.
535 	 *
536 	 * RPC/RDMA marshaling may choose to send payload bearing ops
537 	 * inline, if the result is smaller than the inline threshold.
538 	 * The value of the "size" argument accounts for header
539 	 * requirements but not for the payload in these cases.
540 	 *
541 	 * Likewise, allocate enough space to receive a reply up to the
542 	 * size of the inline threshold.
543 	 *
544 	 * It's unlikely that both the send header and the received
545 	 * reply will be large, but slush is provided here to allow
546 	 * flexibility when marshaling.
547 	 */
548 	min_size = RPCRDMA_INLINE_READ_THRESHOLD(task->tk_rqstp);
549 	min_size += RPCRDMA_INLINE_WRITE_THRESHOLD(task->tk_rqstp);
550 	if (size < min_size)
551 		size = min_size;
552 
553 	rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, size, flags);
554 	if (IS_ERR(rb))
555 		goto out_fail;
556 	rb->rg_owner = req;
557 
558 	r_xprt->rx_stats.hardway_register_count += size;
559 	rpcrdma_free_regbuf(&r_xprt->rx_ia, req->rl_sendbuf);
560 	req->rl_sendbuf = rb;
561 	goto out;
562 
563 out_fail:
564 	rpcrdma_buffer_put(req);
565 	r_xprt->rx_stats.failed_marshal_count++;
566 	return NULL;
567 }
568 
569 /*
570  * This function returns all RDMA resources to the pool.
571  */
572 static void
573 xprt_rdma_free(void *buffer)
574 {
575 	struct rpcrdma_req *req;
576 	struct rpcrdma_xprt *r_xprt;
577 	struct rpcrdma_regbuf *rb;
578 	int i;
579 
580 	if (buffer == NULL)
581 		return;
582 
583 	rb = container_of(buffer, struct rpcrdma_regbuf, rg_base[0]);
584 	req = rb->rg_owner;
585 	r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
586 
587 	dprintk("RPC:       %s: called on 0x%p\n", __func__, req->rl_reply);
588 
589 	for (i = 0; req->rl_nchunks;) {
590 		--req->rl_nchunks;
591 		i += r_xprt->rx_ia.ri_ops->ro_unmap(r_xprt,
592 						    &req->rl_segments[i]);
593 	}
594 
595 	rpcrdma_buffer_put(req);
596 }
597 
598 /*
599  * send_request invokes the meat of RPC RDMA. It must do the following:
600  *  1.  Marshal the RPC request into an RPC RDMA request, which means
601  *	putting a header in front of data, and creating IOVs for RDMA
602  *	from those in the request.
603  *  2.  In marshaling, detect opportunities for RDMA, and use them.
604  *  3.  Post a recv message to set up asynch completion, then send
605  *	the request (rpcrdma_ep_post).
606  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
607  */
608 
609 static int
610 xprt_rdma_send_request(struct rpc_task *task)
611 {
612 	struct rpc_rqst *rqst = task->tk_rqstp;
613 	struct rpc_xprt *xprt = rqst->rq_xprt;
614 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
615 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
616 	int rc = 0;
617 
618 	rc = rpcrdma_marshal_req(rqst);
619 	if (rc < 0)
620 		goto failed_marshal;
621 
622 	if (req->rl_reply == NULL) 		/* e.g. reconnection */
623 		rpcrdma_recv_buffer_get(req);
624 
625 	/* Must suppress retransmit to maintain credits */
626 	if (req->rl_connect_cookie == xprt->connect_cookie)
627 		goto drop_connection;
628 	req->rl_connect_cookie = xprt->connect_cookie;
629 
630 	if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
631 		goto drop_connection;
632 
633 	rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
634 	rqst->rq_bytes_sent = 0;
635 	return 0;
636 
637 failed_marshal:
638 	r_xprt->rx_stats.failed_marshal_count++;
639 	dprintk("RPC:       %s: rpcrdma_marshal_req failed, status %i\n",
640 		__func__, rc);
641 	if (rc == -EIO)
642 		return -EIO;
643 drop_connection:
644 	xprt_disconnect_done(xprt);
645 	return -ENOTCONN;	/* implies disconnect */
646 }
647 
648 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
649 {
650 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
651 	long idle_time = 0;
652 
653 	if (xprt_connected(xprt))
654 		idle_time = (long)(jiffies - xprt->last_used) / HZ;
655 
656 	seq_printf(seq,
657 	  "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
658 	  "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
659 
660 	   0,	/* need a local port? */
661 	   xprt->stat.bind_count,
662 	   xprt->stat.connect_count,
663 	   xprt->stat.connect_time,
664 	   idle_time,
665 	   xprt->stat.sends,
666 	   xprt->stat.recvs,
667 	   xprt->stat.bad_xids,
668 	   xprt->stat.req_u,
669 	   xprt->stat.bklog_u,
670 
671 	   r_xprt->rx_stats.read_chunk_count,
672 	   r_xprt->rx_stats.write_chunk_count,
673 	   r_xprt->rx_stats.reply_chunk_count,
674 	   r_xprt->rx_stats.total_rdma_request,
675 	   r_xprt->rx_stats.total_rdma_reply,
676 	   r_xprt->rx_stats.pullup_copy_count,
677 	   r_xprt->rx_stats.fixup_copy_count,
678 	   r_xprt->rx_stats.hardway_register_count,
679 	   r_xprt->rx_stats.failed_marshal_count,
680 	   r_xprt->rx_stats.bad_reply_count);
681 }
682 
683 static int
684 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
685 {
686 	return -EINVAL;
687 }
688 
689 static void
690 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
691 {
692 }
693 
694 /*
695  * Plumbing for rpc transport switch and kernel module
696  */
697 
698 static struct rpc_xprt_ops xprt_rdma_procs = {
699 	.reserve_xprt		= xprt_reserve_xprt_cong,
700 	.release_xprt		= xprt_release_xprt_cong, /* sunrpc/xprt.c */
701 	.alloc_slot		= xprt_alloc_slot,
702 	.release_request	= xprt_release_rqst_cong,       /* ditto */
703 	.set_retrans_timeout	= xprt_set_retrans_timeout_def, /* ditto */
704 	.rpcbind		= rpcb_getport_async,	/* sunrpc/rpcb_clnt.c */
705 	.set_port		= xprt_rdma_set_port,
706 	.connect		= xprt_rdma_connect,
707 	.buf_alloc		= xprt_rdma_allocate,
708 	.buf_free		= xprt_rdma_free,
709 	.send_request		= xprt_rdma_send_request,
710 	.close			= xprt_rdma_close,
711 	.destroy		= xprt_rdma_destroy,
712 	.print_stats		= xprt_rdma_print_stats,
713 	.enable_swap		= xprt_rdma_enable_swap,
714 	.disable_swap		= xprt_rdma_disable_swap,
715 	.inject_disconnect	= xprt_rdma_inject_disconnect
716 };
717 
718 static struct xprt_class xprt_rdma = {
719 	.list			= LIST_HEAD_INIT(xprt_rdma.list),
720 	.name			= "rdma",
721 	.owner			= THIS_MODULE,
722 	.ident			= XPRT_TRANSPORT_RDMA,
723 	.setup			= xprt_setup_rdma,
724 };
725 
726 void xprt_rdma_cleanup(void)
727 {
728 	int rc;
729 
730 	dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
731 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
732 	if (sunrpc_table_header) {
733 		unregister_sysctl_table(sunrpc_table_header);
734 		sunrpc_table_header = NULL;
735 	}
736 #endif
737 	rc = xprt_unregister_transport(&xprt_rdma);
738 	if (rc)
739 		dprintk("RPC:       %s: xprt_unregister returned %i\n",
740 			__func__, rc);
741 
742 	frwr_destroy_recovery_wq();
743 }
744 
745 int xprt_rdma_init(void)
746 {
747 	int rc;
748 
749 	rc = frwr_alloc_recovery_wq();
750 	if (rc)
751 		return rc;
752 
753 	rc = xprt_register_transport(&xprt_rdma);
754 	if (rc) {
755 		frwr_destroy_recovery_wq();
756 		return rc;
757 	}
758 
759 	dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
760 
761 	dprintk("Defaults:\n");
762 	dprintk("\tSlots %d\n"
763 		"\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
764 		xprt_rdma_slot_table_entries,
765 		xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
766 	dprintk("\tPadding %d\n\tMemreg %d\n",
767 		xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
768 
769 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
770 	if (!sunrpc_table_header)
771 		sunrpc_table_header = register_sysctl_table(sunrpc_table);
772 #endif
773 	return 0;
774 }
775