xref: /linux/net/sunrpc/xprtrdma/transport.c (revision 6d4561110a3e9fa742aeec6717248a491dfb1878)
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/seq_file.h>
53 
54 #include "xprt_rdma.h"
55 
56 #ifdef RPC_DEBUG
57 # define RPCDBG_FACILITY	RPCDBG_TRANS
58 #endif
59 
60 MODULE_LICENSE("Dual BSD/GPL");
61 
62 MODULE_DESCRIPTION("RPC/RDMA Transport for Linux kernel NFS");
63 MODULE_AUTHOR("Network Appliance, Inc.");
64 
65 /*
66  * tunables
67  */
68 
69 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
70 static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
71 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
72 static unsigned int xprt_rdma_inline_write_padding;
73 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
74                 int xprt_rdma_pad_optimize = 0;
75 
76 #ifdef RPC_DEBUG
77 
78 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
79 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
80 static unsigned int zero;
81 static unsigned int max_padding = PAGE_SIZE;
82 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
83 static unsigned int max_memreg = RPCRDMA_LAST - 1;
84 
85 static struct ctl_table_header *sunrpc_table_header;
86 
87 static ctl_table xr_tunables_table[] = {
88 	{
89 		.procname	= "rdma_slot_table_entries",
90 		.data		= &xprt_rdma_slot_table_entries,
91 		.maxlen		= sizeof(unsigned int),
92 		.mode		= 0644,
93 		.proc_handler	= proc_dointvec_minmax,
94 		.extra1		= &min_slot_table_size,
95 		.extra2		= &max_slot_table_size
96 	},
97 	{
98 		.procname	= "rdma_max_inline_read",
99 		.data		= &xprt_rdma_max_inline_read,
100 		.maxlen		= sizeof(unsigned int),
101 		.mode		= 0644,
102 		.proc_handler	= proc_dointvec,
103 	},
104 	{
105 		.procname	= "rdma_max_inline_write",
106 		.data		= &xprt_rdma_max_inline_write,
107 		.maxlen		= sizeof(unsigned int),
108 		.mode		= 0644,
109 		.proc_handler	= proc_dointvec,
110 	},
111 	{
112 		.procname	= "rdma_inline_write_padding",
113 		.data		= &xprt_rdma_inline_write_padding,
114 		.maxlen		= sizeof(unsigned int),
115 		.mode		= 0644,
116 		.proc_handler	= proc_dointvec_minmax,
117 		.extra1		= &zero,
118 		.extra2		= &max_padding,
119 	},
120 	{
121 		.procname	= "rdma_memreg_strategy",
122 		.data		= &xprt_rdma_memreg_strategy,
123 		.maxlen		= sizeof(unsigned int),
124 		.mode		= 0644,
125 		.proc_handler	= proc_dointvec_minmax,
126 		.extra1		= &min_memreg,
127 		.extra2		= &max_memreg,
128 	},
129 	{
130 		.procname	= "rdma_pad_optimize",
131 		.data		= &xprt_rdma_pad_optimize,
132 		.maxlen		= sizeof(unsigned int),
133 		.mode		= 0644,
134 		.proc_handler	= proc_dointvec,
135 	},
136 	{ },
137 };
138 
139 static ctl_table sunrpc_table[] = {
140 	{
141 		.procname	= "sunrpc",
142 		.mode		= 0555,
143 		.child		= xr_tunables_table
144 	},
145 	{ },
146 };
147 
148 #endif
149 
150 static struct rpc_xprt_ops xprt_rdma_procs;	/* forward reference */
151 
152 static void
153 xprt_rdma_format_addresses(struct rpc_xprt *xprt)
154 {
155 	struct sockaddr *sap = (struct sockaddr *)
156 					&rpcx_to_rdmad(xprt).addr;
157 	struct sockaddr_in *sin = (struct sockaddr_in *)sap;
158 	char buf[64];
159 
160 	(void)rpc_ntop(sap, buf, sizeof(buf));
161 	xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
162 
163 	(void)snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
164 	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
165 
166 	xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
167 
168 	(void)snprintf(buf, sizeof(buf), "%02x%02x%02x%02x",
169 				NIPQUAD(sin->sin_addr.s_addr));
170 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
171 
172 	(void)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(&r_xprt->rdma_connect);
241 	flush_scheduled_work();
242 
243 	xprt_clear_connected(xprt);
244 
245 	rpcrdma_buffer_destroy(&r_xprt->rx_buf);
246 	rc = rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
247 	if (rc)
248 		dprintk("RPC:       %s: rpcrdma_ep_destroy returned %i\n",
249 			__func__, rc);
250 	rpcrdma_ia_close(&r_xprt->rx_ia);
251 
252 	xprt_rdma_free_addresses(xprt);
253 
254 	kfree(xprt->slot);
255 	xprt->slot = NULL;
256 	kfree(xprt);
257 
258 	dprintk("RPC:       %s: returning\n", __func__);
259 
260 	module_put(THIS_MODULE);
261 }
262 
263 static const struct rpc_timeout xprt_rdma_default_timeout = {
264 	.to_initval = 60 * HZ,
265 	.to_maxval = 60 * HZ,
266 };
267 
268 /**
269  * xprt_setup_rdma - Set up transport to use RDMA
270  *
271  * @args: rpc transport arguments
272  */
273 static struct rpc_xprt *
274 xprt_setup_rdma(struct xprt_create *args)
275 {
276 	struct rpcrdma_create_data_internal cdata;
277 	struct rpc_xprt *xprt;
278 	struct rpcrdma_xprt *new_xprt;
279 	struct rpcrdma_ep *new_ep;
280 	struct sockaddr_in *sin;
281 	int rc;
282 
283 	if (args->addrlen > sizeof(xprt->addr)) {
284 		dprintk("RPC:       %s: address too large\n", __func__);
285 		return ERR_PTR(-EBADF);
286 	}
287 
288 	xprt = kzalloc(sizeof(struct rpcrdma_xprt), GFP_KERNEL);
289 	if (xprt == NULL) {
290 		dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
291 			__func__);
292 		return ERR_PTR(-ENOMEM);
293 	}
294 
295 	xprt->max_reqs = xprt_rdma_slot_table_entries;
296 	xprt->slot = kcalloc(xprt->max_reqs,
297 				sizeof(struct rpc_rqst), GFP_KERNEL);
298 	if (xprt->slot == NULL) {
299 		dprintk("RPC:       %s: couldn't allocate %d slots\n",
300 			__func__, xprt->max_reqs);
301 		kfree(xprt);
302 		return ERR_PTR(-ENOMEM);
303 	}
304 
305 	/* 60 second timeout, no retries */
306 	xprt->timeout = &xprt_rdma_default_timeout;
307 	xprt->bind_timeout = (60U * HZ);
308 	xprt->connect_timeout = (60U * HZ);
309 	xprt->reestablish_timeout = (5U * HZ);
310 	xprt->idle_timeout = (5U * 60 * HZ);
311 
312 	xprt->resvport = 0;		/* privileged port not needed */
313 	xprt->tsh_size = 0;		/* RPC-RDMA handles framing */
314 	xprt->max_payload = RPCRDMA_MAX_DATA_SEGS * PAGE_SIZE;
315 	xprt->ops = &xprt_rdma_procs;
316 
317 	/*
318 	 * Set up RDMA-specific connect data.
319 	 */
320 
321 	/* Put server RDMA address in local cdata */
322 	memcpy(&cdata.addr, args->dstaddr, args->addrlen);
323 
324 	/* Ensure xprt->addr holds valid server TCP (not RDMA)
325 	 * address, for any side protocols which peek at it */
326 	xprt->prot = IPPROTO_TCP;
327 	xprt->addrlen = args->addrlen;
328 	memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
329 
330 	sin = (struct sockaddr_in *)&cdata.addr;
331 	if (ntohs(sin->sin_port) != 0)
332 		xprt_set_bound(xprt);
333 
334 	dprintk("RPC:       %s: %pI4:%u\n",
335 		__func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port));
336 
337 	/* Set max requests */
338 	cdata.max_requests = xprt->max_reqs;
339 
340 	/* Set some length limits */
341 	cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
342 	cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
343 
344 	cdata.inline_wsize = xprt_rdma_max_inline_write;
345 	if (cdata.inline_wsize > cdata.wsize)
346 		cdata.inline_wsize = cdata.wsize;
347 
348 	cdata.inline_rsize = xprt_rdma_max_inline_read;
349 	if (cdata.inline_rsize > cdata.rsize)
350 		cdata.inline_rsize = cdata.rsize;
351 
352 	cdata.padding = xprt_rdma_inline_write_padding;
353 
354 	/*
355 	 * Create new transport instance, which includes initialized
356 	 *  o ia
357 	 *  o endpoint
358 	 *  o buffers
359 	 */
360 
361 	new_xprt = rpcx_to_rdmax(xprt);
362 
363 	rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
364 				xprt_rdma_memreg_strategy);
365 	if (rc)
366 		goto out1;
367 
368 	/*
369 	 * initialize and create ep
370 	 */
371 	new_xprt->rx_data = cdata;
372 	new_ep = &new_xprt->rx_ep;
373 	new_ep->rep_remote_addr = cdata.addr;
374 
375 	rc = rpcrdma_ep_create(&new_xprt->rx_ep,
376 				&new_xprt->rx_ia, &new_xprt->rx_data);
377 	if (rc)
378 		goto out2;
379 
380 	/*
381 	 * Allocate pre-registered send and receive buffers for headers and
382 	 * any inline data. Also specify any padding which will be provided
383 	 * from a preregistered zero buffer.
384 	 */
385 	rc = rpcrdma_buffer_create(&new_xprt->rx_buf, new_ep, &new_xprt->rx_ia,
386 				&new_xprt->rx_data);
387 	if (rc)
388 		goto out3;
389 
390 	/*
391 	 * Register a callback for connection events. This is necessary because
392 	 * connection loss notification is async. We also catch connection loss
393 	 * when reaping receives.
394 	 */
395 	INIT_DELAYED_WORK(&new_xprt->rdma_connect, xprt_rdma_connect_worker);
396 	new_ep->rep_func = rpcrdma_conn_func;
397 	new_ep->rep_xprt = xprt;
398 
399 	xprt_rdma_format_addresses(xprt);
400 
401 	if (!try_module_get(THIS_MODULE))
402 		goto out4;
403 
404 	return xprt;
405 
406 out4:
407 	xprt_rdma_free_addresses(xprt);
408 	rc = -EINVAL;
409 out3:
410 	(void) rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
411 out2:
412 	rpcrdma_ia_close(&new_xprt->rx_ia);
413 out1:
414 	kfree(xprt->slot);
415 	kfree(xprt);
416 	return ERR_PTR(rc);
417 }
418 
419 /*
420  * Close a connection, during shutdown or timeout/reconnect
421  */
422 static void
423 xprt_rdma_close(struct rpc_xprt *xprt)
424 {
425 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
426 
427 	dprintk("RPC:       %s: closing\n", __func__);
428 	if (r_xprt->rx_ep.rep_connected > 0)
429 		xprt->reestablish_timeout = 0;
430 	xprt_disconnect_done(xprt);
431 	(void) rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
432 }
433 
434 static void
435 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
436 {
437 	struct sockaddr_in *sap;
438 
439 	sap = (struct sockaddr_in *)&xprt->addr;
440 	sap->sin_port = htons(port);
441 	sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
442 	sap->sin_port = htons(port);
443 	dprintk("RPC:       %s: %u\n", __func__, port);
444 }
445 
446 static void
447 xprt_rdma_connect(struct rpc_task *task)
448 {
449 	struct rpc_xprt *xprt = (struct rpc_xprt *)task->tk_xprt;
450 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
451 
452 	if (!xprt_test_and_set_connecting(xprt)) {
453 		if (r_xprt->rx_ep.rep_connected != 0) {
454 			/* Reconnect */
455 			schedule_delayed_work(&r_xprt->rdma_connect,
456 				xprt->reestablish_timeout);
457 			xprt->reestablish_timeout <<= 1;
458 			if (xprt->reestablish_timeout > (30 * HZ))
459 				xprt->reestablish_timeout = (30 * HZ);
460 			else if (xprt->reestablish_timeout < (5 * HZ))
461 				xprt->reestablish_timeout = (5 * HZ);
462 		} else {
463 			schedule_delayed_work(&r_xprt->rdma_connect, 0);
464 			if (!RPC_IS_ASYNC(task))
465 				flush_scheduled_work();
466 		}
467 	}
468 }
469 
470 static int
471 xprt_rdma_reserve_xprt(struct rpc_task *task)
472 {
473 	struct rpc_xprt *xprt = task->tk_xprt;
474 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
475 	int credits = atomic_read(&r_xprt->rx_buf.rb_credits);
476 
477 	/* == RPC_CWNDSCALE @ init, but *after* setup */
478 	if (r_xprt->rx_buf.rb_cwndscale == 0UL) {
479 		r_xprt->rx_buf.rb_cwndscale = xprt->cwnd;
480 		dprintk("RPC:       %s: cwndscale %lu\n", __func__,
481 			r_xprt->rx_buf.rb_cwndscale);
482 		BUG_ON(r_xprt->rx_buf.rb_cwndscale <= 0);
483 	}
484 	xprt->cwnd = credits * r_xprt->rx_buf.rb_cwndscale;
485 	return xprt_reserve_xprt_cong(task);
486 }
487 
488 /*
489  * The RDMA allocate/free functions need the task structure as a place
490  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
491  * sequence. For this reason, the recv buffers are attached to send
492  * buffers for portions of the RPC. Note that the RPC layer allocates
493  * both send and receive buffers in the same call. We may register
494  * the receive buffer portion when using reply chunks.
495  */
496 static void *
497 xprt_rdma_allocate(struct rpc_task *task, size_t size)
498 {
499 	struct rpc_xprt *xprt = task->tk_xprt;
500 	struct rpcrdma_req *req, *nreq;
501 
502 	req = rpcrdma_buffer_get(&rpcx_to_rdmax(xprt)->rx_buf);
503 	BUG_ON(NULL == req);
504 
505 	if (size > req->rl_size) {
506 		dprintk("RPC:       %s: size %zd too large for buffer[%zd]: "
507 			"prog %d vers %d proc %d\n",
508 			__func__, size, req->rl_size,
509 			task->tk_client->cl_prog, task->tk_client->cl_vers,
510 			task->tk_msg.rpc_proc->p_proc);
511 		/*
512 		 * Outgoing length shortage. Our inline write max must have
513 		 * been configured to perform direct i/o.
514 		 *
515 		 * This is therefore a large metadata operation, and the
516 		 * allocate call was made on the maximum possible message,
517 		 * e.g. containing long filename(s) or symlink data. In
518 		 * fact, while these metadata operations *might* carry
519 		 * large outgoing payloads, they rarely *do*. However, we
520 		 * have to commit to the request here, so reallocate and
521 		 * register it now. The data path will never require this
522 		 * reallocation.
523 		 *
524 		 * If the allocation or registration fails, the RPC framework
525 		 * will (doggedly) retry.
526 		 */
527 		if (rpcx_to_rdmax(xprt)->rx_ia.ri_memreg_strategy ==
528 				RPCRDMA_BOUNCEBUFFERS) {
529 			/* forced to "pure inline" */
530 			dprintk("RPC:       %s: too much data (%zd) for inline "
531 					"(r/w max %d/%d)\n", __func__, size,
532 					rpcx_to_rdmad(xprt).inline_rsize,
533 					rpcx_to_rdmad(xprt).inline_wsize);
534 			size = req->rl_size;
535 			rpc_exit(task, -EIO);		/* fail the operation */
536 			rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
537 			goto out;
538 		}
539 		if (task->tk_flags & RPC_TASK_SWAPPER)
540 			nreq = kmalloc(sizeof *req + size, GFP_ATOMIC);
541 		else
542 			nreq = kmalloc(sizeof *req + size, GFP_NOFS);
543 		if (nreq == NULL)
544 			goto outfail;
545 
546 		if (rpcrdma_register_internal(&rpcx_to_rdmax(xprt)->rx_ia,
547 				nreq->rl_base, size + sizeof(struct rpcrdma_req)
548 				- offsetof(struct rpcrdma_req, rl_base),
549 				&nreq->rl_handle, &nreq->rl_iov)) {
550 			kfree(nreq);
551 			goto outfail;
552 		}
553 		rpcx_to_rdmax(xprt)->rx_stats.hardway_register_count += size;
554 		nreq->rl_size = size;
555 		nreq->rl_niovs = 0;
556 		nreq->rl_nchunks = 0;
557 		nreq->rl_buffer = (struct rpcrdma_buffer *)req;
558 		nreq->rl_reply = req->rl_reply;
559 		memcpy(nreq->rl_segments,
560 			req->rl_segments, sizeof nreq->rl_segments);
561 		/* flag the swap with an unused field */
562 		nreq->rl_iov.length = 0;
563 		req->rl_reply = NULL;
564 		req = nreq;
565 	}
566 	dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
567 out:
568 	req->rl_connect_cookie = 0;	/* our reserved value */
569 	return req->rl_xdr_buf;
570 
571 outfail:
572 	rpcrdma_buffer_put(req);
573 	rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
574 	return NULL;
575 }
576 
577 /*
578  * This function returns all RDMA resources to the pool.
579  */
580 static void
581 xprt_rdma_free(void *buffer)
582 {
583 	struct rpcrdma_req *req;
584 	struct rpcrdma_xprt *r_xprt;
585 	struct rpcrdma_rep *rep;
586 	int i;
587 
588 	if (buffer == NULL)
589 		return;
590 
591 	req = container_of(buffer, struct rpcrdma_req, rl_xdr_buf[0]);
592 	if (req->rl_iov.length == 0) {	/* see allocate above */
593 		r_xprt = container_of(((struct rpcrdma_req *) req->rl_buffer)->rl_buffer,
594 				      struct rpcrdma_xprt, rx_buf);
595 	} else
596 		r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
597 	rep = req->rl_reply;
598 
599 	dprintk("RPC:       %s: called on 0x%p%s\n",
600 		__func__, rep, (rep && rep->rr_func) ? " (with waiter)" : "");
601 
602 	/*
603 	 * Finish the deregistration. When using mw bind, this was
604 	 * begun in rpcrdma_reply_handler(). In all other modes, we
605 	 * do it here, in thread context. The process is considered
606 	 * complete when the rr_func vector becomes NULL - this
607 	 * was put in place during rpcrdma_reply_handler() - the wait
608 	 * call below will not block if the dereg is "done". If
609 	 * interrupted, our framework will clean up.
610 	 */
611 	for (i = 0; req->rl_nchunks;) {
612 		--req->rl_nchunks;
613 		i += rpcrdma_deregister_external(
614 			&req->rl_segments[i], r_xprt, NULL);
615 	}
616 
617 	if (rep && wait_event_interruptible(rep->rr_unbind, !rep->rr_func)) {
618 		rep->rr_func = NULL;	/* abandon the callback */
619 		req->rl_reply = NULL;
620 	}
621 
622 	if (req->rl_iov.length == 0) {	/* see allocate above */
623 		struct rpcrdma_req *oreq = (struct rpcrdma_req *)req->rl_buffer;
624 		oreq->rl_reply = req->rl_reply;
625 		(void) rpcrdma_deregister_internal(&r_xprt->rx_ia,
626 						   req->rl_handle,
627 						   &req->rl_iov);
628 		kfree(req);
629 		req = oreq;
630 	}
631 
632 	/* Put back request+reply buffers */
633 	rpcrdma_buffer_put(req);
634 }
635 
636 /*
637  * send_request invokes the meat of RPC RDMA. It must do the following:
638  *  1.  Marshal the RPC request into an RPC RDMA request, which means
639  *	putting a header in front of data, and creating IOVs for RDMA
640  *	from those in the request.
641  *  2.  In marshaling, detect opportunities for RDMA, and use them.
642  *  3.  Post a recv message to set up asynch completion, then send
643  *	the request (rpcrdma_ep_post).
644  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
645  */
646 
647 static int
648 xprt_rdma_send_request(struct rpc_task *task)
649 {
650 	struct rpc_rqst *rqst = task->tk_rqstp;
651 	struct rpc_xprt *xprt = task->tk_xprt;
652 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
653 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
654 
655 	/* marshal the send itself */
656 	if (req->rl_niovs == 0 && rpcrdma_marshal_req(rqst) != 0) {
657 		r_xprt->rx_stats.failed_marshal_count++;
658 		dprintk("RPC:       %s: rpcrdma_marshal_req failed\n",
659 			__func__);
660 		return -EIO;
661 	}
662 
663 	if (req->rl_reply == NULL) 		/* e.g. reconnection */
664 		rpcrdma_recv_buffer_get(req);
665 
666 	if (req->rl_reply) {
667 		req->rl_reply->rr_func = rpcrdma_reply_handler;
668 		/* this need only be done once, but... */
669 		req->rl_reply->rr_xprt = xprt;
670 	}
671 
672 	/* Must suppress retransmit to maintain credits */
673 	if (req->rl_connect_cookie == xprt->connect_cookie)
674 		goto drop_connection;
675 	req->rl_connect_cookie = xprt->connect_cookie;
676 
677 	if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
678 		goto drop_connection;
679 
680 	task->tk_bytes_sent += rqst->rq_snd_buf.len;
681 	rqst->rq_bytes_sent = 0;
682 	return 0;
683 
684 drop_connection:
685 	xprt_disconnect_done(xprt);
686 	return -ENOTCONN;	/* implies disconnect */
687 }
688 
689 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
690 {
691 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
692 	long idle_time = 0;
693 
694 	if (xprt_connected(xprt))
695 		idle_time = (long)(jiffies - xprt->last_used) / HZ;
696 
697 	seq_printf(seq,
698 	  "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
699 	  "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
700 
701 	   0,	/* need a local port? */
702 	   xprt->stat.bind_count,
703 	   xprt->stat.connect_count,
704 	   xprt->stat.connect_time,
705 	   idle_time,
706 	   xprt->stat.sends,
707 	   xprt->stat.recvs,
708 	   xprt->stat.bad_xids,
709 	   xprt->stat.req_u,
710 	   xprt->stat.bklog_u,
711 
712 	   r_xprt->rx_stats.read_chunk_count,
713 	   r_xprt->rx_stats.write_chunk_count,
714 	   r_xprt->rx_stats.reply_chunk_count,
715 	   r_xprt->rx_stats.total_rdma_request,
716 	   r_xprt->rx_stats.total_rdma_reply,
717 	   r_xprt->rx_stats.pullup_copy_count,
718 	   r_xprt->rx_stats.fixup_copy_count,
719 	   r_xprt->rx_stats.hardway_register_count,
720 	   r_xprt->rx_stats.failed_marshal_count,
721 	   r_xprt->rx_stats.bad_reply_count);
722 }
723 
724 /*
725  * Plumbing for rpc transport switch and kernel module
726  */
727 
728 static struct rpc_xprt_ops xprt_rdma_procs = {
729 	.reserve_xprt		= xprt_rdma_reserve_xprt,
730 	.release_xprt		= xprt_release_xprt_cong, /* sunrpc/xprt.c */
731 	.release_request	= xprt_release_rqst_cong,       /* ditto */
732 	.set_retrans_timeout	= xprt_set_retrans_timeout_def, /* ditto */
733 	.rpcbind		= rpcb_getport_async,	/* sunrpc/rpcb_clnt.c */
734 	.set_port		= xprt_rdma_set_port,
735 	.connect		= xprt_rdma_connect,
736 	.buf_alloc		= xprt_rdma_allocate,
737 	.buf_free		= xprt_rdma_free,
738 	.send_request		= xprt_rdma_send_request,
739 	.close			= xprt_rdma_close,
740 	.destroy		= xprt_rdma_destroy,
741 	.print_stats		= xprt_rdma_print_stats
742 };
743 
744 static struct xprt_class xprt_rdma = {
745 	.list			= LIST_HEAD_INIT(xprt_rdma.list),
746 	.name			= "rdma",
747 	.owner			= THIS_MODULE,
748 	.ident			= XPRT_TRANSPORT_RDMA,
749 	.setup			= xprt_setup_rdma,
750 };
751 
752 static void __exit xprt_rdma_cleanup(void)
753 {
754 	int rc;
755 
756 	dprintk(KERN_INFO "RPCRDMA Module Removed, deregister RPC RDMA transport\n");
757 #ifdef RPC_DEBUG
758 	if (sunrpc_table_header) {
759 		unregister_sysctl_table(sunrpc_table_header);
760 		sunrpc_table_header = NULL;
761 	}
762 #endif
763 	rc = xprt_unregister_transport(&xprt_rdma);
764 	if (rc)
765 		dprintk("RPC:       %s: xprt_unregister returned %i\n",
766 			__func__, rc);
767 }
768 
769 static int __init xprt_rdma_init(void)
770 {
771 	int rc;
772 
773 	rc = xprt_register_transport(&xprt_rdma);
774 
775 	if (rc)
776 		return rc;
777 
778 	dprintk(KERN_INFO "RPCRDMA Module Init, register RPC RDMA transport\n");
779 
780 	dprintk(KERN_INFO "Defaults:\n");
781 	dprintk(KERN_INFO "\tSlots %d\n"
782 		"\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
783 		xprt_rdma_slot_table_entries,
784 		xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
785 	dprintk(KERN_INFO "\tPadding %d\n\tMemreg %d\n",
786 		xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
787 
788 #ifdef RPC_DEBUG
789 	if (!sunrpc_table_header)
790 		sunrpc_table_header = register_sysctl_table(sunrpc_table);
791 #endif
792 	return 0;
793 }
794 
795 module_init(xprt_rdma_init);
796 module_exit(xprt_rdma_cleanup);
797