xref: /linux/net/sunrpc/xprtrdma/transport.c (revision 3e44c471a2dab210f7e9b1e5f7d4d54d52df59eb)
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 /*
244  * xprt_rdma_destroy
245  *
246  * Destroy the xprt.
247  * Free all memory associated with the object, including its own.
248  * NOTE: none of the *destroy methods free memory for their top-level
249  * objects, even though they may have allocated it (they do free
250  * private memory). It's up to the caller to handle it. In this
251  * case (RDMA transport), all structure memory is inlined with the
252  * struct rpcrdma_xprt.
253  */
254 static void
255 xprt_rdma_destroy(struct rpc_xprt *xprt)
256 {
257 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
258 
259 	dprintk("RPC:       %s: called\n", __func__);
260 
261 	cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
262 
263 	xprt_clear_connected(xprt);
264 
265 	rpcrdma_buffer_destroy(&r_xprt->rx_buf);
266 	rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
267 	rpcrdma_ia_close(&r_xprt->rx_ia);
268 
269 	xprt_rdma_free_addresses(xprt);
270 
271 	xprt_free(xprt);
272 
273 	dprintk("RPC:       %s: returning\n", __func__);
274 
275 	module_put(THIS_MODULE);
276 }
277 
278 static const struct rpc_timeout xprt_rdma_default_timeout = {
279 	.to_initval = 60 * HZ,
280 	.to_maxval = 60 * HZ,
281 };
282 
283 /**
284  * xprt_setup_rdma - Set up transport to use RDMA
285  *
286  * @args: rpc transport arguments
287  */
288 static struct rpc_xprt *
289 xprt_setup_rdma(struct xprt_create *args)
290 {
291 	struct rpcrdma_create_data_internal cdata;
292 	struct rpc_xprt *xprt;
293 	struct rpcrdma_xprt *new_xprt;
294 	struct rpcrdma_ep *new_ep;
295 	struct sockaddr_in *sin;
296 	int rc;
297 
298 	if (args->addrlen > sizeof(xprt->addr)) {
299 		dprintk("RPC:       %s: address too large\n", __func__);
300 		return ERR_PTR(-EBADF);
301 	}
302 
303 	xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
304 			xprt_rdma_slot_table_entries,
305 			xprt_rdma_slot_table_entries);
306 	if (xprt == NULL) {
307 		dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
308 			__func__);
309 		return ERR_PTR(-ENOMEM);
310 	}
311 
312 	/* 60 second timeout, no retries */
313 	xprt->timeout = &xprt_rdma_default_timeout;
314 	xprt->bind_timeout = RPCRDMA_BIND_TO;
315 	xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
316 	xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
317 
318 	xprt->resvport = 0;		/* privileged port not needed */
319 	xprt->tsh_size = 0;		/* RPC-RDMA handles framing */
320 	xprt->ops = &xprt_rdma_procs;
321 
322 	/*
323 	 * Set up RDMA-specific connect data.
324 	 */
325 
326 	/* Put server RDMA address in local cdata */
327 	memcpy(&cdata.addr, args->dstaddr, args->addrlen);
328 
329 	/* Ensure xprt->addr holds valid server TCP (not RDMA)
330 	 * address, for any side protocols which peek at it */
331 	xprt->prot = IPPROTO_TCP;
332 	xprt->addrlen = args->addrlen;
333 	memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
334 
335 	sin = (struct sockaddr_in *)&cdata.addr;
336 	if (ntohs(sin->sin_port) != 0)
337 		xprt_set_bound(xprt);
338 
339 	dprintk("RPC:       %s: %pI4:%u\n",
340 		__func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port));
341 
342 	/* Set max requests */
343 	cdata.max_requests = xprt->max_reqs;
344 
345 	/* Set some length limits */
346 	cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
347 	cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
348 
349 	cdata.inline_wsize = xprt_rdma_max_inline_write;
350 	if (cdata.inline_wsize > cdata.wsize)
351 		cdata.inline_wsize = cdata.wsize;
352 
353 	cdata.inline_rsize = xprt_rdma_max_inline_read;
354 	if (cdata.inline_rsize > cdata.rsize)
355 		cdata.inline_rsize = cdata.rsize;
356 
357 	cdata.padding = xprt_rdma_inline_write_padding;
358 
359 	/*
360 	 * Create new transport instance, which includes initialized
361 	 *  o ia
362 	 *  o endpoint
363 	 *  o buffers
364 	 */
365 
366 	new_xprt = rpcx_to_rdmax(xprt);
367 
368 	rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
369 				xprt_rdma_memreg_strategy);
370 	if (rc)
371 		goto out1;
372 
373 	/*
374 	 * initialize and create ep
375 	 */
376 	new_xprt->rx_data = cdata;
377 	new_ep = &new_xprt->rx_ep;
378 	new_ep->rep_remote_addr = cdata.addr;
379 
380 	rc = rpcrdma_ep_create(&new_xprt->rx_ep,
381 				&new_xprt->rx_ia, &new_xprt->rx_data);
382 	if (rc)
383 		goto out2;
384 
385 	/*
386 	 * Allocate pre-registered send and receive buffers for headers and
387 	 * any inline data. Also specify any padding which will be provided
388 	 * from a preregistered zero buffer.
389 	 */
390 	rc = rpcrdma_buffer_create(new_xprt);
391 	if (rc)
392 		goto out3;
393 
394 	/*
395 	 * Register a callback for connection events. This is necessary because
396 	 * connection loss notification is async. We also catch connection loss
397 	 * when reaping receives.
398 	 */
399 	INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
400 			  xprt_rdma_connect_worker);
401 
402 	xprt_rdma_format_addresses(xprt);
403 	xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
404 	if (xprt->max_payload == 0)
405 		goto out4;
406 	xprt->max_payload <<= PAGE_SHIFT;
407 	dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
408 		__func__, xprt->max_payload);
409 
410 	if (!try_module_get(THIS_MODULE))
411 		goto out4;
412 
413 	return xprt;
414 
415 out4:
416 	xprt_rdma_free_addresses(xprt);
417 	rc = -EINVAL;
418 out3:
419 	rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
420 out2:
421 	rpcrdma_ia_close(&new_xprt->rx_ia);
422 out1:
423 	xprt_free(xprt);
424 	return ERR_PTR(rc);
425 }
426 
427 /*
428  * Close a connection, during shutdown or timeout/reconnect
429  */
430 static void
431 xprt_rdma_close(struct rpc_xprt *xprt)
432 {
433 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
434 
435 	dprintk("RPC:       %s: closing\n", __func__);
436 	if (r_xprt->rx_ep.rep_connected > 0)
437 		xprt->reestablish_timeout = 0;
438 	xprt_disconnect_done(xprt);
439 	rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
440 }
441 
442 static void
443 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
444 {
445 	struct sockaddr_in *sap;
446 
447 	sap = (struct sockaddr_in *)&xprt->addr;
448 	sap->sin_port = htons(port);
449 	sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
450 	sap->sin_port = htons(port);
451 	dprintk("RPC:       %s: %u\n", __func__, port);
452 }
453 
454 static void
455 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
456 {
457 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
458 
459 	if (r_xprt->rx_ep.rep_connected != 0) {
460 		/* Reconnect */
461 		schedule_delayed_work(&r_xprt->rx_connect_worker,
462 				      xprt->reestablish_timeout);
463 		xprt->reestablish_timeout <<= 1;
464 		if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
465 			xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
466 		else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
467 			xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
468 	} else {
469 		schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
470 		if (!RPC_IS_ASYNC(task))
471 			flush_delayed_work(&r_xprt->rx_connect_worker);
472 	}
473 }
474 
475 /*
476  * The RDMA allocate/free functions need the task structure as a place
477  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
478  * sequence.
479  *
480  * The RPC layer allocates both send and receive buffers in the same call
481  * (rq_send_buf and rq_rcv_buf are both part of a single contiguous buffer).
482  * We may register rq_rcv_buf when using reply chunks.
483  */
484 static void *
485 xprt_rdma_allocate(struct rpc_task *task, size_t size)
486 {
487 	struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
488 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
489 	struct rpcrdma_regbuf *rb;
490 	struct rpcrdma_req *req;
491 	size_t min_size;
492 	gfp_t flags;
493 
494 	req = rpcrdma_buffer_get(&r_xprt->rx_buf);
495 	if (req == NULL)
496 		return NULL;
497 
498 	flags = GFP_NOIO | __GFP_NOWARN;
499 	if (RPC_IS_SWAPPER(task))
500 		flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
501 
502 	if (req->rl_rdmabuf == NULL)
503 		goto out_rdmabuf;
504 	if (req->rl_sendbuf == NULL)
505 		goto out_sendbuf;
506 	if (size > req->rl_sendbuf->rg_size)
507 		goto out_sendbuf;
508 
509 out:
510 	dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
511 	req->rl_connect_cookie = 0;	/* our reserved value */
512 	return req->rl_sendbuf->rg_base;
513 
514 out_rdmabuf:
515 	min_size = RPCRDMA_INLINE_WRITE_THRESHOLD(task->tk_rqstp);
516 	rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, min_size, flags);
517 	if (IS_ERR(rb))
518 		goto out_fail;
519 	req->rl_rdmabuf = rb;
520 
521 out_sendbuf:
522 	/* XDR encoding and RPC/RDMA marshaling of this request has not
523 	 * yet occurred. Thus a lower bound is needed to prevent buffer
524 	 * overrun during marshaling.
525 	 *
526 	 * RPC/RDMA marshaling may choose to send payload bearing ops
527 	 * inline, if the result is smaller than the inline threshold.
528 	 * The value of the "size" argument accounts for header
529 	 * requirements but not for the payload in these cases.
530 	 *
531 	 * Likewise, allocate enough space to receive a reply up to the
532 	 * size of the inline threshold.
533 	 *
534 	 * It's unlikely that both the send header and the received
535 	 * reply will be large, but slush is provided here to allow
536 	 * flexibility when marshaling.
537 	 */
538 	min_size = RPCRDMA_INLINE_READ_THRESHOLD(task->tk_rqstp);
539 	min_size += RPCRDMA_INLINE_WRITE_THRESHOLD(task->tk_rqstp);
540 	if (size < min_size)
541 		size = min_size;
542 
543 	rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, size, flags);
544 	if (IS_ERR(rb))
545 		goto out_fail;
546 	rb->rg_owner = req;
547 
548 	r_xprt->rx_stats.hardway_register_count += size;
549 	rpcrdma_free_regbuf(&r_xprt->rx_ia, req->rl_sendbuf);
550 	req->rl_sendbuf = rb;
551 	goto out;
552 
553 out_fail:
554 	rpcrdma_buffer_put(req);
555 	r_xprt->rx_stats.failed_marshal_count++;
556 	return NULL;
557 }
558 
559 /*
560  * This function returns all RDMA resources to the pool.
561  */
562 static void
563 xprt_rdma_free(void *buffer)
564 {
565 	struct rpcrdma_req *req;
566 	struct rpcrdma_xprt *r_xprt;
567 	struct rpcrdma_regbuf *rb;
568 	int i;
569 
570 	if (buffer == NULL)
571 		return;
572 
573 	rb = container_of(buffer, struct rpcrdma_regbuf, rg_base[0]);
574 	req = rb->rg_owner;
575 	r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
576 
577 	dprintk("RPC:       %s: called on 0x%p\n", __func__, req->rl_reply);
578 
579 	for (i = 0; req->rl_nchunks;) {
580 		--req->rl_nchunks;
581 		i += r_xprt->rx_ia.ri_ops->ro_unmap(r_xprt,
582 						    &req->rl_segments[i]);
583 	}
584 
585 	rpcrdma_buffer_put(req);
586 }
587 
588 /*
589  * send_request invokes the meat of RPC RDMA. It must do the following:
590  *  1.  Marshal the RPC request into an RPC RDMA request, which means
591  *	putting a header in front of data, and creating IOVs for RDMA
592  *	from those in the request.
593  *  2.  In marshaling, detect opportunities for RDMA, and use them.
594  *  3.  Post a recv message to set up asynch completion, then send
595  *	the request (rpcrdma_ep_post).
596  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
597  */
598 
599 static int
600 xprt_rdma_send_request(struct rpc_task *task)
601 {
602 	struct rpc_rqst *rqst = task->tk_rqstp;
603 	struct rpc_xprt *xprt = rqst->rq_xprt;
604 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
605 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
606 	int rc = 0;
607 
608 	rc = rpcrdma_marshal_req(rqst);
609 	if (rc < 0)
610 		goto failed_marshal;
611 
612 	if (req->rl_reply == NULL) 		/* e.g. reconnection */
613 		rpcrdma_recv_buffer_get(req);
614 
615 	if (req->rl_reply) {
616 		req->rl_reply->rr_func = rpcrdma_reply_handler;
617 		/* this need only be done once, but... */
618 		req->rl_reply->rr_xprt = xprt;
619 	}
620 
621 	/* Must suppress retransmit to maintain credits */
622 	if (req->rl_connect_cookie == xprt->connect_cookie)
623 		goto drop_connection;
624 	req->rl_connect_cookie = xprt->connect_cookie;
625 
626 	if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
627 		goto drop_connection;
628 
629 	rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
630 	rqst->rq_bytes_sent = 0;
631 	return 0;
632 
633 failed_marshal:
634 	r_xprt->rx_stats.failed_marshal_count++;
635 	dprintk("RPC:       %s: rpcrdma_marshal_req failed, status %i\n",
636 		__func__, rc);
637 	if (rc == -EIO)
638 		return -EIO;
639 drop_connection:
640 	xprt_disconnect_done(xprt);
641 	return -ENOTCONN;	/* implies disconnect */
642 }
643 
644 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
645 {
646 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
647 	long idle_time = 0;
648 
649 	if (xprt_connected(xprt))
650 		idle_time = (long)(jiffies - xprt->last_used) / HZ;
651 
652 	seq_printf(seq,
653 	  "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
654 	  "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
655 
656 	   0,	/* need a local port? */
657 	   xprt->stat.bind_count,
658 	   xprt->stat.connect_count,
659 	   xprt->stat.connect_time,
660 	   idle_time,
661 	   xprt->stat.sends,
662 	   xprt->stat.recvs,
663 	   xprt->stat.bad_xids,
664 	   xprt->stat.req_u,
665 	   xprt->stat.bklog_u,
666 
667 	   r_xprt->rx_stats.read_chunk_count,
668 	   r_xprt->rx_stats.write_chunk_count,
669 	   r_xprt->rx_stats.reply_chunk_count,
670 	   r_xprt->rx_stats.total_rdma_request,
671 	   r_xprt->rx_stats.total_rdma_reply,
672 	   r_xprt->rx_stats.pullup_copy_count,
673 	   r_xprt->rx_stats.fixup_copy_count,
674 	   r_xprt->rx_stats.hardway_register_count,
675 	   r_xprt->rx_stats.failed_marshal_count,
676 	   r_xprt->rx_stats.bad_reply_count);
677 }
678 
679 /*
680  * Plumbing for rpc transport switch and kernel module
681  */
682 
683 static struct rpc_xprt_ops xprt_rdma_procs = {
684 	.reserve_xprt		= xprt_reserve_xprt_cong,
685 	.release_xprt		= xprt_release_xprt_cong, /* sunrpc/xprt.c */
686 	.alloc_slot		= xprt_alloc_slot,
687 	.release_request	= xprt_release_rqst_cong,       /* ditto */
688 	.set_retrans_timeout	= xprt_set_retrans_timeout_def, /* ditto */
689 	.rpcbind		= rpcb_getport_async,	/* sunrpc/rpcb_clnt.c */
690 	.set_port		= xprt_rdma_set_port,
691 	.connect		= xprt_rdma_connect,
692 	.buf_alloc		= xprt_rdma_allocate,
693 	.buf_free		= xprt_rdma_free,
694 	.send_request		= xprt_rdma_send_request,
695 	.close			= xprt_rdma_close,
696 	.destroy		= xprt_rdma_destroy,
697 	.print_stats		= xprt_rdma_print_stats
698 };
699 
700 static struct xprt_class xprt_rdma = {
701 	.list			= LIST_HEAD_INIT(xprt_rdma.list),
702 	.name			= "rdma",
703 	.owner			= THIS_MODULE,
704 	.ident			= XPRT_TRANSPORT_RDMA,
705 	.setup			= xprt_setup_rdma,
706 };
707 
708 void xprt_rdma_cleanup(void)
709 {
710 	int rc;
711 
712 	dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
713 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
714 	if (sunrpc_table_header) {
715 		unregister_sysctl_table(sunrpc_table_header);
716 		sunrpc_table_header = NULL;
717 	}
718 #endif
719 	rc = xprt_unregister_transport(&xprt_rdma);
720 	if (rc)
721 		dprintk("RPC:       %s: xprt_unregister returned %i\n",
722 			__func__, rc);
723 }
724 
725 int xprt_rdma_init(void)
726 {
727 	int rc;
728 
729 	rc = xprt_register_transport(&xprt_rdma);
730 
731 	if (rc)
732 		return rc;
733 
734 	dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
735 
736 	dprintk("Defaults:\n");
737 	dprintk("\tSlots %d\n"
738 		"\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
739 		xprt_rdma_slot_table_entries,
740 		xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
741 	dprintk("\tPadding %d\n\tMemreg %d\n",
742 		xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
743 
744 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
745 	if (!sunrpc_table_header)
746 		sunrpc_table_header = register_sysctl_table(sunrpc_table);
747 #endif
748 	return 0;
749 }
750