• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/slab.h>
52 #include <linux/seq_file.h>
53 #include <linux/sunrpc/addr.h>
54 
55 #include "xprt_rdma.h"
56 
57 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
58 # define RPCDBG_FACILITY	RPCDBG_TRANS
59 #endif
60 
61 /*
62  * tunables
63  */
64 
65 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
66 unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
67 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
68 static unsigned int xprt_rdma_inline_write_padding;
69 unsigned int xprt_rdma_memreg_strategy		= RPCRDMA_FRMR;
70 int xprt_rdma_pad_optimize;
71 
72 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
73 
74 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
75 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
76 static unsigned int min_inline_size = RPCRDMA_MIN_INLINE;
77 static unsigned int max_inline_size = RPCRDMA_MAX_INLINE;
78 static unsigned int zero;
79 static unsigned int max_padding = PAGE_SIZE;
80 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
81 static unsigned int max_memreg = RPCRDMA_LAST - 1;
82 
83 static struct ctl_table_header *sunrpc_table_header;
84 
85 static struct ctl_table xr_tunables_table[] = {
86 	{
87 		.procname	= "rdma_slot_table_entries",
88 		.data		= &xprt_rdma_slot_table_entries,
89 		.maxlen		= sizeof(unsigned int),
90 		.mode		= 0644,
91 		.proc_handler	= proc_dointvec_minmax,
92 		.extra1		= &min_slot_table_size,
93 		.extra2		= &max_slot_table_size
94 	},
95 	{
96 		.procname	= "rdma_max_inline_read",
97 		.data		= &xprt_rdma_max_inline_read,
98 		.maxlen		= sizeof(unsigned int),
99 		.mode		= 0644,
100 		.proc_handler	= proc_dointvec_minmax,
101 		.extra1		= &min_inline_size,
102 		.extra2		= &max_inline_size,
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_minmax,
110 		.extra1		= &min_inline_size,
111 		.extra2		= &max_inline_size,
112 	},
113 	{
114 		.procname	= "rdma_inline_write_padding",
115 		.data		= &xprt_rdma_inline_write_padding,
116 		.maxlen		= sizeof(unsigned int),
117 		.mode		= 0644,
118 		.proc_handler	= proc_dointvec_minmax,
119 		.extra1		= &zero,
120 		.extra2		= &max_padding,
121 	},
122 	{
123 		.procname	= "rdma_memreg_strategy",
124 		.data		= &xprt_rdma_memreg_strategy,
125 		.maxlen		= sizeof(unsigned int),
126 		.mode		= 0644,
127 		.proc_handler	= proc_dointvec_minmax,
128 		.extra1		= &min_memreg,
129 		.extra2		= &max_memreg,
130 	},
131 	{
132 		.procname	= "rdma_pad_optimize",
133 		.data		= &xprt_rdma_pad_optimize,
134 		.maxlen		= sizeof(unsigned int),
135 		.mode		= 0644,
136 		.proc_handler	= proc_dointvec,
137 	},
138 	{ },
139 };
140 
141 static 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 static const struct rpc_xprt_ops xprt_rdma_procs;
153 
154 static void
xprt_rdma_format_addresses4(struct rpc_xprt * xprt,struct sockaddr * sap)155 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
156 {
157 	struct sockaddr_in *sin = (struct sockaddr_in *)sap;
158 	char buf[20];
159 
160 	snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
161 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
162 
163 	xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
164 }
165 
166 static void
xprt_rdma_format_addresses6(struct rpc_xprt * xprt,struct sockaddr * sap)167 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
168 {
169 	struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
170 	char buf[40];
171 
172 	snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
173 	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
174 
175 	xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
176 }
177 
178 void
xprt_rdma_format_addresses(struct rpc_xprt * xprt,struct sockaddr * sap)179 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
180 {
181 	char buf[128];
182 
183 	switch (sap->sa_family) {
184 	case AF_INET:
185 		xprt_rdma_format_addresses4(xprt, sap);
186 		break;
187 	case AF_INET6:
188 		xprt_rdma_format_addresses6(xprt, sap);
189 		break;
190 	default:
191 		pr_err("rpcrdma: Unrecognized address family\n");
192 		return;
193 	}
194 
195 	(void)rpc_ntop(sap, buf, sizeof(buf));
196 	xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
197 
198 	snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
199 	xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
200 
201 	snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
202 	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
203 
204 	xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
205 }
206 
207 void
xprt_rdma_free_addresses(struct rpc_xprt * xprt)208 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
209 {
210 	unsigned int i;
211 
212 	for (i = 0; i < RPC_DISPLAY_MAX; i++)
213 		switch (i) {
214 		case RPC_DISPLAY_PROTO:
215 		case RPC_DISPLAY_NETID:
216 			continue;
217 		default:
218 			kfree(xprt->address_strings[i]);
219 		}
220 }
221 
222 void
rpcrdma_conn_func(struct rpcrdma_ep * ep)223 rpcrdma_conn_func(struct rpcrdma_ep *ep)
224 {
225 	schedule_delayed_work(&ep->rep_connect_worker, 0);
226 }
227 
228 void
rpcrdma_connect_worker(struct work_struct * work)229 rpcrdma_connect_worker(struct work_struct *work)
230 {
231 	struct rpcrdma_ep *ep =
232 		container_of(work, struct rpcrdma_ep, rep_connect_worker.work);
233 	struct rpcrdma_xprt *r_xprt =
234 		container_of(ep, struct rpcrdma_xprt, rx_ep);
235 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
236 
237 	spin_lock_bh(&xprt->transport_lock);
238 	if (++xprt->connect_cookie == 0)	/* maintain a reserved value */
239 		++xprt->connect_cookie;
240 	if (ep->rep_connected > 0) {
241 		if (!xprt_test_and_set_connected(xprt)) {
242 			xprt->stat.connect_count++;
243 			xprt->stat.connect_time += (long)jiffies -
244 						   xprt->stat.connect_start;
245 			xprt_wake_pending_tasks(xprt, 0);
246 		}
247 	} else {
248 		if (xprt_test_and_clear_connected(xprt))
249 			xprt_wake_pending_tasks(xprt, -ENOTCONN);
250 	}
251 	spin_unlock_bh(&xprt->transport_lock);
252 }
253 
254 static void
xprt_rdma_connect_worker(struct work_struct * work)255 xprt_rdma_connect_worker(struct work_struct *work)
256 {
257 	struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
258 						   rx_connect_worker.work);
259 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
260 	int rc = 0;
261 
262 	xprt_clear_connected(xprt);
263 
264 	dprintk("RPC:       %s: %sconnect\n", __func__,
265 			r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
266 	rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
267 	if (rc)
268 		xprt_wake_pending_tasks(xprt, rc);
269 
270 	dprintk("RPC:       %s: exit\n", __func__);
271 	xprt_clear_connecting(xprt);
272 }
273 
274 static void
xprt_rdma_inject_disconnect(struct rpc_xprt * xprt)275 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
276 {
277 	struct rpcrdma_xprt *r_xprt = container_of(xprt, struct rpcrdma_xprt,
278 						   rx_xprt);
279 
280 	pr_info("rpcrdma: injecting transport disconnect on xprt=%p\n", xprt);
281 	rdma_disconnect(r_xprt->rx_ia.ri_id);
282 }
283 
284 /*
285  * xprt_rdma_destroy
286  *
287  * Destroy the xprt.
288  * Free all memory associated with the object, including its own.
289  * NOTE: none of the *destroy methods free memory for their top-level
290  * objects, even though they may have allocated it (they do free
291  * private memory). It's up to the caller to handle it. In this
292  * case (RDMA transport), all structure memory is inlined with the
293  * struct rpcrdma_xprt.
294  */
295 static void
xprt_rdma_destroy(struct rpc_xprt * xprt)296 xprt_rdma_destroy(struct rpc_xprt *xprt)
297 {
298 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
299 
300 	dprintk("RPC:       %s: called\n", __func__);
301 
302 	cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
303 
304 	xprt_clear_connected(xprt);
305 
306 	rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
307 	rpcrdma_buffer_destroy(&r_xprt->rx_buf);
308 	rpcrdma_ia_close(&r_xprt->rx_ia);
309 
310 	xprt_rdma_free_addresses(xprt);
311 
312 	xprt_free(xprt);
313 
314 	dprintk("RPC:       %s: returning\n", __func__);
315 
316 	module_put(THIS_MODULE);
317 }
318 
319 static const struct rpc_timeout xprt_rdma_default_timeout = {
320 	.to_initval = 60 * HZ,
321 	.to_maxval = 60 * HZ,
322 };
323 
324 /**
325  * xprt_setup_rdma - Set up transport to use RDMA
326  *
327  * @args: rpc transport arguments
328  */
329 static struct rpc_xprt *
xprt_setup_rdma(struct xprt_create * args)330 xprt_setup_rdma(struct xprt_create *args)
331 {
332 	struct rpcrdma_create_data_internal cdata;
333 	struct rpc_xprt *xprt;
334 	struct rpcrdma_xprt *new_xprt;
335 	struct rpcrdma_ep *new_ep;
336 	struct sockaddr *sap;
337 	int rc;
338 
339 	if (args->addrlen > sizeof(xprt->addr)) {
340 		dprintk("RPC:       %s: address too large\n", __func__);
341 		return ERR_PTR(-EBADF);
342 	}
343 
344 	xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
345 			xprt_rdma_slot_table_entries,
346 			xprt_rdma_slot_table_entries);
347 	if (xprt == NULL) {
348 		dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
349 			__func__);
350 		return ERR_PTR(-ENOMEM);
351 	}
352 
353 	/* 60 second timeout, no retries */
354 	xprt->timeout = &xprt_rdma_default_timeout;
355 	xprt->bind_timeout = RPCRDMA_BIND_TO;
356 	xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
357 	xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
358 
359 	xprt->resvport = 0;		/* privileged port not needed */
360 	xprt->tsh_size = 0;		/* RPC-RDMA handles framing */
361 	xprt->ops = &xprt_rdma_procs;
362 
363 	/*
364 	 * Set up RDMA-specific connect data.
365 	 */
366 
367 	sap = (struct sockaddr *)&cdata.addr;
368 	memcpy(sap, args->dstaddr, args->addrlen);
369 
370 	/* Ensure xprt->addr holds valid server TCP (not RDMA)
371 	 * address, for any side protocols which peek at it */
372 	xprt->prot = IPPROTO_TCP;
373 	xprt->addrlen = args->addrlen;
374 	memcpy(&xprt->addr, sap, xprt->addrlen);
375 
376 	if (rpc_get_port(sap))
377 		xprt_set_bound(xprt);
378 
379 	cdata.max_requests = xprt->max_reqs;
380 
381 	cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
382 	cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
383 
384 	cdata.inline_wsize = xprt_rdma_max_inline_write;
385 	if (cdata.inline_wsize > cdata.wsize)
386 		cdata.inline_wsize = cdata.wsize;
387 
388 	cdata.inline_rsize = xprt_rdma_max_inline_read;
389 	if (cdata.inline_rsize > cdata.rsize)
390 		cdata.inline_rsize = cdata.rsize;
391 
392 	cdata.padding = xprt_rdma_inline_write_padding;
393 
394 	/*
395 	 * Create new transport instance, which includes initialized
396 	 *  o ia
397 	 *  o endpoint
398 	 *  o buffers
399 	 */
400 
401 	new_xprt = rpcx_to_rdmax(xprt);
402 
403 	rc = rpcrdma_ia_open(new_xprt, sap);
404 	if (rc)
405 		goto out1;
406 
407 	/*
408 	 * initialize and create ep
409 	 */
410 	new_xprt->rx_data = cdata;
411 	new_ep = &new_xprt->rx_ep;
412 	new_ep->rep_remote_addr = cdata.addr;
413 
414 	rc = rpcrdma_ep_create(&new_xprt->rx_ep,
415 				&new_xprt->rx_ia, &new_xprt->rx_data);
416 	if (rc)
417 		goto out2;
418 
419 	/*
420 	 * Allocate pre-registered send and receive buffers for headers and
421 	 * any inline data. Also specify any padding which will be provided
422 	 * from a preregistered zero buffer.
423 	 */
424 	rc = rpcrdma_buffer_create(new_xprt);
425 	if (rc)
426 		goto out3;
427 
428 	/*
429 	 * Register a callback for connection events. This is necessary because
430 	 * connection loss notification is async. We also catch connection loss
431 	 * when reaping receives.
432 	 */
433 	INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
434 			  xprt_rdma_connect_worker);
435 
436 	xprt_rdma_format_addresses(xprt, sap);
437 	xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
438 	if (xprt->max_payload == 0)
439 		goto out4;
440 	xprt->max_payload <<= PAGE_SHIFT;
441 	dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
442 		__func__, xprt->max_payload);
443 
444 	if (!try_module_get(THIS_MODULE))
445 		goto out4;
446 
447 	dprintk("RPC:       %s: %s:%s\n", __func__,
448 		xprt->address_strings[RPC_DISPLAY_ADDR],
449 		xprt->address_strings[RPC_DISPLAY_PORT]);
450 	return xprt;
451 
452 out4:
453 	xprt_rdma_free_addresses(xprt);
454 	rc = -EINVAL;
455 out3:
456 	rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
457 out2:
458 	rpcrdma_ia_close(&new_xprt->rx_ia);
459 out1:
460 	xprt_free(xprt);
461 	return ERR_PTR(rc);
462 }
463 
464 /**
465  * xprt_rdma_close - Close down RDMA connection
466  * @xprt: generic transport to be closed
467  *
468  * Called during transport shutdown reconnect, or device
469  * removal. Caller holds the transport's write lock.
470  */
471 static void
xprt_rdma_close(struct rpc_xprt * xprt)472 xprt_rdma_close(struct rpc_xprt *xprt)
473 {
474 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
475 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
476 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
477 
478 	dprintk("RPC:       %s: closing xprt %p\n", __func__, xprt);
479 
480 	if (test_and_clear_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags)) {
481 		xprt_clear_connected(xprt);
482 		rpcrdma_ia_remove(ia);
483 		return;
484 	}
485 	if (ep->rep_connected == -ENODEV)
486 		return;
487 	if (ep->rep_connected > 0)
488 		xprt->reestablish_timeout = 0;
489 	xprt_disconnect_done(xprt);
490 	rpcrdma_ep_disconnect(ep, ia);
491 }
492 
493 static void
xprt_rdma_set_port(struct rpc_xprt * xprt,u16 port)494 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
495 {
496 	struct sockaddr_in *sap;
497 
498 	sap = (struct sockaddr_in *)&xprt->addr;
499 	sap->sin_port = htons(port);
500 	sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
501 	sap->sin_port = htons(port);
502 	dprintk("RPC:       %s: %u\n", __func__, port);
503 }
504 
505 /**
506  * xprt_rdma_timer - invoked when an RPC times out
507  * @xprt: controlling RPC transport
508  * @task: RPC task that timed out
509  *
510  * Invoked when the transport is still connected, but an RPC
511  * retransmit timeout occurs.
512  *
513  * Since RDMA connections don't have a keep-alive, forcibly
514  * disconnect and retry to connect. This drives full
515  * detection of the network path, and retransmissions of
516  * all pending RPCs.
517  */
518 static void
xprt_rdma_timer(struct rpc_xprt * xprt,struct rpc_task * task)519 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
520 {
521 	dprintk("RPC: %5u %s: xprt = %p\n", task->tk_pid, __func__, xprt);
522 
523 	xprt_force_disconnect(xprt);
524 }
525 
526 static void
xprt_rdma_connect(struct rpc_xprt * xprt,struct rpc_task * task)527 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
528 {
529 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
530 
531 	if (r_xprt->rx_ep.rep_connected != 0) {
532 		/* Reconnect */
533 		schedule_delayed_work(&r_xprt->rx_connect_worker,
534 				      xprt->reestablish_timeout);
535 		xprt->reestablish_timeout <<= 1;
536 		if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
537 			xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
538 		else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
539 			xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
540 	} else {
541 		schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
542 		if (!RPC_IS_ASYNC(task))
543 			flush_delayed_work(&r_xprt->rx_connect_worker);
544 	}
545 }
546 
547 /* Allocate a fixed-size buffer in which to construct and send the
548  * RPC-over-RDMA header for this request.
549  */
550 static bool
rpcrdma_get_rdmabuf(struct rpcrdma_xprt * r_xprt,struct rpcrdma_req * req,gfp_t flags)551 rpcrdma_get_rdmabuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
552 		    gfp_t flags)
553 {
554 	size_t size = RPCRDMA_HDRBUF_SIZE;
555 	struct rpcrdma_regbuf *rb;
556 
557 	if (req->rl_rdmabuf)
558 		return true;
559 
560 	rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
561 	if (IS_ERR(rb))
562 		return false;
563 
564 	r_xprt->rx_stats.hardway_register_count += size;
565 	req->rl_rdmabuf = rb;
566 	xdr_buf_init(&req->rl_hdrbuf, rb->rg_base, rdmab_length(rb));
567 	return true;
568 }
569 
570 static bool
rpcrdma_get_sendbuf(struct rpcrdma_xprt * r_xprt,struct rpcrdma_req * req,size_t size,gfp_t flags)571 rpcrdma_get_sendbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
572 		    size_t size, gfp_t flags)
573 {
574 	struct rpcrdma_regbuf *rb;
575 
576 	if (req->rl_sendbuf && rdmab_length(req->rl_sendbuf) >= size)
577 		return true;
578 
579 	rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
580 	if (IS_ERR(rb))
581 		return false;
582 
583 	rpcrdma_free_regbuf(req->rl_sendbuf);
584 	r_xprt->rx_stats.hardway_register_count += size;
585 	req->rl_sendbuf = rb;
586 	return true;
587 }
588 
589 /* The rq_rcv_buf is used only if a Reply chunk is necessary.
590  * The decision to use a Reply chunk is made later in
591  * rpcrdma_marshal_req. This buffer is registered at that time.
592  *
593  * Otherwise, the associated RPC Reply arrives in a separate
594  * Receive buffer, arbitrarily chosen by the HCA. The buffer
595  * allocated here for the RPC Reply is not utilized in that
596  * case. See rpcrdma_inline_fixup.
597  *
598  * A regbuf is used here to remember the buffer size.
599  */
600 static bool
rpcrdma_get_recvbuf(struct rpcrdma_xprt * r_xprt,struct rpcrdma_req * req,size_t size,gfp_t flags)601 rpcrdma_get_recvbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
602 		    size_t size, gfp_t flags)
603 {
604 	struct rpcrdma_regbuf *rb;
605 
606 	if (req->rl_recvbuf && rdmab_length(req->rl_recvbuf) >= size)
607 		return true;
608 
609 	rb = rpcrdma_alloc_regbuf(size, DMA_NONE, flags);
610 	if (IS_ERR(rb))
611 		return false;
612 
613 	rpcrdma_free_regbuf(req->rl_recvbuf);
614 	r_xprt->rx_stats.hardway_register_count += size;
615 	req->rl_recvbuf = rb;
616 	return true;
617 }
618 
619 /**
620  * xprt_rdma_allocate - allocate transport resources for an RPC
621  * @task: RPC task
622  *
623  * Return values:
624  *        0:	Success; rq_buffer points to RPC buffer to use
625  *   ENOMEM:	Out of memory, call again later
626  *      EIO:	A permanent error occurred, do not retry
627  *
628  * The RDMA allocate/free functions need the task structure as a place
629  * to hide the struct rpcrdma_req, which is necessary for the actual
630  * send/recv sequence.
631  *
632  * xprt_rdma_allocate provides buffers that are already mapped for
633  * DMA, and a local DMA lkey is provided for each.
634  */
635 static int
xprt_rdma_allocate(struct rpc_task * task)636 xprt_rdma_allocate(struct rpc_task *task)
637 {
638 	struct rpc_rqst *rqst = task->tk_rqstp;
639 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
640 	struct rpcrdma_req *req;
641 	gfp_t flags;
642 
643 	req = rpcrdma_buffer_get(&r_xprt->rx_buf);
644 	if (req == NULL)
645 		return -ENOMEM;
646 
647 	flags = RPCRDMA_DEF_GFP;
648 	if (RPC_IS_SWAPPER(task))
649 		flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
650 
651 	if (!rpcrdma_get_rdmabuf(r_xprt, req, flags))
652 		goto out_fail;
653 	if (!rpcrdma_get_sendbuf(r_xprt, req, rqst->rq_callsize, flags))
654 		goto out_fail;
655 	if (!rpcrdma_get_recvbuf(r_xprt, req, rqst->rq_rcvsize, flags))
656 		goto out_fail;
657 
658 	dprintk("RPC: %5u %s: send size = %zd, recv size = %zd, req = %p\n",
659 		task->tk_pid, __func__, rqst->rq_callsize,
660 		rqst->rq_rcvsize, req);
661 
662 	req->rl_connect_cookie = 0;	/* our reserved value */
663 	rpcrdma_set_xprtdata(rqst, req);
664 	rqst->rq_buffer = req->rl_sendbuf->rg_base;
665 	rqst->rq_rbuffer = req->rl_recvbuf->rg_base;
666 	return 0;
667 
668 out_fail:
669 	rpcrdma_buffer_put(req);
670 	return -ENOMEM;
671 }
672 
673 /**
674  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
675  * @task: RPC task
676  *
677  * Caller guarantees rqst->rq_buffer is non-NULL.
678  */
679 static void
xprt_rdma_free(struct rpc_task * task)680 xprt_rdma_free(struct rpc_task *task)
681 {
682 	struct rpc_rqst *rqst = task->tk_rqstp;
683 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
684 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
685 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
686 
687 	if (req->rl_backchannel)
688 		return;
689 
690 	dprintk("RPC:       %s: called on 0x%p\n", __func__, req->rl_reply);
691 
692 	if (!list_empty(&req->rl_registered))
693 		ia->ri_ops->ro_unmap_sync(r_xprt, &req->rl_registered);
694 	rpcrdma_unmap_sges(ia, req);
695 	rpcrdma_buffer_put(req);
696 }
697 
698 /**
699  * xprt_rdma_send_request - marshal and send an RPC request
700  * @task: RPC task with an RPC message in rq_snd_buf
701  *
702  * Caller holds the transport's write lock.
703  *
704  * Return values:
705  *        0:	The request has been sent
706  * ENOTCONN:	Caller needs to invoke connect logic then call again
707  *  ENOBUFS:	Call again later to send the request
708  *      EIO:	A permanent error occurred. The request was not sent,
709  *		and don't try it again
710  *
711  * send_request invokes the meat of RPC RDMA. It must do the following:
712  *
713  *  1.  Marshal the RPC request into an RPC RDMA request, which means
714  *	putting a header in front of data, and creating IOVs for RDMA
715  *	from those in the request.
716  *  2.  In marshaling, detect opportunities for RDMA, and use them.
717  *  3.  Post a recv message to set up asynch completion, then send
718  *	the request (rpcrdma_ep_post).
719  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
720  */
721 static int
xprt_rdma_send_request(struct rpc_task * task)722 xprt_rdma_send_request(struct rpc_task *task)
723 {
724 	struct rpc_rqst *rqst = task->tk_rqstp;
725 	struct rpc_xprt *xprt = rqst->rq_xprt;
726 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
727 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
728 	int rc = 0;
729 
730 	if (!xprt_connected(xprt))
731 		goto drop_connection;
732 
733 	/* On retransmit, remove any previously registered chunks */
734 	if (unlikely(!list_empty(&req->rl_registered)))
735 		r_xprt->rx_ia.ri_ops->ro_unmap_safe(r_xprt, req, false);
736 
737 	rc = rpcrdma_marshal_req(r_xprt, rqst);
738 	if (rc < 0)
739 		goto failed_marshal;
740 
741 	if (req->rl_reply == NULL) 		/* e.g. reconnection */
742 		rpcrdma_recv_buffer_get(req);
743 
744 	/* Must suppress retransmit to maintain credits */
745 	if (req->rl_connect_cookie == xprt->connect_cookie)
746 		goto drop_connection;
747 	req->rl_connect_cookie = xprt->connect_cookie;
748 
749 	if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
750 		goto drop_connection;
751 
752 	rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
753 	rqst->rq_bytes_sent = 0;
754 	return 0;
755 
756 failed_marshal:
757 	if (rc != -ENOTCONN)
758 		return rc;
759 drop_connection:
760 	xprt_disconnect_done(xprt);
761 	return -ENOTCONN;	/* implies disconnect */
762 }
763 
xprt_rdma_print_stats(struct rpc_xprt * xprt,struct seq_file * seq)764 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
765 {
766 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
767 	long idle_time = 0;
768 
769 	if (xprt_connected(xprt))
770 		idle_time = (long)(jiffies - xprt->last_used) / HZ;
771 
772 	seq_puts(seq, "\txprt:\trdma ");
773 	seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
774 		   0,	/* need a local port? */
775 		   xprt->stat.bind_count,
776 		   xprt->stat.connect_count,
777 		   xprt->stat.connect_time,
778 		   idle_time,
779 		   xprt->stat.sends,
780 		   xprt->stat.recvs,
781 		   xprt->stat.bad_xids,
782 		   xprt->stat.req_u,
783 		   xprt->stat.bklog_u);
784 	seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
785 		   r_xprt->rx_stats.read_chunk_count,
786 		   r_xprt->rx_stats.write_chunk_count,
787 		   r_xprt->rx_stats.reply_chunk_count,
788 		   r_xprt->rx_stats.total_rdma_request,
789 		   r_xprt->rx_stats.total_rdma_reply,
790 		   r_xprt->rx_stats.pullup_copy_count,
791 		   r_xprt->rx_stats.fixup_copy_count,
792 		   r_xprt->rx_stats.hardway_register_count,
793 		   r_xprt->rx_stats.failed_marshal_count,
794 		   r_xprt->rx_stats.bad_reply_count,
795 		   r_xprt->rx_stats.nomsg_call_count);
796 	seq_printf(seq, "%lu %lu %lu %lu\n",
797 		   r_xprt->rx_stats.mrs_recovered,
798 		   r_xprt->rx_stats.mrs_orphaned,
799 		   r_xprt->rx_stats.mrs_allocated,
800 		   r_xprt->rx_stats.local_inv_needed);
801 }
802 
803 static int
xprt_rdma_enable_swap(struct rpc_xprt * xprt)804 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
805 {
806 	return 0;
807 }
808 
809 static void
xprt_rdma_disable_swap(struct rpc_xprt * xprt)810 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
811 {
812 }
813 
814 /*
815  * Plumbing for rpc transport switch and kernel module
816  */
817 
818 static const struct rpc_xprt_ops xprt_rdma_procs = {
819 	.reserve_xprt		= xprt_reserve_xprt_cong,
820 	.release_xprt		= xprt_release_xprt_cong, /* sunrpc/xprt.c */
821 	.alloc_slot		= xprt_alloc_slot,
822 	.release_request	= xprt_release_rqst_cong,       /* ditto */
823 	.set_retrans_timeout	= xprt_set_retrans_timeout_def, /* ditto */
824 	.timer			= xprt_rdma_timer,
825 	.rpcbind		= rpcb_getport_async,	/* sunrpc/rpcb_clnt.c */
826 	.set_port		= xprt_rdma_set_port,
827 	.connect		= xprt_rdma_connect,
828 	.buf_alloc		= xprt_rdma_allocate,
829 	.buf_free		= xprt_rdma_free,
830 	.send_request		= xprt_rdma_send_request,
831 	.close			= xprt_rdma_close,
832 	.destroy		= xprt_rdma_destroy,
833 	.print_stats		= xprt_rdma_print_stats,
834 	.enable_swap		= xprt_rdma_enable_swap,
835 	.disable_swap		= xprt_rdma_disable_swap,
836 	.inject_disconnect	= xprt_rdma_inject_disconnect,
837 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
838 	.bc_setup		= xprt_rdma_bc_setup,
839 	.bc_up			= xprt_rdma_bc_up,
840 	.bc_maxpayload		= xprt_rdma_bc_maxpayload,
841 	.bc_free_rqst		= xprt_rdma_bc_free_rqst,
842 	.bc_destroy		= xprt_rdma_bc_destroy,
843 #endif
844 };
845 
846 static struct xprt_class xprt_rdma = {
847 	.list			= LIST_HEAD_INIT(xprt_rdma.list),
848 	.name			= "rdma",
849 	.owner			= THIS_MODULE,
850 	.ident			= XPRT_TRANSPORT_RDMA,
851 	.setup			= xprt_setup_rdma,
852 };
853 
xprt_rdma_cleanup(void)854 void xprt_rdma_cleanup(void)
855 {
856 	int rc;
857 
858 	dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
859 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
860 	if (sunrpc_table_header) {
861 		unregister_sysctl_table(sunrpc_table_header);
862 		sunrpc_table_header = NULL;
863 	}
864 #endif
865 	rc = xprt_unregister_transport(&xprt_rdma);
866 	if (rc)
867 		dprintk("RPC:       %s: xprt_unregister returned %i\n",
868 			__func__, rc);
869 
870 	rpcrdma_destroy_wq();
871 
872 	rc = xprt_unregister_transport(&xprt_rdma_bc);
873 	if (rc)
874 		dprintk("RPC:       %s: xprt_unregister(bc) returned %i\n",
875 			__func__, rc);
876 }
877 
xprt_rdma_init(void)878 int xprt_rdma_init(void)
879 {
880 	int rc;
881 
882 	rc = rpcrdma_alloc_wq();
883 	if (rc)
884 		return rc;
885 
886 	rc = xprt_register_transport(&xprt_rdma);
887 	if (rc) {
888 		rpcrdma_destroy_wq();
889 		return rc;
890 	}
891 
892 	rc = xprt_register_transport(&xprt_rdma_bc);
893 	if (rc) {
894 		xprt_unregister_transport(&xprt_rdma);
895 		rpcrdma_destroy_wq();
896 		return rc;
897 	}
898 
899 	dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
900 
901 	dprintk("Defaults:\n");
902 	dprintk("\tSlots %d\n"
903 		"\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
904 		xprt_rdma_slot_table_entries,
905 		xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
906 	dprintk("\tPadding %d\n\tMemreg %d\n",
907 		xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
908 
909 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
910 	if (!sunrpc_table_header)
911 		sunrpc_table_header = register_sysctl_table(sunrpc_table);
912 #endif
913 	return 0;
914 }
915