1// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2/*
3 * Copyright (c) 2014-2017 Oracle.  All rights reserved.
4 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
5 *
6 * This software is available to you under a choice of one of two
7 * licenses.  You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the BSD-type
10 * license below:
11 *
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
14 * are met:
15 *
16 *      Redistributions of source code must retain the above copyright
17 *      notice, this list of conditions and the following disclaimer.
18 *
19 *      Redistributions in binary form must reproduce the above
20 *      copyright notice, this list of conditions and the following
21 *      disclaimer in the documentation and/or other materials provided
22 *      with the distribution.
23 *
24 *      Neither the name of the Network Appliance, Inc. nor the names of
25 *      its contributors may be used to endorse or promote products
26 *      derived from this software without specific prior written
27 *      permission.
28 *
29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40 */
41
42/*
43 * transport.c
44 *
45 * This file contains the top-level implementation of an RPC RDMA
46 * transport.
47 *
48 * Naming convention: functions beginning with xprt_ are part of the
49 * transport switch. All others are RPC RDMA internal.
50 */
51
52#include <linux/module.h>
53#include <linux/slab.h>
54#include <linux/seq_file.h>
55#include <linux/smp.h>
56
57#include <linux/sunrpc/addr.h>
58#include <linux/sunrpc/svc_rdma.h>
59
60#include "xprt_rdma.h"
61#include <trace/events/rpcrdma.h>
62
63#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
64# define RPCDBG_FACILITY	RPCDBG_TRANS
65#endif
66
67/*
68 * tunables
69 */
70
71static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
72unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
73unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
74unsigned int xprt_rdma_memreg_strategy		= RPCRDMA_FRWR;
75int xprt_rdma_pad_optimize;
76
77#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
78
79static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
80static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
81static unsigned int min_inline_size = RPCRDMA_MIN_INLINE;
82static unsigned int max_inline_size = RPCRDMA_MAX_INLINE;
83static unsigned int max_padding = PAGE_SIZE;
84static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
85static unsigned int max_memreg = RPCRDMA_LAST - 1;
86static unsigned int dummy;
87
88static struct ctl_table_header *sunrpc_table_header;
89
90static struct ctl_table xr_tunables_table[] = {
91	{
92		.procname	= "rdma_slot_table_entries",
93		.data		= &xprt_rdma_slot_table_entries,
94		.maxlen		= sizeof(unsigned int),
95		.mode		= 0644,
96		.proc_handler	= proc_dointvec_minmax,
97		.extra1		= &min_slot_table_size,
98		.extra2		= &max_slot_table_size
99	},
100	{
101		.procname	= "rdma_max_inline_read",
102		.data		= &xprt_rdma_max_inline_read,
103		.maxlen		= sizeof(unsigned int),
104		.mode		= 0644,
105		.proc_handler	= proc_dointvec_minmax,
106		.extra1		= &min_inline_size,
107		.extra2		= &max_inline_size,
108	},
109	{
110		.procname	= "rdma_max_inline_write",
111		.data		= &xprt_rdma_max_inline_write,
112		.maxlen		= sizeof(unsigned int),
113		.mode		= 0644,
114		.proc_handler	= proc_dointvec_minmax,
115		.extra1		= &min_inline_size,
116		.extra2		= &max_inline_size,
117	},
118	{
119		.procname	= "rdma_inline_write_padding",
120		.data		= &dummy,
121		.maxlen		= sizeof(unsigned int),
122		.mode		= 0644,
123		.proc_handler	= proc_dointvec_minmax,
124		.extra1		= SYSCTL_ZERO,
125		.extra2		= &max_padding,
126	},
127	{
128		.procname	= "rdma_memreg_strategy",
129		.data		= &xprt_rdma_memreg_strategy,
130		.maxlen		= sizeof(unsigned int),
131		.mode		= 0644,
132		.proc_handler	= proc_dointvec_minmax,
133		.extra1		= &min_memreg,
134		.extra2		= &max_memreg,
135	},
136	{
137		.procname	= "rdma_pad_optimize",
138		.data		= &xprt_rdma_pad_optimize,
139		.maxlen		= sizeof(unsigned int),
140		.mode		= 0644,
141		.proc_handler	= proc_dointvec,
142	},
143	{ },
144};
145
146static struct ctl_table sunrpc_table[] = {
147	{
148		.procname	= "sunrpc",
149		.mode		= 0555,
150		.child		= xr_tunables_table
151	},
152	{ },
153};
154
155#endif
156
157static const struct rpc_xprt_ops xprt_rdma_procs;
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
183void
184xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
185{
186	char buf[128];
187
188	switch (sap->sa_family) {
189	case AF_INET:
190		xprt_rdma_format_addresses4(xprt, sap);
191		break;
192	case AF_INET6:
193		xprt_rdma_format_addresses6(xprt, sap);
194		break;
195	default:
196		pr_err("rpcrdma: Unrecognized address family\n");
197		return;
198	}
199
200	(void)rpc_ntop(sap, buf, sizeof(buf));
201	xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
202
203	snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
204	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
205
206	snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
207	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
208
209	xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
210}
211
212void
213xprt_rdma_free_addresses(struct rpc_xprt *xprt)
214{
215	unsigned int i;
216
217	for (i = 0; i < RPC_DISPLAY_MAX; i++)
218		switch (i) {
219		case RPC_DISPLAY_PROTO:
220		case RPC_DISPLAY_NETID:
221			continue;
222		default:
223			kfree(xprt->address_strings[i]);
224		}
225}
226
227/**
228 * xprt_rdma_connect_worker - establish connection in the background
229 * @work: worker thread context
230 *
231 * Requester holds the xprt's send lock to prevent activity on this
232 * transport while a fresh connection is being established. RPC tasks
233 * sleep on the xprt's pending queue waiting for connect to complete.
234 */
235static void
236xprt_rdma_connect_worker(struct work_struct *work)
237{
238	struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
239						   rx_connect_worker.work);
240	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
241	int rc;
242
243	rc = rpcrdma_xprt_connect(r_xprt);
244	xprt_clear_connecting(xprt);
245	if (!rc) {
246		xprt->connect_cookie++;
247		xprt->stat.connect_count++;
248		xprt->stat.connect_time += (long)jiffies -
249					   xprt->stat.connect_start;
250		xprt_set_connected(xprt);
251		rc = -EAGAIN;
252	} else
253		rpcrdma_xprt_disconnect(r_xprt);
254	xprt_unlock_connect(xprt, r_xprt);
255	xprt_wake_pending_tasks(xprt, rc);
256}
257
258/**
259 * xprt_rdma_inject_disconnect - inject a connection fault
260 * @xprt: transport context
261 *
262 * If @xprt is connected, disconnect it to simulate spurious
263 * connection loss. Caller must hold @xprt's send lock to
264 * ensure that data structures and hardware resources are
265 * stable during the rdma_disconnect() call.
266 */
267static void
268xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
269{
270	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
271
272	trace_xprtrdma_op_inject_dsc(r_xprt);
273	rdma_disconnect(r_xprt->rx_ep->re_id);
274}
275
276/**
277 * xprt_rdma_destroy - Full tear down of transport
278 * @xprt: doomed transport context
279 *
280 * Caller guarantees there will be no more calls to us with
281 * this @xprt.
282 */
283static void
284xprt_rdma_destroy(struct rpc_xprt *xprt)
285{
286	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
287
288	cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
289
290	rpcrdma_xprt_disconnect(r_xprt);
291	rpcrdma_buffer_destroy(&r_xprt->rx_buf);
292
293	xprt_rdma_free_addresses(xprt);
294	xprt_free(xprt);
295
296	module_put(THIS_MODULE);
297}
298
299/* 60 second timeout, no retries */
300static const struct rpc_timeout xprt_rdma_default_timeout = {
301	.to_initval = 60 * HZ,
302	.to_maxval = 60 * HZ,
303};
304
305/**
306 * xprt_setup_rdma - Set up transport to use RDMA
307 *
308 * @args: rpc transport arguments
309 */
310static struct rpc_xprt *
311xprt_setup_rdma(struct xprt_create *args)
312{
313	struct rpc_xprt *xprt;
314	struct rpcrdma_xprt *new_xprt;
315	struct sockaddr *sap;
316	int rc;
317
318	if (args->addrlen > sizeof(xprt->addr))
319		return ERR_PTR(-EBADF);
320
321	if (!try_module_get(THIS_MODULE))
322		return ERR_PTR(-EIO);
323
324	xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt), 0,
325			  xprt_rdma_slot_table_entries);
326	if (!xprt) {
327		module_put(THIS_MODULE);
328		return ERR_PTR(-ENOMEM);
329	}
330
331	xprt->timeout = &xprt_rdma_default_timeout;
332	xprt->connect_timeout = xprt->timeout->to_initval;
333	xprt->max_reconnect_timeout = xprt->timeout->to_maxval;
334	xprt->bind_timeout = RPCRDMA_BIND_TO;
335	xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
336	xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
337
338	xprt->resvport = 0;		/* privileged port not needed */
339	xprt->ops = &xprt_rdma_procs;
340
341	/*
342	 * Set up RDMA-specific connect data.
343	 */
344	sap = args->dstaddr;
345
346	/* Ensure xprt->addr holds valid server TCP (not RDMA)
347	 * address, for any side protocols which peek at it */
348	xprt->prot = IPPROTO_TCP;
349	xprt->addrlen = args->addrlen;
350	memcpy(&xprt->addr, sap, xprt->addrlen);
351
352	if (rpc_get_port(sap))
353		xprt_set_bound(xprt);
354	xprt_rdma_format_addresses(xprt, sap);
355
356	new_xprt = rpcx_to_rdmax(xprt);
357	rc = rpcrdma_buffer_create(new_xprt);
358	if (rc) {
359		xprt_rdma_free_addresses(xprt);
360		xprt_free(xprt);
361		module_put(THIS_MODULE);
362		return ERR_PTR(rc);
363	}
364
365	INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
366			  xprt_rdma_connect_worker);
367
368	xprt->max_payload = RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT;
369
370	return xprt;
371}
372
373/**
374 * xprt_rdma_close - close a transport connection
375 * @xprt: transport context
376 *
377 * Called during autoclose or device removal.
378 *
379 * Caller holds @xprt's send lock to prevent activity on this
380 * transport while the connection is torn down.
381 */
382void xprt_rdma_close(struct rpc_xprt *xprt)
383{
384	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
385
386	rpcrdma_xprt_disconnect(r_xprt);
387
388	xprt->reestablish_timeout = 0;
389	++xprt->connect_cookie;
390	xprt_disconnect_done(xprt);
391}
392
393/**
394 * xprt_rdma_set_port - update server port with rpcbind result
395 * @xprt: controlling RPC transport
396 * @port: new port value
397 *
398 * Transport connect status is unchanged.
399 */
400static void
401xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
402{
403	struct sockaddr *sap = (struct sockaddr *)&xprt->addr;
404	char buf[8];
405
406	rpc_set_port(sap, port);
407
408	kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
409	snprintf(buf, sizeof(buf), "%u", port);
410	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
411
412	kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]);
413	snprintf(buf, sizeof(buf), "%4hx", port);
414	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
415}
416
417/**
418 * xprt_rdma_timer - invoked when an RPC times out
419 * @xprt: controlling RPC transport
420 * @task: RPC task that timed out
421 *
422 * Invoked when the transport is still connected, but an RPC
423 * retransmit timeout occurs.
424 *
425 * Since RDMA connections don't have a keep-alive, forcibly
426 * disconnect and retry to connect. This drives full
427 * detection of the network path, and retransmissions of
428 * all pending RPCs.
429 */
430static void
431xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
432{
433	xprt_force_disconnect(xprt);
434}
435
436/**
437 * xprt_rdma_set_connect_timeout - set timeouts for establishing a connection
438 * @xprt: controlling transport instance
439 * @connect_timeout: reconnect timeout after client disconnects
440 * @reconnect_timeout: reconnect timeout after server disconnects
441 *
442 */
443static void xprt_rdma_set_connect_timeout(struct rpc_xprt *xprt,
444					  unsigned long connect_timeout,
445					  unsigned long reconnect_timeout)
446{
447	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
448
449	trace_xprtrdma_op_set_cto(r_xprt, connect_timeout, reconnect_timeout);
450
451	spin_lock(&xprt->transport_lock);
452
453	if (connect_timeout < xprt->connect_timeout) {
454		struct rpc_timeout to;
455		unsigned long initval;
456
457		to = *xprt->timeout;
458		initval = connect_timeout;
459		if (initval < RPCRDMA_INIT_REEST_TO << 1)
460			initval = RPCRDMA_INIT_REEST_TO << 1;
461		to.to_initval = initval;
462		to.to_maxval = initval;
463		r_xprt->rx_timeout = to;
464		xprt->timeout = &r_xprt->rx_timeout;
465		xprt->connect_timeout = connect_timeout;
466	}
467
468	if (reconnect_timeout < xprt->max_reconnect_timeout)
469		xprt->max_reconnect_timeout = reconnect_timeout;
470
471	spin_unlock(&xprt->transport_lock);
472}
473
474/**
475 * xprt_rdma_connect - schedule an attempt to reconnect
476 * @xprt: transport state
477 * @task: RPC scheduler context (unused)
478 *
479 */
480static void
481xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
482{
483	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
484	struct rpcrdma_ep *ep = r_xprt->rx_ep;
485	unsigned long delay;
486
487	WARN_ON_ONCE(!xprt_lock_connect(xprt, task, r_xprt));
488
489	delay = 0;
490	if (ep && ep->re_connect_status != 0) {
491		delay = xprt_reconnect_delay(xprt);
492		xprt_reconnect_backoff(xprt, RPCRDMA_INIT_REEST_TO);
493	}
494	trace_xprtrdma_op_connect(r_xprt, delay);
495	queue_delayed_work(xprtiod_workqueue, &r_xprt->rx_connect_worker,
496			   delay);
497}
498
499/**
500 * xprt_rdma_alloc_slot - allocate an rpc_rqst
501 * @xprt: controlling RPC transport
502 * @task: RPC task requesting a fresh rpc_rqst
503 *
504 * tk_status values:
505 *	%0 if task->tk_rqstp points to a fresh rpc_rqst
506 *	%-EAGAIN if no rpc_rqst is available; queued on backlog
507 */
508static void
509xprt_rdma_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
510{
511	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
512	struct rpcrdma_req *req;
513
514	req = rpcrdma_buffer_get(&r_xprt->rx_buf);
515	if (!req)
516		goto out_sleep;
517	task->tk_rqstp = &req->rl_slot;
518	task->tk_status = 0;
519	return;
520
521out_sleep:
522	task->tk_status = -ENOMEM;
523	xprt_add_backlog(xprt, task);
524}
525
526/**
527 * xprt_rdma_free_slot - release an rpc_rqst
528 * @xprt: controlling RPC transport
529 * @rqst: rpc_rqst to release
530 *
531 */
532static void
533xprt_rdma_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *rqst)
534{
535	struct rpcrdma_xprt *r_xprt =
536		container_of(xprt, struct rpcrdma_xprt, rx_xprt);
537
538	rpcrdma_reply_put(&r_xprt->rx_buf, rpcr_to_rdmar(rqst));
539	if (!xprt_wake_up_backlog(xprt, rqst)) {
540		memset(rqst, 0, sizeof(*rqst));
541		rpcrdma_buffer_put(&r_xprt->rx_buf, rpcr_to_rdmar(rqst));
542	}
543}
544
545static bool rpcrdma_check_regbuf(struct rpcrdma_xprt *r_xprt,
546				 struct rpcrdma_regbuf *rb, size_t size,
547				 gfp_t flags)
548{
549	if (unlikely(rdmab_length(rb) < size)) {
550		if (!rpcrdma_regbuf_realloc(rb, size, flags))
551			return false;
552		r_xprt->rx_stats.hardway_register_count += size;
553	}
554	return true;
555}
556
557/**
558 * xprt_rdma_allocate - allocate transport resources for an RPC
559 * @task: RPC task
560 *
561 * Return values:
562 *        0:	Success; rq_buffer points to RPC buffer to use
563 *   ENOMEM:	Out of memory, call again later
564 *      EIO:	A permanent error occurred, do not retry
565 */
566static int
567xprt_rdma_allocate(struct rpc_task *task)
568{
569	struct rpc_rqst *rqst = task->tk_rqstp;
570	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
571	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
572	gfp_t flags;
573
574	flags = RPCRDMA_DEF_GFP;
575	if (RPC_IS_ASYNC(task))
576		flags = GFP_NOWAIT | __GFP_NOWARN;
577	if (RPC_IS_SWAPPER(task))
578		flags |= __GFP_MEMALLOC;
579
580	if (!rpcrdma_check_regbuf(r_xprt, req->rl_sendbuf, rqst->rq_callsize,
581				  flags))
582		goto out_fail;
583	if (!rpcrdma_check_regbuf(r_xprt, req->rl_recvbuf, rqst->rq_rcvsize,
584				  flags))
585		goto out_fail;
586
587	rqst->rq_buffer = rdmab_data(req->rl_sendbuf);
588	rqst->rq_rbuffer = rdmab_data(req->rl_recvbuf);
589	return 0;
590
591out_fail:
592	return -ENOMEM;
593}
594
595/**
596 * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
597 * @task: RPC task
598 *
599 * Caller guarantees rqst->rq_buffer is non-NULL.
600 */
601static void
602xprt_rdma_free(struct rpc_task *task)
603{
604	struct rpc_rqst *rqst = task->tk_rqstp;
605	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
606	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
607
608	if (!list_empty(&req->rl_registered))
609		frwr_unmap_sync(r_xprt, req);
610
611	/* XXX: If the RPC is completing because of a signal and
612	 * not because a reply was received, we ought to ensure
613	 * that the Send completion has fired, so that memory
614	 * involved with the Send is not still visible to the NIC.
615	 */
616}
617
618/**
619 * xprt_rdma_send_request - marshal and send an RPC request
620 * @rqst: RPC message in rq_snd_buf
621 *
622 * Caller holds the transport's write lock.
623 *
624 * Returns:
625 *	%0 if the RPC message has been sent
626 *	%-ENOTCONN if the caller should reconnect and call again
627 *	%-EAGAIN if the caller should call again
628 *	%-ENOBUFS if the caller should call again after a delay
629 *	%-EMSGSIZE if encoding ran out of buffer space. The request
630 *		was not sent. Do not try to send this message again.
631 *	%-EIO if an I/O error occurred. The request was not sent.
632 *		Do not try to send this message again.
633 */
634static int
635xprt_rdma_send_request(struct rpc_rqst *rqst)
636{
637	struct rpc_xprt *xprt = rqst->rq_xprt;
638	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
639	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
640	int rc = 0;
641
642#if defined(CONFIG_SUNRPC_BACKCHANNEL)
643	if (unlikely(!rqst->rq_buffer))
644		return xprt_rdma_bc_send_reply(rqst);
645#endif	/* CONFIG_SUNRPC_BACKCHANNEL */
646
647	if (!xprt_connected(xprt))
648		return -ENOTCONN;
649
650	if (!xprt_request_get_cong(xprt, rqst))
651		return -EBADSLT;
652
653	rc = rpcrdma_marshal_req(r_xprt, rqst);
654	if (rc < 0)
655		goto failed_marshal;
656
657	/* Must suppress retransmit to maintain credits */
658	if (rqst->rq_connect_cookie == xprt->connect_cookie)
659		goto drop_connection;
660	rqst->rq_xtime = ktime_get();
661
662	if (rpcrdma_post_sends(r_xprt, req))
663		goto drop_connection;
664
665	rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
666
667	/* An RPC with no reply will throw off credit accounting,
668	 * so drop the connection to reset the credit grant.
669	 */
670	if (!rpc_reply_expected(rqst->rq_task))
671		goto drop_connection;
672	return 0;
673
674failed_marshal:
675	if (rc != -ENOTCONN)
676		return rc;
677drop_connection:
678	xprt_rdma_close(xprt);
679	return -ENOTCONN;
680}
681
682void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
683{
684	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
685	long idle_time = 0;
686
687	if (xprt_connected(xprt))
688		idle_time = (long)(jiffies - xprt->last_used) / HZ;
689
690	seq_puts(seq, "\txprt:\trdma ");
691	seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
692		   0,	/* need a local port? */
693		   xprt->stat.bind_count,
694		   xprt->stat.connect_count,
695		   xprt->stat.connect_time / HZ,
696		   idle_time,
697		   xprt->stat.sends,
698		   xprt->stat.recvs,
699		   xprt->stat.bad_xids,
700		   xprt->stat.req_u,
701		   xprt->stat.bklog_u);
702	seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
703		   r_xprt->rx_stats.read_chunk_count,
704		   r_xprt->rx_stats.write_chunk_count,
705		   r_xprt->rx_stats.reply_chunk_count,
706		   r_xprt->rx_stats.total_rdma_request,
707		   r_xprt->rx_stats.total_rdma_reply,
708		   r_xprt->rx_stats.pullup_copy_count,
709		   r_xprt->rx_stats.fixup_copy_count,
710		   r_xprt->rx_stats.hardway_register_count,
711		   r_xprt->rx_stats.failed_marshal_count,
712		   r_xprt->rx_stats.bad_reply_count,
713		   r_xprt->rx_stats.nomsg_call_count);
714	seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n",
715		   r_xprt->rx_stats.mrs_recycled,
716		   r_xprt->rx_stats.mrs_orphaned,
717		   r_xprt->rx_stats.mrs_allocated,
718		   r_xprt->rx_stats.local_inv_needed,
719		   r_xprt->rx_stats.empty_sendctx_q,
720		   r_xprt->rx_stats.reply_waits_for_send);
721}
722
723static int
724xprt_rdma_enable_swap(struct rpc_xprt *xprt)
725{
726	return 0;
727}
728
729static void
730xprt_rdma_disable_swap(struct rpc_xprt *xprt)
731{
732}
733
734/*
735 * Plumbing for rpc transport switch and kernel module
736 */
737
738static const struct rpc_xprt_ops xprt_rdma_procs = {
739	.reserve_xprt		= xprt_reserve_xprt_cong,
740	.release_xprt		= xprt_release_xprt_cong, /* sunrpc/xprt.c */
741	.alloc_slot		= xprt_rdma_alloc_slot,
742	.free_slot		= xprt_rdma_free_slot,
743	.release_request	= xprt_release_rqst_cong,       /* ditto */
744	.wait_for_reply_request	= xprt_wait_for_reply_request_def, /* ditto */
745	.timer			= xprt_rdma_timer,
746	.rpcbind		= rpcb_getport_async,	/* sunrpc/rpcb_clnt.c */
747	.set_port		= xprt_rdma_set_port,
748	.connect		= xprt_rdma_connect,
749	.buf_alloc		= xprt_rdma_allocate,
750	.buf_free		= xprt_rdma_free,
751	.send_request		= xprt_rdma_send_request,
752	.close			= xprt_rdma_close,
753	.destroy		= xprt_rdma_destroy,
754	.set_connect_timeout	= xprt_rdma_set_connect_timeout,
755	.print_stats		= xprt_rdma_print_stats,
756	.enable_swap		= xprt_rdma_enable_swap,
757	.disable_swap		= xprt_rdma_disable_swap,
758	.inject_disconnect	= xprt_rdma_inject_disconnect,
759#if defined(CONFIG_SUNRPC_BACKCHANNEL)
760	.bc_setup		= xprt_rdma_bc_setup,
761	.bc_maxpayload		= xprt_rdma_bc_maxpayload,
762	.bc_num_slots		= xprt_rdma_bc_max_slots,
763	.bc_free_rqst		= xprt_rdma_bc_free_rqst,
764	.bc_destroy		= xprt_rdma_bc_destroy,
765#endif
766};
767
768static struct xprt_class xprt_rdma = {
769	.list			= LIST_HEAD_INIT(xprt_rdma.list),
770	.name			= "rdma",
771	.owner			= THIS_MODULE,
772	.ident			= XPRT_TRANSPORT_RDMA,
773	.setup			= xprt_setup_rdma,
774	.netid			= { "rdma", "rdma6", "" },
775};
776
777void xprt_rdma_cleanup(void)
778{
779#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
780	if (sunrpc_table_header) {
781		unregister_sysctl_table(sunrpc_table_header);
782		sunrpc_table_header = NULL;
783	}
784#endif
785
786	xprt_unregister_transport(&xprt_rdma);
787	xprt_unregister_transport(&xprt_rdma_bc);
788}
789
790int xprt_rdma_init(void)
791{
792	int rc;
793
794	rc = xprt_register_transport(&xprt_rdma);
795	if (rc)
796		return rc;
797
798	rc = xprt_register_transport(&xprt_rdma_bc);
799	if (rc) {
800		xprt_unregister_transport(&xprt_rdma);
801		return rc;
802	}
803
804#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
805	if (!sunrpc_table_header)
806		sunrpc_table_header = register_sysctl_table(sunrpc_table);
807#endif
808	return 0;
809}
810