xref: /openbmc/linux/net/sunrpc/xprtrdma/transport.c (revision 171f1bc7)
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 
55 #include "xprt_rdma.h"
56 
57 #ifdef RPC_DEBUG
58 # define RPCDBG_FACILITY	RPCDBG_TRANS
59 #endif
60 
61 MODULE_LICENSE("Dual BSD/GPL");
62 
63 MODULE_DESCRIPTION("RPC/RDMA Transport for Linux kernel NFS");
64 MODULE_AUTHOR("Network Appliance, Inc.");
65 
66 /*
67  * tunables
68  */
69 
70 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
71 static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
72 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
73 static unsigned int xprt_rdma_inline_write_padding;
74 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
75                 int xprt_rdma_pad_optimize = 0;
76 
77 #ifdef RPC_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 zero;
82 static unsigned int max_padding = PAGE_SIZE;
83 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
84 static unsigned int max_memreg = RPCRDMA_LAST - 1;
85 
86 static struct ctl_table_header *sunrpc_table_header;
87 
88 static ctl_table xr_tunables_table[] = {
89 	{
90 		.procname	= "rdma_slot_table_entries",
91 		.data		= &xprt_rdma_slot_table_entries,
92 		.maxlen		= sizeof(unsigned int),
93 		.mode		= 0644,
94 		.proc_handler	= proc_dointvec_minmax,
95 		.extra1		= &min_slot_table_size,
96 		.extra2		= &max_slot_table_size
97 	},
98 	{
99 		.procname	= "rdma_max_inline_read",
100 		.data		= &xprt_rdma_max_inline_read,
101 		.maxlen		= sizeof(unsigned int),
102 		.mode		= 0644,
103 		.proc_handler	= proc_dointvec,
104 	},
105 	{
106 		.procname	= "rdma_max_inline_write",
107 		.data		= &xprt_rdma_max_inline_write,
108 		.maxlen		= sizeof(unsigned int),
109 		.mode		= 0644,
110 		.proc_handler	= proc_dointvec,
111 	},
112 	{
113 		.procname	= "rdma_inline_write_padding",
114 		.data		= &xprt_rdma_inline_write_padding,
115 		.maxlen		= sizeof(unsigned int),
116 		.mode		= 0644,
117 		.proc_handler	= proc_dointvec_minmax,
118 		.extra1		= &zero,
119 		.extra2		= &max_padding,
120 	},
121 	{
122 		.procname	= "rdma_memreg_strategy",
123 		.data		= &xprt_rdma_memreg_strategy,
124 		.maxlen		= sizeof(unsigned int),
125 		.mode		= 0644,
126 		.proc_handler	= proc_dointvec_minmax,
127 		.extra1		= &min_memreg,
128 		.extra2		= &max_memreg,
129 	},
130 	{
131 		.procname	= "rdma_pad_optimize",
132 		.data		= &xprt_rdma_pad_optimize,
133 		.maxlen		= sizeof(unsigned int),
134 		.mode		= 0644,
135 		.proc_handler	= proc_dointvec,
136 	},
137 	{ },
138 };
139 
140 static ctl_table sunrpc_table[] = {
141 	{
142 		.procname	= "sunrpc",
143 		.mode		= 0555,
144 		.child		= xr_tunables_table
145 	},
146 	{ },
147 };
148 
149 #endif
150 
151 static struct rpc_xprt_ops xprt_rdma_procs;	/* forward reference */
152 
153 static void
154 xprt_rdma_format_addresses(struct rpc_xprt *xprt)
155 {
156 	struct sockaddr *sap = (struct sockaddr *)
157 					&rpcx_to_rdmad(xprt).addr;
158 	struct sockaddr_in *sin = (struct sockaddr_in *)sap;
159 	char buf[64];
160 
161 	(void)rpc_ntop(sap, buf, sizeof(buf));
162 	xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
163 
164 	snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
165 	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
166 
167 	xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
168 
169 	snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
170 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
171 
172 	snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
173 	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
174 
175 	/* netid */
176 	xprt->address_strings[RPC_DISPLAY_NETID] = "rdma";
177 }
178 
179 static void
180 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
181 {
182 	unsigned int i;
183 
184 	for (i = 0; i < RPC_DISPLAY_MAX; i++)
185 		switch (i) {
186 		case RPC_DISPLAY_PROTO:
187 		case RPC_DISPLAY_NETID:
188 			continue;
189 		default:
190 			kfree(xprt->address_strings[i]);
191 		}
192 }
193 
194 static void
195 xprt_rdma_connect_worker(struct work_struct *work)
196 {
197 	struct rpcrdma_xprt *r_xprt =
198 		container_of(work, struct rpcrdma_xprt, rdma_connect.work);
199 	struct rpc_xprt *xprt = &r_xprt->xprt;
200 	int rc = 0;
201 
202 	if (!xprt->shutdown) {
203 		xprt_clear_connected(xprt);
204 
205 		dprintk("RPC:       %s: %sconnect\n", __func__,
206 				r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
207 		rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
208 		if (rc)
209 			goto out;
210 	}
211 	goto out_clear;
212 
213 out:
214 	xprt_wake_pending_tasks(xprt, rc);
215 
216 out_clear:
217 	dprintk("RPC:       %s: exit\n", __func__);
218 	xprt_clear_connecting(xprt);
219 }
220 
221 /*
222  * xprt_rdma_destroy
223  *
224  * Destroy the xprt.
225  * Free all memory associated with the object, including its own.
226  * NOTE: none of the *destroy methods free memory for their top-level
227  * objects, even though they may have allocated it (they do free
228  * private memory). It's up to the caller to handle it. In this
229  * case (RDMA transport), all structure memory is inlined with the
230  * struct rpcrdma_xprt.
231  */
232 static void
233 xprt_rdma_destroy(struct rpc_xprt *xprt)
234 {
235 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
236 	int rc;
237 
238 	dprintk("RPC:       %s: called\n", __func__);
239 
240 	cancel_delayed_work_sync(&r_xprt->rdma_connect);
241 
242 	xprt_clear_connected(xprt);
243 
244 	rpcrdma_buffer_destroy(&r_xprt->rx_buf);
245 	rc = rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
246 	if (rc)
247 		dprintk("RPC:       %s: rpcrdma_ep_destroy returned %i\n",
248 			__func__, rc);
249 	rpcrdma_ia_close(&r_xprt->rx_ia);
250 
251 	xprt_rdma_free_addresses(xprt);
252 
253 	xprt_free(xprt);
254 
255 	dprintk("RPC:       %s: returning\n", __func__);
256 
257 	module_put(THIS_MODULE);
258 }
259 
260 static const struct rpc_timeout xprt_rdma_default_timeout = {
261 	.to_initval = 60 * HZ,
262 	.to_maxval = 60 * HZ,
263 };
264 
265 /**
266  * xprt_setup_rdma - Set up transport to use RDMA
267  *
268  * @args: rpc transport arguments
269  */
270 static struct rpc_xprt *
271 xprt_setup_rdma(struct xprt_create *args)
272 {
273 	struct rpcrdma_create_data_internal cdata;
274 	struct rpc_xprt *xprt;
275 	struct rpcrdma_xprt *new_xprt;
276 	struct rpcrdma_ep *new_ep;
277 	struct sockaddr_in *sin;
278 	int rc;
279 
280 	if (args->addrlen > sizeof(xprt->addr)) {
281 		dprintk("RPC:       %s: address too large\n", __func__);
282 		return ERR_PTR(-EBADF);
283 	}
284 
285 	xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
286 			xprt_rdma_slot_table_entries,
287 			xprt_rdma_slot_table_entries);
288 	if (xprt == NULL) {
289 		dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
290 			__func__);
291 		return ERR_PTR(-ENOMEM);
292 	}
293 
294 	/* 60 second timeout, no retries */
295 	xprt->timeout = &xprt_rdma_default_timeout;
296 	xprt->bind_timeout = (60U * HZ);
297 	xprt->reestablish_timeout = (5U * HZ);
298 	xprt->idle_timeout = (5U * 60 * HZ);
299 
300 	xprt->resvport = 0;		/* privileged port not needed */
301 	xprt->tsh_size = 0;		/* RPC-RDMA handles framing */
302 	xprt->max_payload = RPCRDMA_MAX_DATA_SEGS * PAGE_SIZE;
303 	xprt->ops = &xprt_rdma_procs;
304 
305 	/*
306 	 * Set up RDMA-specific connect data.
307 	 */
308 
309 	/* Put server RDMA address in local cdata */
310 	memcpy(&cdata.addr, args->dstaddr, args->addrlen);
311 
312 	/* Ensure xprt->addr holds valid server TCP (not RDMA)
313 	 * address, for any side protocols which peek at it */
314 	xprt->prot = IPPROTO_TCP;
315 	xprt->addrlen = args->addrlen;
316 	memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
317 
318 	sin = (struct sockaddr_in *)&cdata.addr;
319 	if (ntohs(sin->sin_port) != 0)
320 		xprt_set_bound(xprt);
321 
322 	dprintk("RPC:       %s: %pI4:%u\n",
323 		__func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port));
324 
325 	/* Set max requests */
326 	cdata.max_requests = xprt->max_reqs;
327 
328 	/* Set some length limits */
329 	cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
330 	cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
331 
332 	cdata.inline_wsize = xprt_rdma_max_inline_write;
333 	if (cdata.inline_wsize > cdata.wsize)
334 		cdata.inline_wsize = cdata.wsize;
335 
336 	cdata.inline_rsize = xprt_rdma_max_inline_read;
337 	if (cdata.inline_rsize > cdata.rsize)
338 		cdata.inline_rsize = cdata.rsize;
339 
340 	cdata.padding = xprt_rdma_inline_write_padding;
341 
342 	/*
343 	 * Create new transport instance, which includes initialized
344 	 *  o ia
345 	 *  o endpoint
346 	 *  o buffers
347 	 */
348 
349 	new_xprt = rpcx_to_rdmax(xprt);
350 
351 	rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
352 				xprt_rdma_memreg_strategy);
353 	if (rc)
354 		goto out1;
355 
356 	/*
357 	 * initialize and create ep
358 	 */
359 	new_xprt->rx_data = cdata;
360 	new_ep = &new_xprt->rx_ep;
361 	new_ep->rep_remote_addr = cdata.addr;
362 
363 	rc = rpcrdma_ep_create(&new_xprt->rx_ep,
364 				&new_xprt->rx_ia, &new_xprt->rx_data);
365 	if (rc)
366 		goto out2;
367 
368 	/*
369 	 * Allocate pre-registered send and receive buffers for headers and
370 	 * any inline data. Also specify any padding which will be provided
371 	 * from a preregistered zero buffer.
372 	 */
373 	rc = rpcrdma_buffer_create(&new_xprt->rx_buf, new_ep, &new_xprt->rx_ia,
374 				&new_xprt->rx_data);
375 	if (rc)
376 		goto out3;
377 
378 	/*
379 	 * Register a callback for connection events. This is necessary because
380 	 * connection loss notification is async. We also catch connection loss
381 	 * when reaping receives.
382 	 */
383 	INIT_DELAYED_WORK(&new_xprt->rdma_connect, xprt_rdma_connect_worker);
384 	new_ep->rep_func = rpcrdma_conn_func;
385 	new_ep->rep_xprt = xprt;
386 
387 	xprt_rdma_format_addresses(xprt);
388 
389 	if (!try_module_get(THIS_MODULE))
390 		goto out4;
391 
392 	return xprt;
393 
394 out4:
395 	xprt_rdma_free_addresses(xprt);
396 	rc = -EINVAL;
397 out3:
398 	(void) rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
399 out2:
400 	rpcrdma_ia_close(&new_xprt->rx_ia);
401 out1:
402 	xprt_free(xprt);
403 	return ERR_PTR(rc);
404 }
405 
406 /*
407  * Close a connection, during shutdown or timeout/reconnect
408  */
409 static void
410 xprt_rdma_close(struct rpc_xprt *xprt)
411 {
412 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
413 
414 	dprintk("RPC:       %s: closing\n", __func__);
415 	if (r_xprt->rx_ep.rep_connected > 0)
416 		xprt->reestablish_timeout = 0;
417 	xprt_disconnect_done(xprt);
418 	(void) rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
419 }
420 
421 static void
422 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
423 {
424 	struct sockaddr_in *sap;
425 
426 	sap = (struct sockaddr_in *)&xprt->addr;
427 	sap->sin_port = htons(port);
428 	sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
429 	sap->sin_port = htons(port);
430 	dprintk("RPC:       %s: %u\n", __func__, port);
431 }
432 
433 static void
434 xprt_rdma_connect(struct rpc_task *task)
435 {
436 	struct rpc_xprt *xprt = (struct rpc_xprt *)task->tk_xprt;
437 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
438 
439 	if (r_xprt->rx_ep.rep_connected != 0) {
440 		/* Reconnect */
441 		schedule_delayed_work(&r_xprt->rdma_connect,
442 			xprt->reestablish_timeout);
443 		xprt->reestablish_timeout <<= 1;
444 		if (xprt->reestablish_timeout > (30 * HZ))
445 			xprt->reestablish_timeout = (30 * HZ);
446 		else if (xprt->reestablish_timeout < (5 * HZ))
447 			xprt->reestablish_timeout = (5 * HZ);
448 	} else {
449 		schedule_delayed_work(&r_xprt->rdma_connect, 0);
450 		if (!RPC_IS_ASYNC(task))
451 			flush_delayed_work(&r_xprt->rdma_connect);
452 	}
453 }
454 
455 static int
456 xprt_rdma_reserve_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
457 {
458 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
459 	int credits = atomic_read(&r_xprt->rx_buf.rb_credits);
460 
461 	/* == RPC_CWNDSCALE @ init, but *after* setup */
462 	if (r_xprt->rx_buf.rb_cwndscale == 0UL) {
463 		r_xprt->rx_buf.rb_cwndscale = xprt->cwnd;
464 		dprintk("RPC:       %s: cwndscale %lu\n", __func__,
465 			r_xprt->rx_buf.rb_cwndscale);
466 		BUG_ON(r_xprt->rx_buf.rb_cwndscale <= 0);
467 	}
468 	xprt->cwnd = credits * r_xprt->rx_buf.rb_cwndscale;
469 	return xprt_reserve_xprt_cong(xprt, task);
470 }
471 
472 /*
473  * The RDMA allocate/free functions need the task structure as a place
474  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
475  * sequence. For this reason, the recv buffers are attached to send
476  * buffers for portions of the RPC. Note that the RPC layer allocates
477  * both send and receive buffers in the same call. We may register
478  * the receive buffer portion when using reply chunks.
479  */
480 static void *
481 xprt_rdma_allocate(struct rpc_task *task, size_t size)
482 {
483 	struct rpc_xprt *xprt = task->tk_xprt;
484 	struct rpcrdma_req *req, *nreq;
485 
486 	req = rpcrdma_buffer_get(&rpcx_to_rdmax(xprt)->rx_buf);
487 	BUG_ON(NULL == req);
488 
489 	if (size > req->rl_size) {
490 		dprintk("RPC:       %s: size %zd too large for buffer[%zd]: "
491 			"prog %d vers %d proc %d\n",
492 			__func__, size, req->rl_size,
493 			task->tk_client->cl_prog, task->tk_client->cl_vers,
494 			task->tk_msg.rpc_proc->p_proc);
495 		/*
496 		 * Outgoing length shortage. Our inline write max must have
497 		 * been configured to perform direct i/o.
498 		 *
499 		 * This is therefore a large metadata operation, and the
500 		 * allocate call was made on the maximum possible message,
501 		 * e.g. containing long filename(s) or symlink data. In
502 		 * fact, while these metadata operations *might* carry
503 		 * large outgoing payloads, they rarely *do*. However, we
504 		 * have to commit to the request here, so reallocate and
505 		 * register it now. The data path will never require this
506 		 * reallocation.
507 		 *
508 		 * If the allocation or registration fails, the RPC framework
509 		 * will (doggedly) retry.
510 		 */
511 		if (rpcx_to_rdmax(xprt)->rx_ia.ri_memreg_strategy ==
512 				RPCRDMA_BOUNCEBUFFERS) {
513 			/* forced to "pure inline" */
514 			dprintk("RPC:       %s: too much data (%zd) for inline "
515 					"(r/w max %d/%d)\n", __func__, size,
516 					rpcx_to_rdmad(xprt).inline_rsize,
517 					rpcx_to_rdmad(xprt).inline_wsize);
518 			size = req->rl_size;
519 			rpc_exit(task, -EIO);		/* fail the operation */
520 			rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
521 			goto out;
522 		}
523 		if (task->tk_flags & RPC_TASK_SWAPPER)
524 			nreq = kmalloc(sizeof *req + size, GFP_ATOMIC);
525 		else
526 			nreq = kmalloc(sizeof *req + size, GFP_NOFS);
527 		if (nreq == NULL)
528 			goto outfail;
529 
530 		if (rpcrdma_register_internal(&rpcx_to_rdmax(xprt)->rx_ia,
531 				nreq->rl_base, size + sizeof(struct rpcrdma_req)
532 				- offsetof(struct rpcrdma_req, rl_base),
533 				&nreq->rl_handle, &nreq->rl_iov)) {
534 			kfree(nreq);
535 			goto outfail;
536 		}
537 		rpcx_to_rdmax(xprt)->rx_stats.hardway_register_count += size;
538 		nreq->rl_size = size;
539 		nreq->rl_niovs = 0;
540 		nreq->rl_nchunks = 0;
541 		nreq->rl_buffer = (struct rpcrdma_buffer *)req;
542 		nreq->rl_reply = req->rl_reply;
543 		memcpy(nreq->rl_segments,
544 			req->rl_segments, sizeof nreq->rl_segments);
545 		/* flag the swap with an unused field */
546 		nreq->rl_iov.length = 0;
547 		req->rl_reply = NULL;
548 		req = nreq;
549 	}
550 	dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
551 out:
552 	req->rl_connect_cookie = 0;	/* our reserved value */
553 	return req->rl_xdr_buf;
554 
555 outfail:
556 	rpcrdma_buffer_put(req);
557 	rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
558 	return NULL;
559 }
560 
561 /*
562  * This function returns all RDMA resources to the pool.
563  */
564 static void
565 xprt_rdma_free(void *buffer)
566 {
567 	struct rpcrdma_req *req;
568 	struct rpcrdma_xprt *r_xprt;
569 	struct rpcrdma_rep *rep;
570 	int i;
571 
572 	if (buffer == NULL)
573 		return;
574 
575 	req = container_of(buffer, struct rpcrdma_req, rl_xdr_buf[0]);
576 	if (req->rl_iov.length == 0) {	/* see allocate above */
577 		r_xprt = container_of(((struct rpcrdma_req *) req->rl_buffer)->rl_buffer,
578 				      struct rpcrdma_xprt, rx_buf);
579 	} else
580 		r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
581 	rep = req->rl_reply;
582 
583 	dprintk("RPC:       %s: called on 0x%p%s\n",
584 		__func__, rep, (rep && rep->rr_func) ? " (with waiter)" : "");
585 
586 	/*
587 	 * Finish the deregistration. When using mw bind, this was
588 	 * begun in rpcrdma_reply_handler(). In all other modes, we
589 	 * do it here, in thread context. The process is considered
590 	 * complete when the rr_func vector becomes NULL - this
591 	 * was put in place during rpcrdma_reply_handler() - the wait
592 	 * call below will not block if the dereg is "done". If
593 	 * interrupted, our framework will clean up.
594 	 */
595 	for (i = 0; req->rl_nchunks;) {
596 		--req->rl_nchunks;
597 		i += rpcrdma_deregister_external(
598 			&req->rl_segments[i], r_xprt, NULL);
599 	}
600 
601 	if (rep && wait_event_interruptible(rep->rr_unbind, !rep->rr_func)) {
602 		rep->rr_func = NULL;	/* abandon the callback */
603 		req->rl_reply = NULL;
604 	}
605 
606 	if (req->rl_iov.length == 0) {	/* see allocate above */
607 		struct rpcrdma_req *oreq = (struct rpcrdma_req *)req->rl_buffer;
608 		oreq->rl_reply = req->rl_reply;
609 		(void) rpcrdma_deregister_internal(&r_xprt->rx_ia,
610 						   req->rl_handle,
611 						   &req->rl_iov);
612 		kfree(req);
613 		req = oreq;
614 	}
615 
616 	/* Put back request+reply buffers */
617 	rpcrdma_buffer_put(req);
618 }
619 
620 /*
621  * send_request invokes the meat of RPC RDMA. It must do the following:
622  *  1.  Marshal the RPC request into an RPC RDMA request, which means
623  *	putting a header in front of data, and creating IOVs for RDMA
624  *	from those in the request.
625  *  2.  In marshaling, detect opportunities for RDMA, and use them.
626  *  3.  Post a recv message to set up asynch completion, then send
627  *	the request (rpcrdma_ep_post).
628  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
629  */
630 
631 static int
632 xprt_rdma_send_request(struct rpc_task *task)
633 {
634 	struct rpc_rqst *rqst = task->tk_rqstp;
635 	struct rpc_xprt *xprt = task->tk_xprt;
636 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
637 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
638 
639 	/* marshal the send itself */
640 	if (req->rl_niovs == 0 && rpcrdma_marshal_req(rqst) != 0) {
641 		r_xprt->rx_stats.failed_marshal_count++;
642 		dprintk("RPC:       %s: rpcrdma_marshal_req failed\n",
643 			__func__);
644 		return -EIO;
645 	}
646 
647 	if (req->rl_reply == NULL) 		/* e.g. reconnection */
648 		rpcrdma_recv_buffer_get(req);
649 
650 	if (req->rl_reply) {
651 		req->rl_reply->rr_func = rpcrdma_reply_handler;
652 		/* this need only be done once, but... */
653 		req->rl_reply->rr_xprt = xprt;
654 	}
655 
656 	/* Must suppress retransmit to maintain credits */
657 	if (req->rl_connect_cookie == xprt->connect_cookie)
658 		goto drop_connection;
659 	req->rl_connect_cookie = xprt->connect_cookie;
660 
661 	if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
662 		goto drop_connection;
663 
664 	rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
665 	rqst->rq_bytes_sent = 0;
666 	return 0;
667 
668 drop_connection:
669 	xprt_disconnect_done(xprt);
670 	return -ENOTCONN;	/* implies disconnect */
671 }
672 
673 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
674 {
675 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
676 	long idle_time = 0;
677 
678 	if (xprt_connected(xprt))
679 		idle_time = (long)(jiffies - xprt->last_used) / HZ;
680 
681 	seq_printf(seq,
682 	  "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
683 	  "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
684 
685 	   0,	/* need a local port? */
686 	   xprt->stat.bind_count,
687 	   xprt->stat.connect_count,
688 	   xprt->stat.connect_time,
689 	   idle_time,
690 	   xprt->stat.sends,
691 	   xprt->stat.recvs,
692 	   xprt->stat.bad_xids,
693 	   xprt->stat.req_u,
694 	   xprt->stat.bklog_u,
695 
696 	   r_xprt->rx_stats.read_chunk_count,
697 	   r_xprt->rx_stats.write_chunk_count,
698 	   r_xprt->rx_stats.reply_chunk_count,
699 	   r_xprt->rx_stats.total_rdma_request,
700 	   r_xprt->rx_stats.total_rdma_reply,
701 	   r_xprt->rx_stats.pullup_copy_count,
702 	   r_xprt->rx_stats.fixup_copy_count,
703 	   r_xprt->rx_stats.hardway_register_count,
704 	   r_xprt->rx_stats.failed_marshal_count,
705 	   r_xprt->rx_stats.bad_reply_count);
706 }
707 
708 /*
709  * Plumbing for rpc transport switch and kernel module
710  */
711 
712 static struct rpc_xprt_ops xprt_rdma_procs = {
713 	.reserve_xprt		= xprt_rdma_reserve_xprt,
714 	.release_xprt		= xprt_release_xprt_cong, /* sunrpc/xprt.c */
715 	.release_request	= xprt_release_rqst_cong,       /* ditto */
716 	.set_retrans_timeout	= xprt_set_retrans_timeout_def, /* ditto */
717 	.rpcbind		= rpcb_getport_async,	/* sunrpc/rpcb_clnt.c */
718 	.set_port		= xprt_rdma_set_port,
719 	.connect		= xprt_rdma_connect,
720 	.buf_alloc		= xprt_rdma_allocate,
721 	.buf_free		= xprt_rdma_free,
722 	.send_request		= xprt_rdma_send_request,
723 	.close			= xprt_rdma_close,
724 	.destroy		= xprt_rdma_destroy,
725 	.print_stats		= xprt_rdma_print_stats
726 };
727 
728 static struct xprt_class xprt_rdma = {
729 	.list			= LIST_HEAD_INIT(xprt_rdma.list),
730 	.name			= "rdma",
731 	.owner			= THIS_MODULE,
732 	.ident			= XPRT_TRANSPORT_RDMA,
733 	.setup			= xprt_setup_rdma,
734 };
735 
736 static void __exit xprt_rdma_cleanup(void)
737 {
738 	int rc;
739 
740 	dprintk(KERN_INFO "RPCRDMA Module Removed, deregister RPC RDMA transport\n");
741 #ifdef RPC_DEBUG
742 	if (sunrpc_table_header) {
743 		unregister_sysctl_table(sunrpc_table_header);
744 		sunrpc_table_header = NULL;
745 	}
746 #endif
747 	rc = xprt_unregister_transport(&xprt_rdma);
748 	if (rc)
749 		dprintk("RPC:       %s: xprt_unregister returned %i\n",
750 			__func__, rc);
751 }
752 
753 static int __init xprt_rdma_init(void)
754 {
755 	int rc;
756 
757 	rc = xprt_register_transport(&xprt_rdma);
758 
759 	if (rc)
760 		return rc;
761 
762 	dprintk(KERN_INFO "RPCRDMA Module Init, register RPC RDMA transport\n");
763 
764 	dprintk(KERN_INFO "Defaults:\n");
765 	dprintk(KERN_INFO "\tSlots %d\n"
766 		"\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
767 		xprt_rdma_slot_table_entries,
768 		xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
769 	dprintk(KERN_INFO "\tPadding %d\n\tMemreg %d\n",
770 		xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
771 
772 #ifdef RPC_DEBUG
773 	if (!sunrpc_table_header)
774 		sunrpc_table_header = register_sysctl_table(sunrpc_table);
775 #endif
776 	return 0;
777 }
778 
779 module_init(xprt_rdma_init);
780 module_exit(xprt_rdma_cleanup);
781