xref: /openbmc/linux/net/sunrpc/xprtrdma/transport.c (revision c127f98ba9aba1818a6ca3a1da5a24653a10d966)
1 /*
2  * Copyright (c) 2014-2017 Oracle.  All rights reserved.
3  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the BSD-type
9  * license below:
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  *
15  *      Redistributions of source code must retain the above copyright
16  *      notice, this list of conditions and the following disclaimer.
17  *
18  *      Redistributions in binary form must reproduce the above
19  *      copyright notice, this list of conditions and the following
20  *      disclaimer in the documentation and/or other materials provided
21  *      with the distribution.
22  *
23  *      Neither the name of the Network Appliance, Inc. nor the names of
24  *      its contributors may be used to endorse or promote products
25  *      derived from this software without specific prior written
26  *      permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39  */
40 
41 /*
42  * transport.c
43  *
44  * This file contains the top-level implementation of an RPC RDMA
45  * transport.
46  *
47  * Naming convention: functions beginning with xprt_ are part of the
48  * transport switch. All others are RPC RDMA internal.
49  */
50 
51 #include <linux/module.h>
52 #include <linux/slab.h>
53 #include <linux/seq_file.h>
54 #include <linux/sunrpc/addr.h>
55 #include <linux/smp.h>
56 
57 #include "xprt_rdma.h"
58 
59 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
60 # define RPCDBG_FACILITY	RPCDBG_TRANS
61 #endif
62 
63 /*
64  * tunables
65  */
66 
67 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
68 unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
69 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
70 static unsigned int xprt_rdma_inline_write_padding;
71 unsigned int xprt_rdma_memreg_strategy		= RPCRDMA_FRMR;
72 int xprt_rdma_pad_optimize;
73 
74 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
75 
76 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
77 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
78 static unsigned int min_inline_size = RPCRDMA_MIN_INLINE;
79 static unsigned int max_inline_size = RPCRDMA_MAX_INLINE;
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 struct 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_minmax,
103 		.extra1		= &min_inline_size,
104 		.extra2		= &max_inline_size,
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_minmax,
112 		.extra1		= &min_inline_size,
113 		.extra2		= &max_inline_size,
114 	},
115 	{
116 		.procname	= "rdma_inline_write_padding",
117 		.data		= &xprt_rdma_inline_write_padding,
118 		.maxlen		= sizeof(unsigned int),
119 		.mode		= 0644,
120 		.proc_handler	= proc_dointvec_minmax,
121 		.extra1		= &zero,
122 		.extra2		= &max_padding,
123 	},
124 	{
125 		.procname	= "rdma_memreg_strategy",
126 		.data		= &xprt_rdma_memreg_strategy,
127 		.maxlen		= sizeof(unsigned int),
128 		.mode		= 0644,
129 		.proc_handler	= proc_dointvec_minmax,
130 		.extra1		= &min_memreg,
131 		.extra2		= &max_memreg,
132 	},
133 	{
134 		.procname	= "rdma_pad_optimize",
135 		.data		= &xprt_rdma_pad_optimize,
136 		.maxlen		= sizeof(unsigned int),
137 		.mode		= 0644,
138 		.proc_handler	= proc_dointvec,
139 	},
140 	{ },
141 };
142 
143 static struct ctl_table sunrpc_table[] = {
144 	{
145 		.procname	= "sunrpc",
146 		.mode		= 0555,
147 		.child		= xr_tunables_table
148 	},
149 	{ },
150 };
151 
152 #endif
153 
154 static const struct rpc_xprt_ops xprt_rdma_procs;
155 
156 static void
157 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
158 {
159 	struct sockaddr_in *sin = (struct sockaddr_in *)sap;
160 	char buf[20];
161 
162 	snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
163 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
164 
165 	xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
166 }
167 
168 static void
169 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
170 {
171 	struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
172 	char buf[40];
173 
174 	snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
175 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
176 
177 	xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
178 }
179 
180 void
181 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
182 {
183 	char buf[128];
184 
185 	switch (sap->sa_family) {
186 	case AF_INET:
187 		xprt_rdma_format_addresses4(xprt, sap);
188 		break;
189 	case AF_INET6:
190 		xprt_rdma_format_addresses6(xprt, sap);
191 		break;
192 	default:
193 		pr_err("rpcrdma: Unrecognized address family\n");
194 		return;
195 	}
196 
197 	(void)rpc_ntop(sap, buf, sizeof(buf));
198 	xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
199 
200 	snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
201 	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
202 
203 	snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
204 	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
205 
206 	xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
207 }
208 
209 void
210 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
211 {
212 	unsigned int i;
213 
214 	for (i = 0; i < RPC_DISPLAY_MAX; i++)
215 		switch (i) {
216 		case RPC_DISPLAY_PROTO:
217 		case RPC_DISPLAY_NETID:
218 			continue;
219 		default:
220 			kfree(xprt->address_strings[i]);
221 		}
222 }
223 
224 void
225 rpcrdma_conn_func(struct rpcrdma_ep *ep)
226 {
227 	schedule_delayed_work(&ep->rep_connect_worker, 0);
228 }
229 
230 void
231 rpcrdma_connect_worker(struct work_struct *work)
232 {
233 	struct rpcrdma_ep *ep =
234 		container_of(work, struct rpcrdma_ep, rep_connect_worker.work);
235 	struct rpcrdma_xprt *r_xprt =
236 		container_of(ep, struct rpcrdma_xprt, rx_ep);
237 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
238 
239 	spin_lock_bh(&xprt->transport_lock);
240 	if (++xprt->connect_cookie == 0)	/* maintain a reserved value */
241 		++xprt->connect_cookie;
242 	if (ep->rep_connected > 0) {
243 		if (!xprt_test_and_set_connected(xprt))
244 			xprt_wake_pending_tasks(xprt, 0);
245 	} else {
246 		if (xprt_test_and_clear_connected(xprt))
247 			xprt_wake_pending_tasks(xprt, -ENOTCONN);
248 	}
249 	spin_unlock_bh(&xprt->transport_lock);
250 }
251 
252 static void
253 xprt_rdma_connect_worker(struct work_struct *work)
254 {
255 	struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
256 						   rx_connect_worker.work);
257 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
258 	int rc = 0;
259 
260 	xprt_clear_connected(xprt);
261 
262 	dprintk("RPC:       %s: %sconnect\n", __func__,
263 			r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
264 	rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
265 	if (rc)
266 		xprt_wake_pending_tasks(xprt, rc);
267 
268 	dprintk("RPC:       %s: exit\n", __func__);
269 	xprt_clear_connecting(xprt);
270 }
271 
272 static void
273 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
274 {
275 	struct rpcrdma_xprt *r_xprt = container_of(xprt, struct rpcrdma_xprt,
276 						   rx_xprt);
277 
278 	pr_info("rpcrdma: injecting transport disconnect on xprt=%p\n", xprt);
279 	rdma_disconnect(r_xprt->rx_ia.ri_id);
280 }
281 
282 /*
283  * xprt_rdma_destroy
284  *
285  * Destroy the xprt.
286  * Free all memory associated with the object, including its own.
287  * NOTE: none of the *destroy methods free memory for their top-level
288  * objects, even though they may have allocated it (they do free
289  * private memory). It's up to the caller to handle it. In this
290  * case (RDMA transport), all structure memory is inlined with the
291  * struct rpcrdma_xprt.
292  */
293 static void
294 xprt_rdma_destroy(struct rpc_xprt *xprt)
295 {
296 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
297 
298 	dprintk("RPC:       %s: called\n", __func__);
299 
300 	cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
301 
302 	xprt_clear_connected(xprt);
303 
304 	rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
305 	rpcrdma_buffer_destroy(&r_xprt->rx_buf);
306 	rpcrdma_ia_close(&r_xprt->rx_ia);
307 
308 	xprt_rdma_free_addresses(xprt);
309 
310 	xprt_free(xprt);
311 
312 	dprintk("RPC:       %s: returning\n", __func__);
313 
314 	module_put(THIS_MODULE);
315 }
316 
317 static const struct rpc_timeout xprt_rdma_default_timeout = {
318 	.to_initval = 60 * HZ,
319 	.to_maxval = 60 * HZ,
320 };
321 
322 /**
323  * xprt_setup_rdma - Set up transport to use RDMA
324  *
325  * @args: rpc transport arguments
326  */
327 static struct rpc_xprt *
328 xprt_setup_rdma(struct xprt_create *args)
329 {
330 	struct rpcrdma_create_data_internal cdata;
331 	struct rpc_xprt *xprt;
332 	struct rpcrdma_xprt *new_xprt;
333 	struct rpcrdma_ep *new_ep;
334 	struct sockaddr *sap;
335 	int rc;
336 
337 	if (args->addrlen > sizeof(xprt->addr)) {
338 		dprintk("RPC:       %s: address too large\n", __func__);
339 		return ERR_PTR(-EBADF);
340 	}
341 
342 	xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
343 			xprt_rdma_slot_table_entries,
344 			xprt_rdma_slot_table_entries);
345 	if (xprt == NULL) {
346 		dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
347 			__func__);
348 		return ERR_PTR(-ENOMEM);
349 	}
350 
351 	/* 60 second timeout, no retries */
352 	xprt->timeout = &xprt_rdma_default_timeout;
353 	xprt->bind_timeout = RPCRDMA_BIND_TO;
354 	xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
355 	xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
356 
357 	xprt->resvport = 0;		/* privileged port not needed */
358 	xprt->tsh_size = 0;		/* RPC-RDMA handles framing */
359 	xprt->ops = &xprt_rdma_procs;
360 
361 	/*
362 	 * Set up RDMA-specific connect data.
363 	 */
364 
365 	sap = (struct sockaddr *)&cdata.addr;
366 	memcpy(sap, args->dstaddr, args->addrlen);
367 
368 	/* Ensure xprt->addr holds valid server TCP (not RDMA)
369 	 * address, for any side protocols which peek at it */
370 	xprt->prot = IPPROTO_TCP;
371 	xprt->addrlen = args->addrlen;
372 	memcpy(&xprt->addr, sap, xprt->addrlen);
373 
374 	if (rpc_get_port(sap))
375 		xprt_set_bound(xprt);
376 
377 	cdata.max_requests = xprt->max_reqs;
378 
379 	cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
380 	cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
381 
382 	cdata.inline_wsize = xprt_rdma_max_inline_write;
383 	if (cdata.inline_wsize > cdata.wsize)
384 		cdata.inline_wsize = cdata.wsize;
385 
386 	cdata.inline_rsize = xprt_rdma_max_inline_read;
387 	if (cdata.inline_rsize > cdata.rsize)
388 		cdata.inline_rsize = cdata.rsize;
389 
390 	cdata.padding = xprt_rdma_inline_write_padding;
391 
392 	/*
393 	 * Create new transport instance, which includes initialized
394 	 *  o ia
395 	 *  o endpoint
396 	 *  o buffers
397 	 */
398 
399 	new_xprt = rpcx_to_rdmax(xprt);
400 
401 	rc = rpcrdma_ia_open(new_xprt, sap);
402 	if (rc)
403 		goto out1;
404 
405 	/*
406 	 * initialize and create ep
407 	 */
408 	new_xprt->rx_data = cdata;
409 	new_ep = &new_xprt->rx_ep;
410 	new_ep->rep_remote_addr = cdata.addr;
411 
412 	rc = rpcrdma_ep_create(&new_xprt->rx_ep,
413 				&new_xprt->rx_ia, &new_xprt->rx_data);
414 	if (rc)
415 		goto out2;
416 
417 	/*
418 	 * Allocate pre-registered send and receive buffers for headers and
419 	 * any inline data. Also specify any padding which will be provided
420 	 * from a preregistered zero buffer.
421 	 */
422 	rc = rpcrdma_buffer_create(new_xprt);
423 	if (rc)
424 		goto out3;
425 
426 	/*
427 	 * Register a callback for connection events. This is necessary because
428 	 * connection loss notification is async. We also catch connection loss
429 	 * when reaping receives.
430 	 */
431 	INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
432 			  xprt_rdma_connect_worker);
433 
434 	xprt_rdma_format_addresses(xprt, sap);
435 	xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
436 	if (xprt->max_payload == 0)
437 		goto out4;
438 	xprt->max_payload <<= PAGE_SHIFT;
439 	dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
440 		__func__, xprt->max_payload);
441 
442 	if (!try_module_get(THIS_MODULE))
443 		goto out4;
444 
445 	dprintk("RPC:       %s: %s:%s\n", __func__,
446 		xprt->address_strings[RPC_DISPLAY_ADDR],
447 		xprt->address_strings[RPC_DISPLAY_PORT]);
448 	return xprt;
449 
450 out4:
451 	xprt_rdma_free_addresses(xprt);
452 	rc = -EINVAL;
453 out3:
454 	rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
455 out2:
456 	rpcrdma_ia_close(&new_xprt->rx_ia);
457 out1:
458 	xprt_free(xprt);
459 	return ERR_PTR(rc);
460 }
461 
462 /**
463  * xprt_rdma_close - Close down RDMA connection
464  * @xprt: generic transport to be closed
465  *
466  * Called during transport shutdown reconnect, or device
467  * removal. Caller holds the transport's write lock.
468  */
469 static void
470 xprt_rdma_close(struct rpc_xprt *xprt)
471 {
472 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
473 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
474 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
475 
476 	dprintk("RPC:       %s: closing xprt %p\n", __func__, xprt);
477 
478 	if (test_and_clear_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags)) {
479 		xprt_clear_connected(xprt);
480 		rpcrdma_ia_remove(ia);
481 		return;
482 	}
483 	if (ep->rep_connected == -ENODEV)
484 		return;
485 	if (ep->rep_connected > 0)
486 		xprt->reestablish_timeout = 0;
487 	xprt_disconnect_done(xprt);
488 	rpcrdma_ep_disconnect(ep, ia);
489 }
490 
491 static void
492 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
493 {
494 	struct sockaddr_in *sap;
495 
496 	sap = (struct sockaddr_in *)&xprt->addr;
497 	sap->sin_port = htons(port);
498 	sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
499 	sap->sin_port = htons(port);
500 	dprintk("RPC:       %s: %u\n", __func__, port);
501 }
502 
503 /**
504  * xprt_rdma_timer - invoked when an RPC times out
505  * @xprt: controlling RPC transport
506  * @task: RPC task that timed out
507  *
508  * Invoked when the transport is still connected, but an RPC
509  * retransmit timeout occurs.
510  *
511  * Since RDMA connections don't have a keep-alive, forcibly
512  * disconnect and retry to connect. This drives full
513  * detection of the network path, and retransmissions of
514  * all pending RPCs.
515  */
516 static void
517 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
518 {
519 	dprintk("RPC: %5u %s: xprt = %p\n", task->tk_pid, __func__, xprt);
520 
521 	xprt_force_disconnect(xprt);
522 }
523 
524 static void
525 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
526 {
527 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
528 
529 	if (r_xprt->rx_ep.rep_connected != 0) {
530 		/* Reconnect */
531 		schedule_delayed_work(&r_xprt->rx_connect_worker,
532 				      xprt->reestablish_timeout);
533 		xprt->reestablish_timeout <<= 1;
534 		if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
535 			xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
536 		else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
537 			xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
538 	} else {
539 		schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
540 		if (!RPC_IS_ASYNC(task))
541 			flush_delayed_work(&r_xprt->rx_connect_worker);
542 	}
543 }
544 
545 /* Allocate a fixed-size buffer in which to construct and send the
546  * RPC-over-RDMA header for this request.
547  */
548 static bool
549 rpcrdma_get_rdmabuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
550 		    gfp_t flags)
551 {
552 	size_t size = RPCRDMA_HDRBUF_SIZE;
553 	struct rpcrdma_regbuf *rb;
554 
555 	if (req->rl_rdmabuf)
556 		return true;
557 
558 	rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
559 	if (IS_ERR(rb))
560 		return false;
561 
562 	r_xprt->rx_stats.hardway_register_count += size;
563 	req->rl_rdmabuf = rb;
564 	xdr_buf_init(&req->rl_hdrbuf, rb->rg_base, rdmab_length(rb));
565 	return true;
566 }
567 
568 static bool
569 rpcrdma_get_sendbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
570 		    size_t size, gfp_t flags)
571 {
572 	struct rpcrdma_regbuf *rb;
573 
574 	if (req->rl_sendbuf && rdmab_length(req->rl_sendbuf) >= size)
575 		return true;
576 
577 	rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
578 	if (IS_ERR(rb))
579 		return false;
580 
581 	rpcrdma_free_regbuf(req->rl_sendbuf);
582 	r_xprt->rx_stats.hardway_register_count += size;
583 	req->rl_sendbuf = rb;
584 	return true;
585 }
586 
587 /* The rq_rcv_buf is used only if a Reply chunk is necessary.
588  * The decision to use a Reply chunk is made later in
589  * rpcrdma_marshal_req. This buffer is registered at that time.
590  *
591  * Otherwise, the associated RPC Reply arrives in a separate
592  * Receive buffer, arbitrarily chosen by the HCA. The buffer
593  * allocated here for the RPC Reply is not utilized in that
594  * case. See rpcrdma_inline_fixup.
595  *
596  * A regbuf is used here to remember the buffer size.
597  */
598 static bool
599 rpcrdma_get_recvbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
600 		    size_t size, gfp_t flags)
601 {
602 	struct rpcrdma_regbuf *rb;
603 
604 	if (req->rl_recvbuf && rdmab_length(req->rl_recvbuf) >= size)
605 		return true;
606 
607 	rb = rpcrdma_alloc_regbuf(size, DMA_NONE, flags);
608 	if (IS_ERR(rb))
609 		return false;
610 
611 	rpcrdma_free_regbuf(req->rl_recvbuf);
612 	r_xprt->rx_stats.hardway_register_count += size;
613 	req->rl_recvbuf = rb;
614 	return true;
615 }
616 
617 /**
618  * xprt_rdma_allocate - allocate transport resources for an RPC
619  * @task: RPC task
620  *
621  * Return values:
622  *        0:	Success; rq_buffer points to RPC buffer to use
623  *   ENOMEM:	Out of memory, call again later
624  *      EIO:	A permanent error occurred, do not retry
625  *
626  * The RDMA allocate/free functions need the task structure as a place
627  * to hide the struct rpcrdma_req, which is necessary for the actual
628  * send/recv sequence.
629  *
630  * xprt_rdma_allocate provides buffers that are already mapped for
631  * DMA, and a local DMA lkey is provided for each.
632  */
633 static int
634 xprt_rdma_allocate(struct rpc_task *task)
635 {
636 	struct rpc_rqst *rqst = task->tk_rqstp;
637 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
638 	struct rpcrdma_req *req;
639 	gfp_t flags;
640 
641 	req = rpcrdma_buffer_get(&r_xprt->rx_buf);
642 	if (req == NULL)
643 		return -ENOMEM;
644 
645 	flags = RPCRDMA_DEF_GFP;
646 	if (RPC_IS_SWAPPER(task))
647 		flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
648 
649 	if (!rpcrdma_get_rdmabuf(r_xprt, req, flags))
650 		goto out_fail;
651 	if (!rpcrdma_get_sendbuf(r_xprt, req, rqst->rq_callsize, flags))
652 		goto out_fail;
653 	if (!rpcrdma_get_recvbuf(r_xprt, req, rqst->rq_rcvsize, flags))
654 		goto out_fail;
655 
656 	dprintk("RPC: %5u %s: send size = %zd, recv size = %zd, req = %p\n",
657 		task->tk_pid, __func__, rqst->rq_callsize,
658 		rqst->rq_rcvsize, req);
659 
660 	req->rl_cpu = smp_processor_id();
661 	req->rl_connect_cookie = 0;	/* our reserved value */
662 	rpcrdma_set_xprtdata(rqst, req);
663 	rqst->rq_buffer = req->rl_sendbuf->rg_base;
664 	rqst->rq_rbuffer = req->rl_recvbuf->rg_base;
665 	return 0;
666 
667 out_fail:
668 	rpcrdma_buffer_put(req);
669 	return -ENOMEM;
670 }
671 
672 /**
673  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
674  * @task: RPC task
675  *
676  * Caller guarantees rqst->rq_buffer is non-NULL.
677  */
678 static void
679 xprt_rdma_free(struct rpc_task *task)
680 {
681 	struct rpc_rqst *rqst = task->tk_rqstp;
682 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
683 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
684 
685 	if (test_bit(RPCRDMA_REQ_F_BACKCHANNEL, &req->rl_flags))
686 		return;
687 
688 	dprintk("RPC:       %s: called on 0x%p\n", __func__, req->rl_reply);
689 
690 	if (test_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags))
691 		rpcrdma_release_rqst(r_xprt, req);
692 	rpcrdma_buffer_put(req);
693 }
694 
695 /**
696  * xprt_rdma_send_request - marshal and send an RPC request
697  * @task: RPC task with an RPC message in rq_snd_buf
698  *
699  * Caller holds the transport's write lock.
700  *
701  * Return values:
702  *        0:	The request has been sent
703  * ENOTCONN:	Caller needs to invoke connect logic then call again
704  *  ENOBUFS:	Call again later to send the request
705  *      EIO:	A permanent error occurred. The request was not sent,
706  *		and don't try it again
707  *
708  * send_request invokes the meat of RPC RDMA. It must do the following:
709  *
710  *  1.  Marshal the RPC request into an RPC RDMA request, which means
711  *	putting a header in front of data, and creating IOVs for RDMA
712  *	from those in the request.
713  *  2.  In marshaling, detect opportunities for RDMA, and use them.
714  *  3.  Post a recv message to set up asynch completion, then send
715  *	the request (rpcrdma_ep_post).
716  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
717  */
718 static int
719 xprt_rdma_send_request(struct rpc_task *task)
720 {
721 	struct rpc_rqst *rqst = task->tk_rqstp;
722 	struct rpc_xprt *xprt = rqst->rq_xprt;
723 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
724 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
725 	int rc = 0;
726 
727 	if (!xprt_connected(xprt))
728 		goto drop_connection;
729 
730 	/* On retransmit, remove any previously registered chunks */
731 	if (unlikely(!list_empty(&req->rl_registered)))
732 		r_xprt->rx_ia.ri_ops->ro_unmap_sync(r_xprt,
733 						    &req->rl_registered);
734 
735 	rc = rpcrdma_marshal_req(r_xprt, rqst);
736 	if (rc < 0)
737 		goto failed_marshal;
738 
739 	if (req->rl_reply == NULL) 		/* e.g. reconnection */
740 		rpcrdma_recv_buffer_get(req);
741 
742 	/* Must suppress retransmit to maintain credits */
743 	if (req->rl_connect_cookie == xprt->connect_cookie)
744 		goto drop_connection;
745 	req->rl_connect_cookie = xprt->connect_cookie;
746 
747 	set_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags);
748 	if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
749 		goto drop_connection;
750 
751 	rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
752 	rqst->rq_bytes_sent = 0;
753 	return 0;
754 
755 failed_marshal:
756 	if (rc != -ENOTCONN)
757 		return rc;
758 drop_connection:
759 	xprt_disconnect_done(xprt);
760 	return -ENOTCONN;	/* implies disconnect */
761 }
762 
763 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
764 {
765 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
766 	long idle_time = 0;
767 
768 	if (xprt_connected(xprt))
769 		idle_time = (long)(jiffies - xprt->last_used) / HZ;
770 
771 	seq_puts(seq, "\txprt:\trdma ");
772 	seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
773 		   0,	/* need a local port? */
774 		   xprt->stat.bind_count,
775 		   xprt->stat.connect_count,
776 		   xprt->stat.connect_time,
777 		   idle_time,
778 		   xprt->stat.sends,
779 		   xprt->stat.recvs,
780 		   xprt->stat.bad_xids,
781 		   xprt->stat.req_u,
782 		   xprt->stat.bklog_u);
783 	seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
784 		   r_xprt->rx_stats.read_chunk_count,
785 		   r_xprt->rx_stats.write_chunk_count,
786 		   r_xprt->rx_stats.reply_chunk_count,
787 		   r_xprt->rx_stats.total_rdma_request,
788 		   r_xprt->rx_stats.total_rdma_reply,
789 		   r_xprt->rx_stats.pullup_copy_count,
790 		   r_xprt->rx_stats.fixup_copy_count,
791 		   r_xprt->rx_stats.hardway_register_count,
792 		   r_xprt->rx_stats.failed_marshal_count,
793 		   r_xprt->rx_stats.bad_reply_count,
794 		   r_xprt->rx_stats.nomsg_call_count);
795 	seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n",
796 		   r_xprt->rx_stats.mrs_recovered,
797 		   r_xprt->rx_stats.mrs_orphaned,
798 		   r_xprt->rx_stats.mrs_allocated,
799 		   r_xprt->rx_stats.local_inv_needed,
800 		   r_xprt->rx_stats.empty_sendctx_q,
801 		   r_xprt->rx_stats.reply_waits_for_send);
802 }
803 
804 static int
805 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
806 {
807 	return 0;
808 }
809 
810 static void
811 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
812 {
813 }
814 
815 /*
816  * Plumbing for rpc transport switch and kernel module
817  */
818 
819 static const struct rpc_xprt_ops xprt_rdma_procs = {
820 	.reserve_xprt		= xprt_reserve_xprt_cong,
821 	.release_xprt		= xprt_release_xprt_cong, /* sunrpc/xprt.c */
822 	.alloc_slot		= xprt_alloc_slot,
823 	.release_request	= xprt_release_rqst_cong,       /* ditto */
824 	.set_retrans_timeout	= xprt_set_retrans_timeout_def, /* ditto */
825 	.timer			= xprt_rdma_timer,
826 	.rpcbind		= rpcb_getport_async,	/* sunrpc/rpcb_clnt.c */
827 	.set_port		= xprt_rdma_set_port,
828 	.connect		= xprt_rdma_connect,
829 	.buf_alloc		= xprt_rdma_allocate,
830 	.buf_free		= xprt_rdma_free,
831 	.send_request		= xprt_rdma_send_request,
832 	.close			= xprt_rdma_close,
833 	.destroy		= xprt_rdma_destroy,
834 	.print_stats		= xprt_rdma_print_stats,
835 	.enable_swap		= xprt_rdma_enable_swap,
836 	.disable_swap		= xprt_rdma_disable_swap,
837 	.inject_disconnect	= xprt_rdma_inject_disconnect,
838 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
839 	.bc_setup		= xprt_rdma_bc_setup,
840 	.bc_up			= xprt_rdma_bc_up,
841 	.bc_maxpayload		= xprt_rdma_bc_maxpayload,
842 	.bc_free_rqst		= xprt_rdma_bc_free_rqst,
843 	.bc_destroy		= xprt_rdma_bc_destroy,
844 #endif
845 };
846 
847 static struct xprt_class xprt_rdma = {
848 	.list			= LIST_HEAD_INIT(xprt_rdma.list),
849 	.name			= "rdma",
850 	.owner			= THIS_MODULE,
851 	.ident			= XPRT_TRANSPORT_RDMA,
852 	.setup			= xprt_setup_rdma,
853 };
854 
855 void xprt_rdma_cleanup(void)
856 {
857 	int rc;
858 
859 	dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
860 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
861 	if (sunrpc_table_header) {
862 		unregister_sysctl_table(sunrpc_table_header);
863 		sunrpc_table_header = NULL;
864 	}
865 #endif
866 	rc = xprt_unregister_transport(&xprt_rdma);
867 	if (rc)
868 		dprintk("RPC:       %s: xprt_unregister returned %i\n",
869 			__func__, rc);
870 
871 	rpcrdma_destroy_wq();
872 
873 	rc = xprt_unregister_transport(&xprt_rdma_bc);
874 	if (rc)
875 		dprintk("RPC:       %s: xprt_unregister(bc) returned %i\n",
876 			__func__, rc);
877 }
878 
879 int xprt_rdma_init(void)
880 {
881 	int rc;
882 
883 	rc = rpcrdma_alloc_wq();
884 	if (rc)
885 		return rc;
886 
887 	rc = xprt_register_transport(&xprt_rdma);
888 	if (rc) {
889 		rpcrdma_destroy_wq();
890 		return rc;
891 	}
892 
893 	rc = xprt_register_transport(&xprt_rdma_bc);
894 	if (rc) {
895 		xprt_unregister_transport(&xprt_rdma);
896 		rpcrdma_destroy_wq();
897 		return rc;
898 	}
899 
900 	dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
901 
902 	dprintk("Defaults:\n");
903 	dprintk("\tSlots %d\n"
904 		"\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
905 		xprt_rdma_slot_table_entries,
906 		xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
907 	dprintk("\tPadding %d\n\tMemreg %d\n",
908 		xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
909 
910 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
911 	if (!sunrpc_table_header)
912 		sunrpc_table_header = register_sysctl_table(sunrpc_table);
913 #endif
914 	return 0;
915 }
916