xref: /openbmc/linux/net/sunrpc/xprtrdma/transport.c (revision ee89bd6b)
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 #ifdef RPC_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 = 0;
77 
78 #ifdef RPC_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 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 ctl_table sunrpc_table[] = {
142 	{
143 		.procname	= "sunrpc",
144 		.mode		= 0555,
145 		.child		= xr_tunables_table
146 	},
147 	{ },
148 };
149 
150 #endif
151 
152 static struct rpc_xprt_ops xprt_rdma_procs;	/* forward reference */
153 
154 static void
155 xprt_rdma_format_addresses(struct rpc_xprt *xprt)
156 {
157 	struct sockaddr *sap = (struct sockaddr *)
158 					&rpcx_to_rdmad(xprt).addr;
159 	struct sockaddr_in *sin = (struct sockaddr_in *)sap;
160 	char buf[64];
161 
162 	(void)rpc_ntop(sap, buf, sizeof(buf));
163 	xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
164 
165 	snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
166 	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
167 
168 	xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
169 
170 	snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
171 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
172 
173 	snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
174 	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
175 
176 	/* netid */
177 	xprt->address_strings[RPC_DISPLAY_NETID] = "rdma";
178 }
179 
180 static void
181 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
182 {
183 	unsigned int i;
184 
185 	for (i = 0; i < RPC_DISPLAY_MAX; i++)
186 		switch (i) {
187 		case RPC_DISPLAY_PROTO:
188 		case RPC_DISPLAY_NETID:
189 			continue;
190 		default:
191 			kfree(xprt->address_strings[i]);
192 		}
193 }
194 
195 static void
196 xprt_rdma_connect_worker(struct work_struct *work)
197 {
198 	struct rpcrdma_xprt *r_xprt =
199 		container_of(work, struct rpcrdma_xprt, rdma_connect.work);
200 	struct rpc_xprt *xprt = &r_xprt->xprt;
201 	int rc = 0;
202 
203 	current->flags |= PF_FSTRANS;
204 	xprt_clear_connected(xprt);
205 
206 	dprintk("RPC:       %s: %sconnect\n", __func__,
207 			r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
208 	rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
209 	if (rc)
210 		xprt_wake_pending_tasks(xprt, rc);
211 
212 	dprintk("RPC:       %s: exit\n", __func__);
213 	xprt_clear_connecting(xprt);
214 	current->flags &= ~PF_FSTRANS;
215 }
216 
217 /*
218  * xprt_rdma_destroy
219  *
220  * Destroy the xprt.
221  * Free all memory associated with the object, including its own.
222  * NOTE: none of the *destroy methods free memory for their top-level
223  * objects, even though they may have allocated it (they do free
224  * private memory). It's up to the caller to handle it. In this
225  * case (RDMA transport), all structure memory is inlined with the
226  * struct rpcrdma_xprt.
227  */
228 static void
229 xprt_rdma_destroy(struct rpc_xprt *xprt)
230 {
231 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
232 	int rc;
233 
234 	dprintk("RPC:       %s: called\n", __func__);
235 
236 	cancel_delayed_work_sync(&r_xprt->rdma_connect);
237 
238 	xprt_clear_connected(xprt);
239 
240 	rpcrdma_buffer_destroy(&r_xprt->rx_buf);
241 	rc = rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
242 	if (rc)
243 		dprintk("RPC:       %s: rpcrdma_ep_destroy returned %i\n",
244 			__func__, rc);
245 	rpcrdma_ia_close(&r_xprt->rx_ia);
246 
247 	xprt_rdma_free_addresses(xprt);
248 
249 	xprt_free(xprt);
250 
251 	dprintk("RPC:       %s: returning\n", __func__);
252 
253 	module_put(THIS_MODULE);
254 }
255 
256 static const struct rpc_timeout xprt_rdma_default_timeout = {
257 	.to_initval = 60 * HZ,
258 	.to_maxval = 60 * HZ,
259 };
260 
261 /**
262  * xprt_setup_rdma - Set up transport to use RDMA
263  *
264  * @args: rpc transport arguments
265  */
266 static struct rpc_xprt *
267 xprt_setup_rdma(struct xprt_create *args)
268 {
269 	struct rpcrdma_create_data_internal cdata;
270 	struct rpc_xprt *xprt;
271 	struct rpcrdma_xprt *new_xprt;
272 	struct rpcrdma_ep *new_ep;
273 	struct sockaddr_in *sin;
274 	int rc;
275 
276 	if (args->addrlen > sizeof(xprt->addr)) {
277 		dprintk("RPC:       %s: address too large\n", __func__);
278 		return ERR_PTR(-EBADF);
279 	}
280 
281 	xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
282 			xprt_rdma_slot_table_entries,
283 			xprt_rdma_slot_table_entries);
284 	if (xprt == NULL) {
285 		dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
286 			__func__);
287 		return ERR_PTR(-ENOMEM);
288 	}
289 
290 	/* 60 second timeout, no retries */
291 	xprt->timeout = &xprt_rdma_default_timeout;
292 	xprt->bind_timeout = (60U * HZ);
293 	xprt->reestablish_timeout = (5U * HZ);
294 	xprt->idle_timeout = (5U * 60 * HZ);
295 
296 	xprt->resvport = 0;		/* privileged port not needed */
297 	xprt->tsh_size = 0;		/* RPC-RDMA handles framing */
298 	xprt->max_payload = RPCRDMA_MAX_DATA_SEGS * PAGE_SIZE;
299 	xprt->ops = &xprt_rdma_procs;
300 
301 	/*
302 	 * Set up RDMA-specific connect data.
303 	 */
304 
305 	/* Put server RDMA address in local cdata */
306 	memcpy(&cdata.addr, args->dstaddr, args->addrlen);
307 
308 	/* Ensure xprt->addr holds valid server TCP (not RDMA)
309 	 * address, for any side protocols which peek at it */
310 	xprt->prot = IPPROTO_TCP;
311 	xprt->addrlen = args->addrlen;
312 	memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
313 
314 	sin = (struct sockaddr_in *)&cdata.addr;
315 	if (ntohs(sin->sin_port) != 0)
316 		xprt_set_bound(xprt);
317 
318 	dprintk("RPC:       %s: %pI4:%u\n",
319 		__func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port));
320 
321 	/* Set max requests */
322 	cdata.max_requests = xprt->max_reqs;
323 
324 	/* Set some length limits */
325 	cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
326 	cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
327 
328 	cdata.inline_wsize = xprt_rdma_max_inline_write;
329 	if (cdata.inline_wsize > cdata.wsize)
330 		cdata.inline_wsize = cdata.wsize;
331 
332 	cdata.inline_rsize = xprt_rdma_max_inline_read;
333 	if (cdata.inline_rsize > cdata.rsize)
334 		cdata.inline_rsize = cdata.rsize;
335 
336 	cdata.padding = xprt_rdma_inline_write_padding;
337 
338 	/*
339 	 * Create new transport instance, which includes initialized
340 	 *  o ia
341 	 *  o endpoint
342 	 *  o buffers
343 	 */
344 
345 	new_xprt = rpcx_to_rdmax(xprt);
346 
347 	rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
348 				xprt_rdma_memreg_strategy);
349 	if (rc)
350 		goto out1;
351 
352 	/*
353 	 * initialize and create ep
354 	 */
355 	new_xprt->rx_data = cdata;
356 	new_ep = &new_xprt->rx_ep;
357 	new_ep->rep_remote_addr = cdata.addr;
358 
359 	rc = rpcrdma_ep_create(&new_xprt->rx_ep,
360 				&new_xprt->rx_ia, &new_xprt->rx_data);
361 	if (rc)
362 		goto out2;
363 
364 	/*
365 	 * Allocate pre-registered send and receive buffers for headers and
366 	 * any inline data. Also specify any padding which will be provided
367 	 * from a preregistered zero buffer.
368 	 */
369 	rc = rpcrdma_buffer_create(&new_xprt->rx_buf, new_ep, &new_xprt->rx_ia,
370 				&new_xprt->rx_data);
371 	if (rc)
372 		goto out3;
373 
374 	/*
375 	 * Register a callback for connection events. This is necessary because
376 	 * connection loss notification is async. We also catch connection loss
377 	 * when reaping receives.
378 	 */
379 	INIT_DELAYED_WORK(&new_xprt->rdma_connect, xprt_rdma_connect_worker);
380 	new_ep->rep_func = rpcrdma_conn_func;
381 	new_ep->rep_xprt = xprt;
382 
383 	xprt_rdma_format_addresses(xprt);
384 
385 	if (!try_module_get(THIS_MODULE))
386 		goto out4;
387 
388 	return xprt;
389 
390 out4:
391 	xprt_rdma_free_addresses(xprt);
392 	rc = -EINVAL;
393 out3:
394 	(void) rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
395 out2:
396 	rpcrdma_ia_close(&new_xprt->rx_ia);
397 out1:
398 	xprt_free(xprt);
399 	return ERR_PTR(rc);
400 }
401 
402 /*
403  * Close a connection, during shutdown or timeout/reconnect
404  */
405 static void
406 xprt_rdma_close(struct rpc_xprt *xprt)
407 {
408 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
409 
410 	dprintk("RPC:       %s: closing\n", __func__);
411 	if (r_xprt->rx_ep.rep_connected > 0)
412 		xprt->reestablish_timeout = 0;
413 	xprt_disconnect_done(xprt);
414 	(void) rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
415 }
416 
417 static void
418 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
419 {
420 	struct sockaddr_in *sap;
421 
422 	sap = (struct sockaddr_in *)&xprt->addr;
423 	sap->sin_port = htons(port);
424 	sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
425 	sap->sin_port = htons(port);
426 	dprintk("RPC:       %s: %u\n", __func__, port);
427 }
428 
429 static void
430 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
431 {
432 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
433 
434 	if (r_xprt->rx_ep.rep_connected != 0) {
435 		/* Reconnect */
436 		schedule_delayed_work(&r_xprt->rdma_connect,
437 			xprt->reestablish_timeout);
438 		xprt->reestablish_timeout <<= 1;
439 		if (xprt->reestablish_timeout > (30 * HZ))
440 			xprt->reestablish_timeout = (30 * HZ);
441 		else if (xprt->reestablish_timeout < (5 * HZ))
442 			xprt->reestablish_timeout = (5 * HZ);
443 	} else {
444 		schedule_delayed_work(&r_xprt->rdma_connect, 0);
445 		if (!RPC_IS_ASYNC(task))
446 			flush_delayed_work(&r_xprt->rdma_connect);
447 	}
448 }
449 
450 static int
451 xprt_rdma_reserve_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
452 {
453 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
454 	int credits = atomic_read(&r_xprt->rx_buf.rb_credits);
455 
456 	/* == RPC_CWNDSCALE @ init, but *after* setup */
457 	if (r_xprt->rx_buf.rb_cwndscale == 0UL) {
458 		r_xprt->rx_buf.rb_cwndscale = xprt->cwnd;
459 		dprintk("RPC:       %s: cwndscale %lu\n", __func__,
460 			r_xprt->rx_buf.rb_cwndscale);
461 		BUG_ON(r_xprt->rx_buf.rb_cwndscale <= 0);
462 	}
463 	xprt->cwnd = credits * r_xprt->rx_buf.rb_cwndscale;
464 	return xprt_reserve_xprt_cong(xprt, task);
465 }
466 
467 /*
468  * The RDMA allocate/free functions need the task structure as a place
469  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
470  * sequence. For this reason, the recv buffers are attached to send
471  * buffers for portions of the RPC. Note that the RPC layer allocates
472  * both send and receive buffers in the same call. We may register
473  * the receive buffer portion when using reply chunks.
474  */
475 static void *
476 xprt_rdma_allocate(struct rpc_task *task, size_t size)
477 {
478 	struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
479 	struct rpcrdma_req *req, *nreq;
480 
481 	req = rpcrdma_buffer_get(&rpcx_to_rdmax(xprt)->rx_buf);
482 	BUG_ON(NULL == req);
483 
484 	if (size > req->rl_size) {
485 		dprintk("RPC:       %s: size %zd too large for buffer[%zd]: "
486 			"prog %d vers %d proc %d\n",
487 			__func__, size, req->rl_size,
488 			task->tk_client->cl_prog, task->tk_client->cl_vers,
489 			task->tk_msg.rpc_proc->p_proc);
490 		/*
491 		 * Outgoing length shortage. Our inline write max must have
492 		 * been configured to perform direct i/o.
493 		 *
494 		 * This is therefore a large metadata operation, and the
495 		 * allocate call was made on the maximum possible message,
496 		 * e.g. containing long filename(s) or symlink data. In
497 		 * fact, while these metadata operations *might* carry
498 		 * large outgoing payloads, they rarely *do*. However, we
499 		 * have to commit to the request here, so reallocate and
500 		 * register it now. The data path will never require this
501 		 * reallocation.
502 		 *
503 		 * If the allocation or registration fails, the RPC framework
504 		 * will (doggedly) retry.
505 		 */
506 		if (rpcx_to_rdmax(xprt)->rx_ia.ri_memreg_strategy ==
507 				RPCRDMA_BOUNCEBUFFERS) {
508 			/* forced to "pure inline" */
509 			dprintk("RPC:       %s: too much data (%zd) for inline "
510 					"(r/w max %d/%d)\n", __func__, size,
511 					rpcx_to_rdmad(xprt).inline_rsize,
512 					rpcx_to_rdmad(xprt).inline_wsize);
513 			size = req->rl_size;
514 			rpc_exit(task, -EIO);		/* fail the operation */
515 			rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
516 			goto out;
517 		}
518 		if (task->tk_flags & RPC_TASK_SWAPPER)
519 			nreq = kmalloc(sizeof *req + size, GFP_ATOMIC);
520 		else
521 			nreq = kmalloc(sizeof *req + size, GFP_NOFS);
522 		if (nreq == NULL)
523 			goto outfail;
524 
525 		if (rpcrdma_register_internal(&rpcx_to_rdmax(xprt)->rx_ia,
526 				nreq->rl_base, size + sizeof(struct rpcrdma_req)
527 				- offsetof(struct rpcrdma_req, rl_base),
528 				&nreq->rl_handle, &nreq->rl_iov)) {
529 			kfree(nreq);
530 			goto outfail;
531 		}
532 		rpcx_to_rdmax(xprt)->rx_stats.hardway_register_count += size;
533 		nreq->rl_size = size;
534 		nreq->rl_niovs = 0;
535 		nreq->rl_nchunks = 0;
536 		nreq->rl_buffer = (struct rpcrdma_buffer *)req;
537 		nreq->rl_reply = req->rl_reply;
538 		memcpy(nreq->rl_segments,
539 			req->rl_segments, sizeof nreq->rl_segments);
540 		/* flag the swap with an unused field */
541 		nreq->rl_iov.length = 0;
542 		req->rl_reply = NULL;
543 		req = nreq;
544 	}
545 	dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
546 out:
547 	req->rl_connect_cookie = 0;	/* our reserved value */
548 	return req->rl_xdr_buf;
549 
550 outfail:
551 	rpcrdma_buffer_put(req);
552 	rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
553 	return NULL;
554 }
555 
556 /*
557  * This function returns all RDMA resources to the pool.
558  */
559 static void
560 xprt_rdma_free(void *buffer)
561 {
562 	struct rpcrdma_req *req;
563 	struct rpcrdma_xprt *r_xprt;
564 	struct rpcrdma_rep *rep;
565 	int i;
566 
567 	if (buffer == NULL)
568 		return;
569 
570 	req = container_of(buffer, struct rpcrdma_req, rl_xdr_buf[0]);
571 	if (req->rl_iov.length == 0) {	/* see allocate above */
572 		r_xprt = container_of(((struct rpcrdma_req *) req->rl_buffer)->rl_buffer,
573 				      struct rpcrdma_xprt, rx_buf);
574 	} else
575 		r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
576 	rep = req->rl_reply;
577 
578 	dprintk("RPC:       %s: called on 0x%p%s\n",
579 		__func__, rep, (rep && rep->rr_func) ? " (with waiter)" : "");
580 
581 	/*
582 	 * Finish the deregistration. When using mw bind, this was
583 	 * begun in rpcrdma_reply_handler(). In all other modes, we
584 	 * do it here, in thread context. The process is considered
585 	 * complete when the rr_func vector becomes NULL - this
586 	 * was put in place during rpcrdma_reply_handler() - the wait
587 	 * call below will not block if the dereg is "done". If
588 	 * interrupted, our framework will clean up.
589 	 */
590 	for (i = 0; req->rl_nchunks;) {
591 		--req->rl_nchunks;
592 		i += rpcrdma_deregister_external(
593 			&req->rl_segments[i], r_xprt, NULL);
594 	}
595 
596 	if (rep && wait_event_interruptible(rep->rr_unbind, !rep->rr_func)) {
597 		rep->rr_func = NULL;	/* abandon the callback */
598 		req->rl_reply = NULL;
599 	}
600 
601 	if (req->rl_iov.length == 0) {	/* see allocate above */
602 		struct rpcrdma_req *oreq = (struct rpcrdma_req *)req->rl_buffer;
603 		oreq->rl_reply = req->rl_reply;
604 		(void) rpcrdma_deregister_internal(&r_xprt->rx_ia,
605 						   req->rl_handle,
606 						   &req->rl_iov);
607 		kfree(req);
608 		req = oreq;
609 	}
610 
611 	/* Put back request+reply buffers */
612 	rpcrdma_buffer_put(req);
613 }
614 
615 /*
616  * send_request invokes the meat of RPC RDMA. It must do the following:
617  *  1.  Marshal the RPC request into an RPC RDMA request, which means
618  *	putting a header in front of data, and creating IOVs for RDMA
619  *	from those in the request.
620  *  2.  In marshaling, detect opportunities for RDMA, and use them.
621  *  3.  Post a recv message to set up asynch completion, then send
622  *	the request (rpcrdma_ep_post).
623  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
624  */
625 
626 static int
627 xprt_rdma_send_request(struct rpc_task *task)
628 {
629 	struct rpc_rqst *rqst = task->tk_rqstp;
630 	struct rpc_xprt *xprt = rqst->rq_xprt;
631 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
632 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
633 
634 	/* marshal the send itself */
635 	if (req->rl_niovs == 0 && rpcrdma_marshal_req(rqst) != 0) {
636 		r_xprt->rx_stats.failed_marshal_count++;
637 		dprintk("RPC:       %s: rpcrdma_marshal_req failed\n",
638 			__func__);
639 		return -EIO;
640 	}
641 
642 	if (req->rl_reply == NULL) 		/* e.g. reconnection */
643 		rpcrdma_recv_buffer_get(req);
644 
645 	if (req->rl_reply) {
646 		req->rl_reply->rr_func = rpcrdma_reply_handler;
647 		/* this need only be done once, but... */
648 		req->rl_reply->rr_xprt = xprt;
649 	}
650 
651 	/* Must suppress retransmit to maintain credits */
652 	if (req->rl_connect_cookie == xprt->connect_cookie)
653 		goto drop_connection;
654 	req->rl_connect_cookie = xprt->connect_cookie;
655 
656 	if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
657 		goto drop_connection;
658 
659 	rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
660 	rqst->rq_bytes_sent = 0;
661 	return 0;
662 
663 drop_connection:
664 	xprt_disconnect_done(xprt);
665 	return -ENOTCONN;	/* implies disconnect */
666 }
667 
668 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
669 {
670 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
671 	long idle_time = 0;
672 
673 	if (xprt_connected(xprt))
674 		idle_time = (long)(jiffies - xprt->last_used) / HZ;
675 
676 	seq_printf(seq,
677 	  "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
678 	  "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
679 
680 	   0,	/* need a local port? */
681 	   xprt->stat.bind_count,
682 	   xprt->stat.connect_count,
683 	   xprt->stat.connect_time,
684 	   idle_time,
685 	   xprt->stat.sends,
686 	   xprt->stat.recvs,
687 	   xprt->stat.bad_xids,
688 	   xprt->stat.req_u,
689 	   xprt->stat.bklog_u,
690 
691 	   r_xprt->rx_stats.read_chunk_count,
692 	   r_xprt->rx_stats.write_chunk_count,
693 	   r_xprt->rx_stats.reply_chunk_count,
694 	   r_xprt->rx_stats.total_rdma_request,
695 	   r_xprt->rx_stats.total_rdma_reply,
696 	   r_xprt->rx_stats.pullup_copy_count,
697 	   r_xprt->rx_stats.fixup_copy_count,
698 	   r_xprt->rx_stats.hardway_register_count,
699 	   r_xprt->rx_stats.failed_marshal_count,
700 	   r_xprt->rx_stats.bad_reply_count);
701 }
702 
703 /*
704  * Plumbing for rpc transport switch and kernel module
705  */
706 
707 static struct rpc_xprt_ops xprt_rdma_procs = {
708 	.reserve_xprt		= xprt_rdma_reserve_xprt,
709 	.release_xprt		= xprt_release_xprt_cong, /* sunrpc/xprt.c */
710 	.alloc_slot		= xprt_alloc_slot,
711 	.release_request	= xprt_release_rqst_cong,       /* ditto */
712 	.set_retrans_timeout	= xprt_set_retrans_timeout_def, /* ditto */
713 	.rpcbind		= rpcb_getport_async,	/* sunrpc/rpcb_clnt.c */
714 	.set_port		= xprt_rdma_set_port,
715 	.connect		= xprt_rdma_connect,
716 	.buf_alloc		= xprt_rdma_allocate,
717 	.buf_free		= xprt_rdma_free,
718 	.send_request		= xprt_rdma_send_request,
719 	.close			= xprt_rdma_close,
720 	.destroy		= xprt_rdma_destroy,
721 	.print_stats		= xprt_rdma_print_stats
722 };
723 
724 static struct xprt_class xprt_rdma = {
725 	.list			= LIST_HEAD_INIT(xprt_rdma.list),
726 	.name			= "rdma",
727 	.owner			= THIS_MODULE,
728 	.ident			= XPRT_TRANSPORT_RDMA,
729 	.setup			= xprt_setup_rdma,
730 };
731 
732 static void __exit xprt_rdma_cleanup(void)
733 {
734 	int rc;
735 
736 	dprintk(KERN_INFO "RPCRDMA Module Removed, deregister RPC RDMA transport\n");
737 #ifdef RPC_DEBUG
738 	if (sunrpc_table_header) {
739 		unregister_sysctl_table(sunrpc_table_header);
740 		sunrpc_table_header = NULL;
741 	}
742 #endif
743 	rc = xprt_unregister_transport(&xprt_rdma);
744 	if (rc)
745 		dprintk("RPC:       %s: xprt_unregister returned %i\n",
746 			__func__, rc);
747 }
748 
749 static int __init xprt_rdma_init(void)
750 {
751 	int rc;
752 
753 	rc = xprt_register_transport(&xprt_rdma);
754 
755 	if (rc)
756 		return rc;
757 
758 	dprintk(KERN_INFO "RPCRDMA Module Init, register RPC RDMA transport\n");
759 
760 	dprintk(KERN_INFO "Defaults:\n");
761 	dprintk(KERN_INFO "\tSlots %d\n"
762 		"\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
763 		xprt_rdma_slot_table_entries,
764 		xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
765 	dprintk(KERN_INFO "\tPadding %d\n\tMemreg %d\n",
766 		xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
767 
768 #ifdef RPC_DEBUG
769 	if (!sunrpc_table_header)
770 		sunrpc_table_header = register_sysctl_table(sunrpc_table);
771 #endif
772 	return 0;
773 }
774 
775 module_init(xprt_rdma_init);
776 module_exit(xprt_rdma_cleanup);
777