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
62MODULE_LICENSE("Dual BSD/GPL");
63
64MODULE_DESCRIPTION("RPC/RDMA Transport for Linux kernel NFS");
65MODULE_AUTHOR("Network Appliance, Inc.");
66
67/*
68 * tunables
69 */
70
71static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
72static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
73static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
74static unsigned int xprt_rdma_inline_write_padding;
75static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
76		int xprt_rdma_pad_optimize = 1;
77
78#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
79
80static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
81static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
82static unsigned int zero;
83static unsigned int max_padding = PAGE_SIZE;
84static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
85static unsigned int max_memreg = RPCRDMA_LAST - 1;
86
87static struct ctl_table_header *sunrpc_table_header;
88
89static 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
141static 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
157static struct rpc_xprt_ops xprt_rdma_procs;	/* forward reference */
158
159static void
160xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
161{
162	struct sockaddr_in *sin = (struct sockaddr_in *)sap;
163	char buf[20];
164
165	snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
166	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
167
168	xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
169}
170
171static void
172xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
173{
174	struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
175	char buf[40];
176
177	snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
178	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
179
180	xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
181}
182
183static void
184xprt_rdma_format_addresses(struct rpc_xprt *xprt)
185{
186	struct sockaddr *sap = (struct sockaddr *)
187					&rpcx_to_rdmad(xprt).addr;
188	char buf[128];
189
190	switch (sap->sa_family) {
191	case AF_INET:
192		xprt_rdma_format_addresses4(xprt, sap);
193		break;
194	case AF_INET6:
195		xprt_rdma_format_addresses6(xprt, sap);
196		break;
197	default:
198		pr_err("rpcrdma: Unrecognized address family\n");
199		return;
200	}
201
202	(void)rpc_ntop(sap, buf, sizeof(buf));
203	xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
204
205	snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
206	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
207
208	snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
209	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
210
211	xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
212}
213
214static void
215xprt_rdma_free_addresses(struct rpc_xprt *xprt)
216{
217	unsigned int i;
218
219	for (i = 0; i < RPC_DISPLAY_MAX; i++)
220		switch (i) {
221		case RPC_DISPLAY_PROTO:
222		case RPC_DISPLAY_NETID:
223			continue;
224		default:
225			kfree(xprt->address_strings[i]);
226		}
227}
228
229static void
230xprt_rdma_connect_worker(struct work_struct *work)
231{
232	struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
233						   rx_connect_worker.work);
234	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
235	int rc = 0;
236
237	xprt_clear_connected(xprt);
238
239	dprintk("RPC:       %s: %sconnect\n", __func__,
240			r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
241	rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
242	if (rc)
243		xprt_wake_pending_tasks(xprt, rc);
244
245	dprintk("RPC:       %s: exit\n", __func__);
246	xprt_clear_connecting(xprt);
247}
248
249/*
250 * xprt_rdma_destroy
251 *
252 * Destroy the xprt.
253 * Free all memory associated with the object, including its own.
254 * NOTE: none of the *destroy methods free memory for their top-level
255 * objects, even though they may have allocated it (they do free
256 * private memory). It's up to the caller to handle it. In this
257 * case (RDMA transport), all structure memory is inlined with the
258 * struct rpcrdma_xprt.
259 */
260static void
261xprt_rdma_destroy(struct rpc_xprt *xprt)
262{
263	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
264
265	dprintk("RPC:       %s: called\n", __func__);
266
267	cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
268
269	xprt_clear_connected(xprt);
270
271	rpcrdma_buffer_destroy(&r_xprt->rx_buf);
272	rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
273	rpcrdma_ia_close(&r_xprt->rx_ia);
274
275	xprt_rdma_free_addresses(xprt);
276
277	xprt_free(xprt);
278
279	dprintk("RPC:       %s: returning\n", __func__);
280
281	module_put(THIS_MODULE);
282}
283
284static const struct rpc_timeout xprt_rdma_default_timeout = {
285	.to_initval = 60 * HZ,
286	.to_maxval = 60 * HZ,
287};
288
289/**
290 * xprt_setup_rdma - Set up transport to use RDMA
291 *
292 * @args: rpc transport arguments
293 */
294static struct rpc_xprt *
295xprt_setup_rdma(struct xprt_create *args)
296{
297	struct rpcrdma_create_data_internal cdata;
298	struct rpc_xprt *xprt;
299	struct rpcrdma_xprt *new_xprt;
300	struct rpcrdma_ep *new_ep;
301	struct sockaddr_in *sin;
302	int rc;
303
304	if (args->addrlen > sizeof(xprt->addr)) {
305		dprintk("RPC:       %s: address too large\n", __func__);
306		return ERR_PTR(-EBADF);
307	}
308
309	xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
310			xprt_rdma_slot_table_entries,
311			xprt_rdma_slot_table_entries);
312	if (xprt == NULL) {
313		dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
314			__func__);
315		return ERR_PTR(-ENOMEM);
316	}
317
318	/* 60 second timeout, no retries */
319	xprt->timeout = &xprt_rdma_default_timeout;
320	xprt->bind_timeout = RPCRDMA_BIND_TO;
321	xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
322	xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
323
324	xprt->resvport = 0;		/* privileged port not needed */
325	xprt->tsh_size = 0;		/* RPC-RDMA handles framing */
326	xprt->ops = &xprt_rdma_procs;
327
328	/*
329	 * Set up RDMA-specific connect data.
330	 */
331
332	/* Put server RDMA address in local cdata */
333	memcpy(&cdata.addr, args->dstaddr, args->addrlen);
334
335	/* Ensure xprt->addr holds valid server TCP (not RDMA)
336	 * address, for any side protocols which peek at it */
337	xprt->prot = IPPROTO_TCP;
338	xprt->addrlen = args->addrlen;
339	memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
340
341	sin = (struct sockaddr_in *)&cdata.addr;
342	if (ntohs(sin->sin_port) != 0)
343		xprt_set_bound(xprt);
344
345	dprintk("RPC:       %s: %pI4:%u\n",
346		__func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port));
347
348	/* Set max requests */
349	cdata.max_requests = xprt->max_reqs;
350
351	/* Set some length limits */
352	cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
353	cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
354
355	cdata.inline_wsize = xprt_rdma_max_inline_write;
356	if (cdata.inline_wsize > cdata.wsize)
357		cdata.inline_wsize = cdata.wsize;
358
359	cdata.inline_rsize = xprt_rdma_max_inline_read;
360	if (cdata.inline_rsize > cdata.rsize)
361		cdata.inline_rsize = cdata.rsize;
362
363	cdata.padding = xprt_rdma_inline_write_padding;
364
365	/*
366	 * Create new transport instance, which includes initialized
367	 *  o ia
368	 *  o endpoint
369	 *  o buffers
370	 */
371
372	new_xprt = rpcx_to_rdmax(xprt);
373
374	rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
375				xprt_rdma_memreg_strategy);
376	if (rc)
377		goto out1;
378
379	/*
380	 * initialize and create ep
381	 */
382	new_xprt->rx_data = cdata;
383	new_ep = &new_xprt->rx_ep;
384	new_ep->rep_remote_addr = cdata.addr;
385
386	rc = rpcrdma_ep_create(&new_xprt->rx_ep,
387				&new_xprt->rx_ia, &new_xprt->rx_data);
388	if (rc)
389		goto out2;
390
391	/*
392	 * Allocate pre-registered send and receive buffers for headers and
393	 * any inline data. Also specify any padding which will be provided
394	 * from a preregistered zero buffer.
395	 */
396	rc = rpcrdma_buffer_create(new_xprt);
397	if (rc)
398		goto out3;
399
400	/*
401	 * Register a callback for connection events. This is necessary because
402	 * connection loss notification is async. We also catch connection loss
403	 * when reaping receives.
404	 */
405	INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
406			  xprt_rdma_connect_worker);
407
408	xprt_rdma_format_addresses(xprt);
409	xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
410	if (xprt->max_payload == 0)
411		goto out4;
412	xprt->max_payload <<= PAGE_SHIFT;
413	dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
414		__func__, xprt->max_payload);
415
416	if (!try_module_get(THIS_MODULE))
417		goto out4;
418
419	return xprt;
420
421out4:
422	xprt_rdma_free_addresses(xprt);
423	rc = -EINVAL;
424out3:
425	rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
426out2:
427	rpcrdma_ia_close(&new_xprt->rx_ia);
428out1:
429	xprt_free(xprt);
430	return ERR_PTR(rc);
431}
432
433/*
434 * Close a connection, during shutdown or timeout/reconnect
435 */
436static void
437xprt_rdma_close(struct rpc_xprt *xprt)
438{
439	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
440
441	dprintk("RPC:       %s: closing\n", __func__);
442	if (r_xprt->rx_ep.rep_connected > 0)
443		xprt->reestablish_timeout = 0;
444	xprt_disconnect_done(xprt);
445	rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
446}
447
448static void
449xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
450{
451	struct sockaddr_in *sap;
452
453	sap = (struct sockaddr_in *)&xprt->addr;
454	sap->sin_port = htons(port);
455	sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
456	sap->sin_port = htons(port);
457	dprintk("RPC:       %s: %u\n", __func__, port);
458}
459
460static void
461xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
462{
463	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
464
465	if (r_xprt->rx_ep.rep_connected != 0) {
466		/* Reconnect */
467		schedule_delayed_work(&r_xprt->rx_connect_worker,
468				      xprt->reestablish_timeout);
469		xprt->reestablish_timeout <<= 1;
470		if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
471			xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
472		else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
473			xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
474	} else {
475		schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
476		if (!RPC_IS_ASYNC(task))
477			flush_delayed_work(&r_xprt->rx_connect_worker);
478	}
479}
480
481/*
482 * The RDMA allocate/free functions need the task structure as a place
483 * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
484 * sequence.
485 *
486 * The RPC layer allocates both send and receive buffers in the same call
487 * (rq_send_buf and rq_rcv_buf are both part of a single contiguous buffer).
488 * We may register rq_rcv_buf when using reply chunks.
489 */
490static void *
491xprt_rdma_allocate(struct rpc_task *task, size_t size)
492{
493	struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
494	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
495	struct rpcrdma_regbuf *rb;
496	struct rpcrdma_req *req;
497	size_t min_size;
498	gfp_t flags;
499
500	req = rpcrdma_buffer_get(&r_xprt->rx_buf);
501	if (req == NULL)
502		return NULL;
503
504	flags = GFP_NOIO | __GFP_NOWARN;
505	if (RPC_IS_SWAPPER(task))
506		flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
507
508	if (req->rl_rdmabuf == NULL)
509		goto out_rdmabuf;
510	if (req->rl_sendbuf == NULL)
511		goto out_sendbuf;
512	if (size > req->rl_sendbuf->rg_size)
513		goto out_sendbuf;
514
515out:
516	dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
517	req->rl_connect_cookie = 0;	/* our reserved value */
518	return req->rl_sendbuf->rg_base;
519
520out_rdmabuf:
521	min_size = RPCRDMA_INLINE_WRITE_THRESHOLD(task->tk_rqstp);
522	rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, min_size, flags);
523	if (IS_ERR(rb))
524		goto out_fail;
525	req->rl_rdmabuf = rb;
526
527out_sendbuf:
528	/* XDR encoding and RPC/RDMA marshaling of this request has not
529	 * yet occurred. Thus a lower bound is needed to prevent buffer
530	 * overrun during marshaling.
531	 *
532	 * RPC/RDMA marshaling may choose to send payload bearing ops
533	 * inline, if the result is smaller than the inline threshold.
534	 * The value of the "size" argument accounts for header
535	 * requirements but not for the payload in these cases.
536	 *
537	 * Likewise, allocate enough space to receive a reply up to the
538	 * size of the inline threshold.
539	 *
540	 * It's unlikely that both the send header and the received
541	 * reply will be large, but slush is provided here to allow
542	 * flexibility when marshaling.
543	 */
544	min_size = RPCRDMA_INLINE_READ_THRESHOLD(task->tk_rqstp);
545	min_size += RPCRDMA_INLINE_WRITE_THRESHOLD(task->tk_rqstp);
546	if (size < min_size)
547		size = min_size;
548
549	rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, size, flags);
550	if (IS_ERR(rb))
551		goto out_fail;
552	rb->rg_owner = req;
553
554	r_xprt->rx_stats.hardway_register_count += size;
555	rpcrdma_free_regbuf(&r_xprt->rx_ia, req->rl_sendbuf);
556	req->rl_sendbuf = rb;
557	goto out;
558
559out_fail:
560	rpcrdma_buffer_put(req);
561	r_xprt->rx_stats.failed_marshal_count++;
562	return NULL;
563}
564
565/*
566 * This function returns all RDMA resources to the pool.
567 */
568static void
569xprt_rdma_free(void *buffer)
570{
571	struct rpcrdma_req *req;
572	struct rpcrdma_xprt *r_xprt;
573	struct rpcrdma_regbuf *rb;
574	int i;
575
576	if (buffer == NULL)
577		return;
578
579	rb = container_of(buffer, struct rpcrdma_regbuf, rg_base[0]);
580	req = rb->rg_owner;
581	r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
582
583	dprintk("RPC:       %s: called on 0x%p\n", __func__, req->rl_reply);
584
585	for (i = 0; req->rl_nchunks;) {
586		--req->rl_nchunks;
587		i += r_xprt->rx_ia.ri_ops->ro_unmap(r_xprt,
588						    &req->rl_segments[i]);
589	}
590
591	rpcrdma_buffer_put(req);
592}
593
594/*
595 * send_request invokes the meat of RPC RDMA. It must do the following:
596 *  1.  Marshal the RPC request into an RPC RDMA request, which means
597 *	putting a header in front of data, and creating IOVs for RDMA
598 *	from those in the request.
599 *  2.  In marshaling, detect opportunities for RDMA, and use them.
600 *  3.  Post a recv message to set up asynch completion, then send
601 *	the request (rpcrdma_ep_post).
602 *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
603 */
604
605static int
606xprt_rdma_send_request(struct rpc_task *task)
607{
608	struct rpc_rqst *rqst = task->tk_rqstp;
609	struct rpc_xprt *xprt = rqst->rq_xprt;
610	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
611	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
612	int rc = 0;
613
614	rc = rpcrdma_marshal_req(rqst);
615	if (rc < 0)
616		goto failed_marshal;
617
618	if (req->rl_reply == NULL) 		/* e.g. reconnection */
619		rpcrdma_recv_buffer_get(req);
620
621	if (req->rl_reply) {
622		req->rl_reply->rr_func = rpcrdma_reply_handler;
623		/* this need only be done once, but... */
624		req->rl_reply->rr_xprt = xprt;
625	}
626
627	/* Must suppress retransmit to maintain credits */
628	if (req->rl_connect_cookie == xprt->connect_cookie)
629		goto drop_connection;
630	req->rl_connect_cookie = xprt->connect_cookie;
631
632	if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
633		goto drop_connection;
634
635	rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
636	rqst->rq_bytes_sent = 0;
637	return 0;
638
639failed_marshal:
640	r_xprt->rx_stats.failed_marshal_count++;
641	dprintk("RPC:       %s: rpcrdma_marshal_req failed, status %i\n",
642		__func__, rc);
643	if (rc == -EIO)
644		return -EIO;
645drop_connection:
646	xprt_disconnect_done(xprt);
647	return -ENOTCONN;	/* implies disconnect */
648}
649
650static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
651{
652	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
653	long idle_time = 0;
654
655	if (xprt_connected(xprt))
656		idle_time = (long)(jiffies - xprt->last_used) / HZ;
657
658	seq_printf(seq,
659	  "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
660	  "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
661
662	   0,	/* need a local port? */
663	   xprt->stat.bind_count,
664	   xprt->stat.connect_count,
665	   xprt->stat.connect_time,
666	   idle_time,
667	   xprt->stat.sends,
668	   xprt->stat.recvs,
669	   xprt->stat.bad_xids,
670	   xprt->stat.req_u,
671	   xprt->stat.bklog_u,
672
673	   r_xprt->rx_stats.read_chunk_count,
674	   r_xprt->rx_stats.write_chunk_count,
675	   r_xprt->rx_stats.reply_chunk_count,
676	   r_xprt->rx_stats.total_rdma_request,
677	   r_xprt->rx_stats.total_rdma_reply,
678	   r_xprt->rx_stats.pullup_copy_count,
679	   r_xprt->rx_stats.fixup_copy_count,
680	   r_xprt->rx_stats.hardway_register_count,
681	   r_xprt->rx_stats.failed_marshal_count,
682	   r_xprt->rx_stats.bad_reply_count);
683}
684
685/*
686 * Plumbing for rpc transport switch and kernel module
687 */
688
689static struct rpc_xprt_ops xprt_rdma_procs = {
690	.reserve_xprt		= xprt_reserve_xprt_cong,
691	.release_xprt		= xprt_release_xprt_cong, /* sunrpc/xprt.c */
692	.alloc_slot		= xprt_alloc_slot,
693	.release_request	= xprt_release_rqst_cong,       /* ditto */
694	.set_retrans_timeout	= xprt_set_retrans_timeout_def, /* ditto */
695	.rpcbind		= rpcb_getport_async,	/* sunrpc/rpcb_clnt.c */
696	.set_port		= xprt_rdma_set_port,
697	.connect		= xprt_rdma_connect,
698	.buf_alloc		= xprt_rdma_allocate,
699	.buf_free		= xprt_rdma_free,
700	.send_request		= xprt_rdma_send_request,
701	.close			= xprt_rdma_close,
702	.destroy		= xprt_rdma_destroy,
703	.print_stats		= xprt_rdma_print_stats
704};
705
706static struct xprt_class xprt_rdma = {
707	.list			= LIST_HEAD_INIT(xprt_rdma.list),
708	.name			= "rdma",
709	.owner			= THIS_MODULE,
710	.ident			= XPRT_TRANSPORT_RDMA,
711	.setup			= xprt_setup_rdma,
712};
713
714static void __exit xprt_rdma_cleanup(void)
715{
716	int rc;
717
718	dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
719#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
720	if (sunrpc_table_header) {
721		unregister_sysctl_table(sunrpc_table_header);
722		sunrpc_table_header = NULL;
723	}
724#endif
725	rc = xprt_unregister_transport(&xprt_rdma);
726	if (rc)
727		dprintk("RPC:       %s: xprt_unregister returned %i\n",
728			__func__, rc);
729}
730
731static int __init xprt_rdma_init(void)
732{
733	int rc;
734
735	rc = xprt_register_transport(&xprt_rdma);
736
737	if (rc)
738		return rc;
739
740	dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
741
742	dprintk("Defaults:\n");
743	dprintk("\tSlots %d\n"
744		"\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
745		xprt_rdma_slot_table_entries,
746		xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
747	dprintk("\tPadding %d\n\tMemreg %d\n",
748		xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
749
750#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
751	if (!sunrpc_table_header)
752		sunrpc_table_header = register_sysctl_table(sunrpc_table);
753#endif
754	return 0;
755}
756
757module_init(xprt_rdma_init);
758module_exit(xprt_rdma_cleanup);
759