xref: /linux/net/sunrpc/xprtrdma/transport.c (revision 93d90ad708b8da6efc0e487b66111aa9db7f70c7)
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/init.h>
52 #include <linux/slab.h>
53 #include <linux/seq_file.h>
54 #include <linux/sunrpc/addr.h>
55 
56 #include "xprt_rdma.h"
57 
58 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
59 # define RPCDBG_FACILITY	RPCDBG_TRANS
60 #endif
61 
62 MODULE_LICENSE("Dual BSD/GPL");
63 
64 MODULE_DESCRIPTION("RPC/RDMA Transport for Linux kernel NFS");
65 MODULE_AUTHOR("Network Appliance, Inc.");
66 
67 /*
68  * tunables
69  */
70 
71 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
72 static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
73 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
74 static unsigned int xprt_rdma_inline_write_padding;
75 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
76 		int xprt_rdma_pad_optimize = 1;
77 
78 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
79 
80 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
81 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
82 static unsigned int zero;
83 static unsigned int max_padding = PAGE_SIZE;
84 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
85 static unsigned int max_memreg = RPCRDMA_LAST - 1;
86 
87 static struct ctl_table_header *sunrpc_table_header;
88 
89 static struct ctl_table xr_tunables_table[] = {
90 	{
91 		.procname	= "rdma_slot_table_entries",
92 		.data		= &xprt_rdma_slot_table_entries,
93 		.maxlen		= sizeof(unsigned int),
94 		.mode		= 0644,
95 		.proc_handler	= proc_dointvec_minmax,
96 		.extra1		= &min_slot_table_size,
97 		.extra2		= &max_slot_table_size
98 	},
99 	{
100 		.procname	= "rdma_max_inline_read",
101 		.data		= &xprt_rdma_max_inline_read,
102 		.maxlen		= sizeof(unsigned int),
103 		.mode		= 0644,
104 		.proc_handler	= proc_dointvec,
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,
112 	},
113 	{
114 		.procname	= "rdma_inline_write_padding",
115 		.data		= &xprt_rdma_inline_write_padding,
116 		.maxlen		= sizeof(unsigned int),
117 		.mode		= 0644,
118 		.proc_handler	= proc_dointvec_minmax,
119 		.extra1		= &zero,
120 		.extra2		= &max_padding,
121 	},
122 	{
123 		.procname	= "rdma_memreg_strategy",
124 		.data		= &xprt_rdma_memreg_strategy,
125 		.maxlen		= sizeof(unsigned int),
126 		.mode		= 0644,
127 		.proc_handler	= proc_dointvec_minmax,
128 		.extra1		= &min_memreg,
129 		.extra2		= &max_memreg,
130 	},
131 	{
132 		.procname	= "rdma_pad_optimize",
133 		.data		= &xprt_rdma_pad_optimize,
134 		.maxlen		= sizeof(unsigned int),
135 		.mode		= 0644,
136 		.proc_handler	= proc_dointvec,
137 	},
138 	{ },
139 };
140 
141 static struct ctl_table sunrpc_table[] = {
142 	{
143 		.procname	= "sunrpc",
144 		.mode		= 0555,
145 		.child		= xr_tunables_table
146 	},
147 	{ },
148 };
149 
150 #endif
151 
152 #define RPCRDMA_BIND_TO		(60U * HZ)
153 #define RPCRDMA_INIT_REEST_TO	(5U * HZ)
154 #define RPCRDMA_MAX_REEST_TO	(30U * HZ)
155 #define RPCRDMA_IDLE_DISC_TO	(5U * 60 * HZ)
156 
157 static struct rpc_xprt_ops xprt_rdma_procs;	/* forward reference */
158 
159 static void
160 xprt_rdma_format_addresses(struct rpc_xprt *xprt)
161 {
162 	struct sockaddr *sap = (struct sockaddr *)
163 					&rpcx_to_rdmad(xprt).addr;
164 	struct sockaddr_in *sin = (struct sockaddr_in *)sap;
165 	char buf[64];
166 
167 	(void)rpc_ntop(sap, buf, sizeof(buf));
168 	xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
169 
170 	snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
171 	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
172 
173 	xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
174 
175 	snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
176 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
177 
178 	snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
179 	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
180 
181 	/* netid */
182 	xprt->address_strings[RPC_DISPLAY_NETID] = "rdma";
183 }
184 
185 static void
186 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
187 {
188 	unsigned int i;
189 
190 	for (i = 0; i < RPC_DISPLAY_MAX; i++)
191 		switch (i) {
192 		case RPC_DISPLAY_PROTO:
193 		case RPC_DISPLAY_NETID:
194 			continue;
195 		default:
196 			kfree(xprt->address_strings[i]);
197 		}
198 }
199 
200 static void
201 xprt_rdma_connect_worker(struct work_struct *work)
202 {
203 	struct rpcrdma_xprt *r_xprt =
204 		container_of(work, struct rpcrdma_xprt, rdma_connect.work);
205 	struct rpc_xprt *xprt = &r_xprt->xprt;
206 	int rc = 0;
207 
208 	xprt_clear_connected(xprt);
209 
210 	dprintk("RPC:       %s: %sconnect\n", __func__,
211 			r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
212 	rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
213 	if (rc)
214 		xprt_wake_pending_tasks(xprt, rc);
215 
216 	dprintk("RPC:       %s: exit\n", __func__);
217 	xprt_clear_connecting(xprt);
218 }
219 
220 /*
221  * xprt_rdma_destroy
222  *
223  * Destroy the xprt.
224  * Free all memory associated with the object, including its own.
225  * NOTE: none of the *destroy methods free memory for their top-level
226  * objects, even though they may have allocated it (they do free
227  * private memory). It's up to the caller to handle it. In this
228  * case (RDMA transport), all structure memory is inlined with the
229  * struct rpcrdma_xprt.
230  */
231 static void
232 xprt_rdma_destroy(struct rpc_xprt *xprt)
233 {
234 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
235 
236 	dprintk("RPC:       %s: called\n", __func__);
237 
238 	cancel_delayed_work_sync(&r_xprt->rdma_connect);
239 
240 	xprt_clear_connected(xprt);
241 
242 	rpcrdma_buffer_destroy(&r_xprt->rx_buf);
243 	rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
244 	rpcrdma_ia_close(&r_xprt->rx_ia);
245 
246 	xprt_rdma_free_addresses(xprt);
247 
248 	xprt_free(xprt);
249 
250 	dprintk("RPC:       %s: returning\n", __func__);
251 
252 	module_put(THIS_MODULE);
253 }
254 
255 static const struct rpc_timeout xprt_rdma_default_timeout = {
256 	.to_initval = 60 * HZ,
257 	.to_maxval = 60 * HZ,
258 };
259 
260 /**
261  * xprt_setup_rdma - Set up transport to use RDMA
262  *
263  * @args: rpc transport arguments
264  */
265 static struct rpc_xprt *
266 xprt_setup_rdma(struct xprt_create *args)
267 {
268 	struct rpcrdma_create_data_internal cdata;
269 	struct rpc_xprt *xprt;
270 	struct rpcrdma_xprt *new_xprt;
271 	struct rpcrdma_ep *new_ep;
272 	struct sockaddr_in *sin;
273 	int rc;
274 
275 	if (args->addrlen > sizeof(xprt->addr)) {
276 		dprintk("RPC:       %s: address too large\n", __func__);
277 		return ERR_PTR(-EBADF);
278 	}
279 
280 	xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
281 			xprt_rdma_slot_table_entries,
282 			xprt_rdma_slot_table_entries);
283 	if (xprt == NULL) {
284 		dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
285 			__func__);
286 		return ERR_PTR(-ENOMEM);
287 	}
288 
289 	/* 60 second timeout, no retries */
290 	xprt->timeout = &xprt_rdma_default_timeout;
291 	xprt->bind_timeout = RPCRDMA_BIND_TO;
292 	xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
293 	xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
294 
295 	xprt->resvport = 0;		/* privileged port not needed */
296 	xprt->tsh_size = 0;		/* RPC-RDMA handles framing */
297 	xprt->ops = &xprt_rdma_procs;
298 
299 	/*
300 	 * Set up RDMA-specific connect data.
301 	 */
302 
303 	/* Put server RDMA address in local cdata */
304 	memcpy(&cdata.addr, args->dstaddr, args->addrlen);
305 
306 	/* Ensure xprt->addr holds valid server TCP (not RDMA)
307 	 * address, for any side protocols which peek at it */
308 	xprt->prot = IPPROTO_TCP;
309 	xprt->addrlen = args->addrlen;
310 	memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
311 
312 	sin = (struct sockaddr_in *)&cdata.addr;
313 	if (ntohs(sin->sin_port) != 0)
314 		xprt_set_bound(xprt);
315 
316 	dprintk("RPC:       %s: %pI4:%u\n",
317 		__func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port));
318 
319 	/* Set max requests */
320 	cdata.max_requests = xprt->max_reqs;
321 
322 	/* Set some length limits */
323 	cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
324 	cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
325 
326 	cdata.inline_wsize = xprt_rdma_max_inline_write;
327 	if (cdata.inline_wsize > cdata.wsize)
328 		cdata.inline_wsize = cdata.wsize;
329 
330 	cdata.inline_rsize = xprt_rdma_max_inline_read;
331 	if (cdata.inline_rsize > cdata.rsize)
332 		cdata.inline_rsize = cdata.rsize;
333 
334 	cdata.padding = xprt_rdma_inline_write_padding;
335 
336 	/*
337 	 * Create new transport instance, which includes initialized
338 	 *  o ia
339 	 *  o endpoint
340 	 *  o buffers
341 	 */
342 
343 	new_xprt = rpcx_to_rdmax(xprt);
344 
345 	rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
346 				xprt_rdma_memreg_strategy);
347 	if (rc)
348 		goto out1;
349 
350 	/*
351 	 * initialize and create ep
352 	 */
353 	new_xprt->rx_data = cdata;
354 	new_ep = &new_xprt->rx_ep;
355 	new_ep->rep_remote_addr = cdata.addr;
356 
357 	rc = rpcrdma_ep_create(&new_xprt->rx_ep,
358 				&new_xprt->rx_ia, &new_xprt->rx_data);
359 	if (rc)
360 		goto out2;
361 
362 	/*
363 	 * Allocate pre-registered send and receive buffers for headers and
364 	 * any inline data. Also specify any padding which will be provided
365 	 * from a preregistered zero buffer.
366 	 */
367 	rc = rpcrdma_buffer_create(&new_xprt->rx_buf, new_ep, &new_xprt->rx_ia,
368 				&new_xprt->rx_data);
369 	if (rc)
370 		goto out3;
371 
372 	/*
373 	 * Register a callback for connection events. This is necessary because
374 	 * connection loss notification is async. We also catch connection loss
375 	 * when reaping receives.
376 	 */
377 	INIT_DELAYED_WORK(&new_xprt->rdma_connect, xprt_rdma_connect_worker);
378 	new_ep->rep_func = rpcrdma_conn_func;
379 	new_ep->rep_xprt = xprt;
380 
381 	xprt_rdma_format_addresses(xprt);
382 	xprt->max_payload = rpcrdma_max_payload(new_xprt);
383 	dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
384 		__func__, xprt->max_payload);
385 
386 	if (!try_module_get(THIS_MODULE))
387 		goto out4;
388 
389 	return xprt;
390 
391 out4:
392 	xprt_rdma_free_addresses(xprt);
393 	rc = -EINVAL;
394 out3:
395 	rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
396 out2:
397 	rpcrdma_ia_close(&new_xprt->rx_ia);
398 out1:
399 	xprt_free(xprt);
400 	return ERR_PTR(rc);
401 }
402 
403 /*
404  * Close a connection, during shutdown or timeout/reconnect
405  */
406 static void
407 xprt_rdma_close(struct rpc_xprt *xprt)
408 {
409 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
410 
411 	dprintk("RPC:       %s: closing\n", __func__);
412 	if (r_xprt->rx_ep.rep_connected > 0)
413 		xprt->reestablish_timeout = 0;
414 	xprt_disconnect_done(xprt);
415 	rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
416 }
417 
418 static void
419 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
420 {
421 	struct sockaddr_in *sap;
422 
423 	sap = (struct sockaddr_in *)&xprt->addr;
424 	sap->sin_port = htons(port);
425 	sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
426 	sap->sin_port = htons(port);
427 	dprintk("RPC:       %s: %u\n", __func__, port);
428 }
429 
430 static void
431 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
432 {
433 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
434 
435 	if (r_xprt->rx_ep.rep_connected != 0) {
436 		/* Reconnect */
437 		schedule_delayed_work(&r_xprt->rdma_connect,
438 			xprt->reestablish_timeout);
439 		xprt->reestablish_timeout <<= 1;
440 		if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
441 			xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
442 		else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
443 			xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
444 	} else {
445 		schedule_delayed_work(&r_xprt->rdma_connect, 0);
446 		if (!RPC_IS_ASYNC(task))
447 			flush_delayed_work(&r_xprt->rdma_connect);
448 	}
449 }
450 
451 /*
452  * The RDMA allocate/free functions need the task structure as a place
453  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
454  * sequence. For this reason, the recv buffers are attached to send
455  * buffers for portions of the RPC. Note that the RPC layer allocates
456  * both send and receive buffers in the same call. We may register
457  * the receive buffer portion when using reply chunks.
458  */
459 static void *
460 xprt_rdma_allocate(struct rpc_task *task, size_t size)
461 {
462 	struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
463 	struct rpcrdma_req *req, *nreq;
464 
465 	req = rpcrdma_buffer_get(&rpcx_to_rdmax(xprt)->rx_buf);
466 	if (req == NULL)
467 		return NULL;
468 
469 	if (size > req->rl_size) {
470 		dprintk("RPC:       %s: size %zd too large for buffer[%zd]: "
471 			"prog %d vers %d proc %d\n",
472 			__func__, size, req->rl_size,
473 			task->tk_client->cl_prog, task->tk_client->cl_vers,
474 			task->tk_msg.rpc_proc->p_proc);
475 		/*
476 		 * Outgoing length shortage. Our inline write max must have
477 		 * been configured to perform direct i/o.
478 		 *
479 		 * This is therefore a large metadata operation, and the
480 		 * allocate call was made on the maximum possible message,
481 		 * e.g. containing long filename(s) or symlink data. In
482 		 * fact, while these metadata operations *might* carry
483 		 * large outgoing payloads, they rarely *do*. However, we
484 		 * have to commit to the request here, so reallocate and
485 		 * register it now. The data path will never require this
486 		 * reallocation.
487 		 *
488 		 * If the allocation or registration fails, the RPC framework
489 		 * will (doggedly) retry.
490 		 */
491 		if (task->tk_flags & RPC_TASK_SWAPPER)
492 			nreq = kmalloc(sizeof *req + size, GFP_ATOMIC);
493 		else
494 			nreq = kmalloc(sizeof *req + size, GFP_NOFS);
495 		if (nreq == NULL)
496 			goto outfail;
497 
498 		if (rpcrdma_register_internal(&rpcx_to_rdmax(xprt)->rx_ia,
499 				nreq->rl_base, size + sizeof(struct rpcrdma_req)
500 				- offsetof(struct rpcrdma_req, rl_base),
501 				&nreq->rl_handle, &nreq->rl_iov)) {
502 			kfree(nreq);
503 			goto outfail;
504 		}
505 		rpcx_to_rdmax(xprt)->rx_stats.hardway_register_count += size;
506 		nreq->rl_size = size;
507 		nreq->rl_niovs = 0;
508 		nreq->rl_nchunks = 0;
509 		nreq->rl_buffer = (struct rpcrdma_buffer *)req;
510 		nreq->rl_reply = req->rl_reply;
511 		memcpy(nreq->rl_segments,
512 			req->rl_segments, sizeof nreq->rl_segments);
513 		/* flag the swap with an unused field */
514 		nreq->rl_iov.length = 0;
515 		req->rl_reply = NULL;
516 		req = nreq;
517 	}
518 	dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
519 	req->rl_connect_cookie = 0;	/* our reserved value */
520 	return req->rl_xdr_buf;
521 
522 outfail:
523 	rpcrdma_buffer_put(req);
524 	rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
525 	return NULL;
526 }
527 
528 /*
529  * This function returns all RDMA resources to the pool.
530  */
531 static void
532 xprt_rdma_free(void *buffer)
533 {
534 	struct rpcrdma_req *req;
535 	struct rpcrdma_xprt *r_xprt;
536 	struct rpcrdma_rep *rep;
537 	int i;
538 
539 	if (buffer == NULL)
540 		return;
541 
542 	req = container_of(buffer, struct rpcrdma_req, rl_xdr_buf[0]);
543 	if (req->rl_iov.length == 0) {	/* see allocate above */
544 		r_xprt = container_of(((struct rpcrdma_req *) req->rl_buffer)->rl_buffer,
545 				      struct rpcrdma_xprt, rx_buf);
546 	} else
547 		r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
548 	rep = req->rl_reply;
549 
550 	dprintk("RPC:       %s: called on 0x%p%s\n",
551 		__func__, rep, (rep && rep->rr_func) ? " (with waiter)" : "");
552 
553 	/*
554 	 * Finish the deregistration.  The process is considered
555 	 * complete when the rr_func vector becomes NULL - this
556 	 * was put in place during rpcrdma_reply_handler() - the wait
557 	 * call below will not block if the dereg is "done". If
558 	 * interrupted, our framework will clean up.
559 	 */
560 	for (i = 0; req->rl_nchunks;) {
561 		--req->rl_nchunks;
562 		i += rpcrdma_deregister_external(
563 			&req->rl_segments[i], r_xprt);
564 	}
565 
566 	if (req->rl_iov.length == 0) {	/* see allocate above */
567 		struct rpcrdma_req *oreq = (struct rpcrdma_req *)req->rl_buffer;
568 		oreq->rl_reply = req->rl_reply;
569 		(void) rpcrdma_deregister_internal(&r_xprt->rx_ia,
570 						   req->rl_handle,
571 						   &req->rl_iov);
572 		kfree(req);
573 		req = oreq;
574 	}
575 
576 	/* Put back request+reply buffers */
577 	rpcrdma_buffer_put(req);
578 }
579 
580 /*
581  * send_request invokes the meat of RPC RDMA. It must do the following:
582  *  1.  Marshal the RPC request into an RPC RDMA request, which means
583  *	putting a header in front of data, and creating IOVs for RDMA
584  *	from those in the request.
585  *  2.  In marshaling, detect opportunities for RDMA, and use them.
586  *  3.  Post a recv message to set up asynch completion, then send
587  *	the request (rpcrdma_ep_post).
588  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
589  */
590 
591 static int
592 xprt_rdma_send_request(struct rpc_task *task)
593 {
594 	struct rpc_rqst *rqst = task->tk_rqstp;
595 	struct rpc_xprt *xprt = rqst->rq_xprt;
596 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
597 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
598 	int rc = 0;
599 
600 	if (req->rl_niovs == 0)
601 		rc = rpcrdma_marshal_req(rqst);
602 	else if (r_xprt->rx_ia.ri_memreg_strategy != RPCRDMA_ALLPHYSICAL)
603 		rc = rpcrdma_marshal_chunks(rqst, 0);
604 	if (rc < 0)
605 		goto failed_marshal;
606 
607 	if (req->rl_reply == NULL) 		/* e.g. reconnection */
608 		rpcrdma_recv_buffer_get(req);
609 
610 	if (req->rl_reply) {
611 		req->rl_reply->rr_func = rpcrdma_reply_handler;
612 		/* this need only be done once, but... */
613 		req->rl_reply->rr_xprt = xprt;
614 	}
615 
616 	/* Must suppress retransmit to maintain credits */
617 	if (req->rl_connect_cookie == xprt->connect_cookie)
618 		goto drop_connection;
619 	req->rl_connect_cookie = xprt->connect_cookie;
620 
621 	if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
622 		goto drop_connection;
623 
624 	rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
625 	rqst->rq_bytes_sent = 0;
626 	return 0;
627 
628 failed_marshal:
629 	r_xprt->rx_stats.failed_marshal_count++;
630 	dprintk("RPC:       %s: rpcrdma_marshal_req failed, status %i\n",
631 		__func__, rc);
632 	if (rc == -EIO)
633 		return -EIO;
634 drop_connection:
635 	xprt_disconnect_done(xprt);
636 	return -ENOTCONN;	/* implies disconnect */
637 }
638 
639 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
640 {
641 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
642 	long idle_time = 0;
643 
644 	if (xprt_connected(xprt))
645 		idle_time = (long)(jiffies - xprt->last_used) / HZ;
646 
647 	seq_printf(seq,
648 	  "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
649 	  "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
650 
651 	   0,	/* need a local port? */
652 	   xprt->stat.bind_count,
653 	   xprt->stat.connect_count,
654 	   xprt->stat.connect_time,
655 	   idle_time,
656 	   xprt->stat.sends,
657 	   xprt->stat.recvs,
658 	   xprt->stat.bad_xids,
659 	   xprt->stat.req_u,
660 	   xprt->stat.bklog_u,
661 
662 	   r_xprt->rx_stats.read_chunk_count,
663 	   r_xprt->rx_stats.write_chunk_count,
664 	   r_xprt->rx_stats.reply_chunk_count,
665 	   r_xprt->rx_stats.total_rdma_request,
666 	   r_xprt->rx_stats.total_rdma_reply,
667 	   r_xprt->rx_stats.pullup_copy_count,
668 	   r_xprt->rx_stats.fixup_copy_count,
669 	   r_xprt->rx_stats.hardway_register_count,
670 	   r_xprt->rx_stats.failed_marshal_count,
671 	   r_xprt->rx_stats.bad_reply_count);
672 }
673 
674 /*
675  * Plumbing for rpc transport switch and kernel module
676  */
677 
678 static struct rpc_xprt_ops xprt_rdma_procs = {
679 	.reserve_xprt		= xprt_reserve_xprt_cong,
680 	.release_xprt		= xprt_release_xprt_cong, /* sunrpc/xprt.c */
681 	.alloc_slot		= xprt_alloc_slot,
682 	.release_request	= xprt_release_rqst_cong,       /* ditto */
683 	.set_retrans_timeout	= xprt_set_retrans_timeout_def, /* ditto */
684 	.rpcbind		= rpcb_getport_async,	/* sunrpc/rpcb_clnt.c */
685 	.set_port		= xprt_rdma_set_port,
686 	.connect		= xprt_rdma_connect,
687 	.buf_alloc		= xprt_rdma_allocate,
688 	.buf_free		= xprt_rdma_free,
689 	.send_request		= xprt_rdma_send_request,
690 	.close			= xprt_rdma_close,
691 	.destroy		= xprt_rdma_destroy,
692 	.print_stats		= xprt_rdma_print_stats
693 };
694 
695 static struct xprt_class xprt_rdma = {
696 	.list			= LIST_HEAD_INIT(xprt_rdma.list),
697 	.name			= "rdma",
698 	.owner			= THIS_MODULE,
699 	.ident			= XPRT_TRANSPORT_RDMA,
700 	.setup			= xprt_setup_rdma,
701 };
702 
703 static void __exit xprt_rdma_cleanup(void)
704 {
705 	int rc;
706 
707 	dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
708 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
709 	if (sunrpc_table_header) {
710 		unregister_sysctl_table(sunrpc_table_header);
711 		sunrpc_table_header = NULL;
712 	}
713 #endif
714 	rc = xprt_unregister_transport(&xprt_rdma);
715 	if (rc)
716 		dprintk("RPC:       %s: xprt_unregister returned %i\n",
717 			__func__, rc);
718 }
719 
720 static int __init xprt_rdma_init(void)
721 {
722 	int rc;
723 
724 	rc = xprt_register_transport(&xprt_rdma);
725 
726 	if (rc)
727 		return rc;
728 
729 	dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
730 
731 	dprintk("Defaults:\n");
732 	dprintk("\tSlots %d\n"
733 		"\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
734 		xprt_rdma_slot_table_entries,
735 		xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
736 	dprintk("\tPadding %d\n\tMemreg %d\n",
737 		xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
738 
739 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
740 	if (!sunrpc_table_header)
741 		sunrpc_table_header = register_sysctl_table(sunrpc_table);
742 #endif
743 	return 0;
744 }
745 
746 module_init(xprt_rdma_init);
747 module_exit(xprt_rdma_cleanup);
748