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 /*
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 unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
70 unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRWR;
71 int xprt_rdma_pad_optimize;
72 static struct xprt_class xprt_rdma;
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 max_padding = PAGE_SIZE;
81 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
82 static unsigned int max_memreg = RPCRDMA_LAST - 1;
83 static unsigned int dummy;
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 = &dummy,
118 .maxlen = sizeof(unsigned int),
119 .mode = 0644,
120 .proc_handler = proc_dointvec_minmax,
121 .extra1 = SYSCTL_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
xprt_rdma_format_addresses4(struct rpc_xprt * xprt,struct sockaddr * sap)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
xprt_rdma_format_addresses6(struct rpc_xprt * xprt,struct sockaddr * sap)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
xprt_rdma_format_addresses(struct rpc_xprt * xprt,struct sockaddr * sap)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
xprt_rdma_free_addresses(struct rpc_xprt * xprt)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 /**
225 * xprt_rdma_connect_worker - establish connection in the background
226 * @work: worker thread context
227 *
228 * Requester holds the xprt's send lock to prevent activity on this
229 * transport while a fresh connection is being established. RPC tasks
230 * sleep on the xprt's pending queue waiting for connect to complete.
231 */
232 static void
xprt_rdma_connect_worker(struct work_struct * work)233 xprt_rdma_connect_worker(struct work_struct *work)
234 {
235 struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
236 rx_connect_worker.work);
237 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
238 unsigned int pflags = current->flags;
239 int rc;
240
241 if (atomic_read(&xprt->swapper))
242 current->flags |= PF_MEMALLOC;
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 current_restore_flags(pflags, PF_MEMALLOC);
257 }
258
259 /**
260 * xprt_rdma_inject_disconnect - inject a connection fault
261 * @xprt: transport context
262 *
263 * If @xprt is connected, disconnect it to simulate spurious
264 * connection loss. Caller must hold @xprt's send lock to
265 * ensure that data structures and hardware resources are
266 * stable during the rdma_disconnect() call.
267 */
268 static void
xprt_rdma_inject_disconnect(struct rpc_xprt * xprt)269 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
270 {
271 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
272
273 trace_xprtrdma_op_inject_dsc(r_xprt);
274 rdma_disconnect(r_xprt->rx_ep->re_id);
275 }
276
277 /**
278 * xprt_rdma_destroy - Full tear down of transport
279 * @xprt: doomed transport context
280 *
281 * Caller guarantees there will be no more calls to us with
282 * this @xprt.
283 */
284 static void
xprt_rdma_destroy(struct rpc_xprt * xprt)285 xprt_rdma_destroy(struct rpc_xprt *xprt)
286 {
287 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
288
289 cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
290
291 rpcrdma_xprt_disconnect(r_xprt);
292 rpcrdma_buffer_destroy(&r_xprt->rx_buf);
293
294 xprt_rdma_free_addresses(xprt);
295 xprt_free(xprt);
296
297 module_put(THIS_MODULE);
298 }
299
300 /* 60 second timeout, no retries */
301 static const struct rpc_timeout xprt_rdma_default_timeout = {
302 .to_initval = 60 * HZ,
303 .to_maxval = 60 * HZ,
304 };
305
306 /**
307 * xprt_setup_rdma - Set up transport to use RDMA
308 *
309 * @args: rpc transport arguments
310 */
311 static struct rpc_xprt *
xprt_setup_rdma(struct xprt_create * args)312 xprt_setup_rdma(struct xprt_create *args)
313 {
314 struct rpc_xprt *xprt;
315 struct rpcrdma_xprt *new_xprt;
316 struct sockaddr *sap;
317 int rc;
318
319 if (args->addrlen > sizeof(xprt->addr))
320 return ERR_PTR(-EBADF);
321
322 if (!try_module_get(THIS_MODULE))
323 return ERR_PTR(-EIO);
324
325 xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt), 0,
326 xprt_rdma_slot_table_entries);
327 if (!xprt) {
328 module_put(THIS_MODULE);
329 return ERR_PTR(-ENOMEM);
330 }
331
332 xprt->timeout = &xprt_rdma_default_timeout;
333 xprt->connect_timeout = xprt->timeout->to_initval;
334 xprt->max_reconnect_timeout = xprt->timeout->to_maxval;
335 xprt->bind_timeout = RPCRDMA_BIND_TO;
336 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
337 xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
338
339 xprt->resvport = 0; /* privileged port not needed */
340 xprt->ops = &xprt_rdma_procs;
341
342 /*
343 * Set up RDMA-specific connect data.
344 */
345 sap = args->dstaddr;
346
347 /* Ensure xprt->addr holds valid server TCP (not RDMA)
348 * address, for any side protocols which peek at it */
349 xprt->prot = IPPROTO_TCP;
350 xprt->xprt_class = &xprt_rdma;
351 xprt->addrlen = args->addrlen;
352 memcpy(&xprt->addr, sap, xprt->addrlen);
353
354 if (rpc_get_port(sap))
355 xprt_set_bound(xprt);
356 xprt_rdma_format_addresses(xprt, sap);
357
358 new_xprt = rpcx_to_rdmax(xprt);
359 rc = rpcrdma_buffer_create(new_xprt);
360 if (rc) {
361 xprt_rdma_free_addresses(xprt);
362 xprt_free(xprt);
363 module_put(THIS_MODULE);
364 return ERR_PTR(rc);
365 }
366
367 INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
368 xprt_rdma_connect_worker);
369
370 xprt->max_payload = RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT;
371
372 return xprt;
373 }
374
375 /**
376 * xprt_rdma_close - close a transport connection
377 * @xprt: transport context
378 *
379 * Called during autoclose or device removal.
380 *
381 * Caller holds @xprt's send lock to prevent activity on this
382 * transport while the connection is torn down.
383 */
xprt_rdma_close(struct rpc_xprt * xprt)384 void xprt_rdma_close(struct rpc_xprt *xprt)
385 {
386 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
387
388 rpcrdma_xprt_disconnect(r_xprt);
389
390 xprt->reestablish_timeout = 0;
391 ++xprt->connect_cookie;
392 xprt_disconnect_done(xprt);
393 }
394
395 /**
396 * xprt_rdma_set_port - update server port with rpcbind result
397 * @xprt: controlling RPC transport
398 * @port: new port value
399 *
400 * Transport connect status is unchanged.
401 */
402 static void
xprt_rdma_set_port(struct rpc_xprt * xprt,u16 port)403 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
404 {
405 struct sockaddr *sap = (struct sockaddr *)&xprt->addr;
406 char buf[8];
407
408 rpc_set_port(sap, port);
409
410 kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
411 snprintf(buf, sizeof(buf), "%u", port);
412 xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
413
414 kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]);
415 snprintf(buf, sizeof(buf), "%4hx", port);
416 xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
417 }
418
419 /**
420 * xprt_rdma_timer - invoked when an RPC times out
421 * @xprt: controlling RPC transport
422 * @task: RPC task that timed out
423 *
424 * Invoked when the transport is still connected, but an RPC
425 * retransmit timeout occurs.
426 *
427 * Since RDMA connections don't have a keep-alive, forcibly
428 * disconnect and retry to connect. This drives full
429 * detection of the network path, and retransmissions of
430 * all pending RPCs.
431 */
432 static void
xprt_rdma_timer(struct rpc_xprt * xprt,struct rpc_task * task)433 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
434 {
435 xprt_force_disconnect(xprt);
436 }
437
438 /**
439 * xprt_rdma_set_connect_timeout - set timeouts for establishing a connection
440 * @xprt: controlling transport instance
441 * @connect_timeout: reconnect timeout after client disconnects
442 * @reconnect_timeout: reconnect timeout after server disconnects
443 *
444 */
xprt_rdma_set_connect_timeout(struct rpc_xprt * xprt,unsigned long connect_timeout,unsigned long reconnect_timeout)445 static void xprt_rdma_set_connect_timeout(struct rpc_xprt *xprt,
446 unsigned long connect_timeout,
447 unsigned long reconnect_timeout)
448 {
449 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
450
451 trace_xprtrdma_op_set_cto(r_xprt, connect_timeout, reconnect_timeout);
452
453 spin_lock(&xprt->transport_lock);
454
455 if (connect_timeout < xprt->connect_timeout) {
456 struct rpc_timeout to;
457 unsigned long initval;
458
459 to = *xprt->timeout;
460 initval = connect_timeout;
461 if (initval < RPCRDMA_INIT_REEST_TO << 1)
462 initval = RPCRDMA_INIT_REEST_TO << 1;
463 to.to_initval = initval;
464 to.to_maxval = initval;
465 r_xprt->rx_timeout = to;
466 xprt->timeout = &r_xprt->rx_timeout;
467 xprt->connect_timeout = connect_timeout;
468 }
469
470 if (reconnect_timeout < xprt->max_reconnect_timeout)
471 xprt->max_reconnect_timeout = reconnect_timeout;
472
473 spin_unlock(&xprt->transport_lock);
474 }
475
476 /**
477 * xprt_rdma_connect - schedule an attempt to reconnect
478 * @xprt: transport state
479 * @task: RPC scheduler context (unused)
480 *
481 */
482 static void
xprt_rdma_connect(struct rpc_xprt * xprt,struct rpc_task * task)483 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
484 {
485 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
486 struct rpcrdma_ep *ep = r_xprt->rx_ep;
487 unsigned long delay;
488
489 WARN_ON_ONCE(!xprt_lock_connect(xprt, task, r_xprt));
490
491 delay = 0;
492 if (ep && ep->re_connect_status != 0) {
493 delay = xprt_reconnect_delay(xprt);
494 xprt_reconnect_backoff(xprt, RPCRDMA_INIT_REEST_TO);
495 }
496 trace_xprtrdma_op_connect(r_xprt, delay);
497 queue_delayed_work(system_long_wq, &r_xprt->rx_connect_worker, delay);
498 }
499
500 /**
501 * xprt_rdma_alloc_slot - allocate an rpc_rqst
502 * @xprt: controlling RPC transport
503 * @task: RPC task requesting a fresh rpc_rqst
504 *
505 * tk_status values:
506 * %0 if task->tk_rqstp points to a fresh rpc_rqst
507 * %-EAGAIN if no rpc_rqst is available; queued on backlog
508 */
509 static void
xprt_rdma_alloc_slot(struct rpc_xprt * xprt,struct rpc_task * task)510 xprt_rdma_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
511 {
512 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
513 struct rpcrdma_req *req;
514
515 req = rpcrdma_buffer_get(&r_xprt->rx_buf);
516 if (!req)
517 goto out_sleep;
518 task->tk_rqstp = &req->rl_slot;
519 task->tk_status = 0;
520 return;
521
522 out_sleep:
523 task->tk_status = -ENOMEM;
524 xprt_add_backlog(xprt, task);
525 }
526
527 /**
528 * xprt_rdma_free_slot - release an rpc_rqst
529 * @xprt: controlling RPC transport
530 * @rqst: rpc_rqst to release
531 *
532 */
533 static void
xprt_rdma_free_slot(struct rpc_xprt * xprt,struct rpc_rqst * rqst)534 xprt_rdma_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *rqst)
535 {
536 struct rpcrdma_xprt *r_xprt =
537 container_of(xprt, struct rpcrdma_xprt, rx_xprt);
538
539 rpcrdma_reply_put(&r_xprt->rx_buf, rpcr_to_rdmar(rqst));
540 if (!xprt_wake_up_backlog(xprt, rqst)) {
541 memset(rqst, 0, sizeof(*rqst));
542 rpcrdma_buffer_put(&r_xprt->rx_buf, rpcr_to_rdmar(rqst));
543 }
544 }
545
rpcrdma_check_regbuf(struct rpcrdma_xprt * r_xprt,struct rpcrdma_regbuf * rb,size_t size,gfp_t flags)546 static bool rpcrdma_check_regbuf(struct rpcrdma_xprt *r_xprt,
547 struct rpcrdma_regbuf *rb, size_t size,
548 gfp_t flags)
549 {
550 if (unlikely(rdmab_length(rb) < size)) {
551 if (!rpcrdma_regbuf_realloc(rb, size, flags))
552 return false;
553 r_xprt->rx_stats.hardway_register_count += size;
554 }
555 return true;
556 }
557
558 /**
559 * xprt_rdma_allocate - allocate transport resources for an RPC
560 * @task: RPC task
561 *
562 * Return values:
563 * 0: Success; rq_buffer points to RPC buffer to use
564 * ENOMEM: Out of memory, call again later
565 * EIO: A permanent error occurred, do not retry
566 */
567 static int
xprt_rdma_allocate(struct rpc_task * task)568 xprt_rdma_allocate(struct rpc_task *task)
569 {
570 struct rpc_rqst *rqst = task->tk_rqstp;
571 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
572 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
573 gfp_t flags = rpc_task_gfp_mask();
574
575 if (!rpcrdma_check_regbuf(r_xprt, req->rl_sendbuf, rqst->rq_callsize,
576 flags))
577 goto out_fail;
578 if (!rpcrdma_check_regbuf(r_xprt, req->rl_recvbuf, rqst->rq_rcvsize,
579 flags))
580 goto out_fail;
581
582 rqst->rq_buffer = rdmab_data(req->rl_sendbuf);
583 rqst->rq_rbuffer = rdmab_data(req->rl_recvbuf);
584 return 0;
585
586 out_fail:
587 return -ENOMEM;
588 }
589
590 /**
591 * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
592 * @task: RPC task
593 *
594 * Caller guarantees rqst->rq_buffer is non-NULL.
595 */
596 static void
xprt_rdma_free(struct rpc_task * task)597 xprt_rdma_free(struct rpc_task *task)
598 {
599 struct rpc_rqst *rqst = task->tk_rqstp;
600 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
601
602 if (unlikely(!list_empty(&req->rl_registered))) {
603 trace_xprtrdma_mrs_zap(task);
604 frwr_unmap_sync(rpcx_to_rdmax(rqst->rq_xprt), req);
605 }
606
607 /* XXX: If the RPC is completing because of a signal and
608 * not because a reply was received, we ought to ensure
609 * that the Send completion has fired, so that memory
610 * involved with the Send is not still visible to the NIC.
611 */
612 }
613
614 /**
615 * xprt_rdma_send_request - marshal and send an RPC request
616 * @rqst: RPC message in rq_snd_buf
617 *
618 * Caller holds the transport's write lock.
619 *
620 * Returns:
621 * %0 if the RPC message has been sent
622 * %-ENOTCONN if the caller should reconnect and call again
623 * %-EAGAIN if the caller should call again
624 * %-ENOBUFS if the caller should call again after a delay
625 * %-EMSGSIZE if encoding ran out of buffer space. The request
626 * was not sent. Do not try to send this message again.
627 * %-EIO if an I/O error occurred. The request was not sent.
628 * Do not try to send this message again.
629 */
630 static int
xprt_rdma_send_request(struct rpc_rqst * rqst)631 xprt_rdma_send_request(struct rpc_rqst *rqst)
632 {
633 struct rpc_xprt *xprt = rqst->rq_xprt;
634 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
635 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
636 int rc = 0;
637
638 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
639 if (unlikely(!rqst->rq_buffer))
640 return xprt_rdma_bc_send_reply(rqst);
641 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
642
643 if (!xprt_connected(xprt))
644 return -ENOTCONN;
645
646 if (!xprt_request_get_cong(xprt, rqst))
647 return -EBADSLT;
648
649 rc = rpcrdma_marshal_req(r_xprt, rqst);
650 if (rc < 0)
651 goto failed_marshal;
652
653 /* Must suppress retransmit to maintain credits */
654 if (rqst->rq_connect_cookie == xprt->connect_cookie)
655 goto drop_connection;
656 rqst->rq_xtime = ktime_get();
657
658 if (frwr_send(r_xprt, req))
659 goto drop_connection;
660
661 rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
662
663 /* An RPC with no reply will throw off credit accounting,
664 * so drop the connection to reset the credit grant.
665 */
666 if (!rpc_reply_expected(rqst->rq_task))
667 goto drop_connection;
668 return 0;
669
670 failed_marshal:
671 if (rc != -ENOTCONN)
672 return rc;
673 drop_connection:
674 xprt_rdma_close(xprt);
675 return -ENOTCONN;
676 }
677
xprt_rdma_print_stats(struct rpc_xprt * xprt,struct seq_file * seq)678 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
679 {
680 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
681 long idle_time = 0;
682
683 if (xprt_connected(xprt))
684 idle_time = (long)(jiffies - xprt->last_used) / HZ;
685
686 seq_puts(seq, "\txprt:\trdma ");
687 seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
688 0, /* need a local port? */
689 xprt->stat.bind_count,
690 xprt->stat.connect_count,
691 xprt->stat.connect_time / HZ,
692 idle_time,
693 xprt->stat.sends,
694 xprt->stat.recvs,
695 xprt->stat.bad_xids,
696 xprt->stat.req_u,
697 xprt->stat.bklog_u);
698 seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
699 r_xprt->rx_stats.read_chunk_count,
700 r_xprt->rx_stats.write_chunk_count,
701 r_xprt->rx_stats.reply_chunk_count,
702 r_xprt->rx_stats.total_rdma_request,
703 r_xprt->rx_stats.total_rdma_reply,
704 r_xprt->rx_stats.pullup_copy_count,
705 r_xprt->rx_stats.fixup_copy_count,
706 r_xprt->rx_stats.hardway_register_count,
707 r_xprt->rx_stats.failed_marshal_count,
708 r_xprt->rx_stats.bad_reply_count,
709 r_xprt->rx_stats.nomsg_call_count);
710 seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n",
711 r_xprt->rx_stats.mrs_recycled,
712 r_xprt->rx_stats.mrs_orphaned,
713 r_xprt->rx_stats.mrs_allocated,
714 r_xprt->rx_stats.local_inv_needed,
715 r_xprt->rx_stats.empty_sendctx_q,
716 r_xprt->rx_stats.reply_waits_for_send);
717 }
718
719 static int
xprt_rdma_enable_swap(struct rpc_xprt * xprt)720 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
721 {
722 return 0;
723 }
724
725 static void
xprt_rdma_disable_swap(struct rpc_xprt * xprt)726 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
727 {
728 }
729
730 /*
731 * Plumbing for rpc transport switch and kernel module
732 */
733
734 static const struct rpc_xprt_ops xprt_rdma_procs = {
735 .reserve_xprt = xprt_reserve_xprt_cong,
736 .release_xprt = xprt_release_xprt_cong, /* sunrpc/xprt.c */
737 .alloc_slot = xprt_rdma_alloc_slot,
738 .free_slot = xprt_rdma_free_slot,
739 .release_request = xprt_release_rqst_cong, /* ditto */
740 .wait_for_reply_request = xprt_wait_for_reply_request_def, /* ditto */
741 .timer = xprt_rdma_timer,
742 .rpcbind = rpcb_getport_async, /* sunrpc/rpcb_clnt.c */
743 .set_port = xprt_rdma_set_port,
744 .connect = xprt_rdma_connect,
745 .buf_alloc = xprt_rdma_allocate,
746 .buf_free = xprt_rdma_free,
747 .send_request = xprt_rdma_send_request,
748 .close = xprt_rdma_close,
749 .destroy = xprt_rdma_destroy,
750 .set_connect_timeout = xprt_rdma_set_connect_timeout,
751 .print_stats = xprt_rdma_print_stats,
752 .enable_swap = xprt_rdma_enable_swap,
753 .disable_swap = xprt_rdma_disable_swap,
754 .inject_disconnect = xprt_rdma_inject_disconnect,
755 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
756 .bc_setup = xprt_rdma_bc_setup,
757 .bc_maxpayload = xprt_rdma_bc_maxpayload,
758 .bc_num_slots = xprt_rdma_bc_max_slots,
759 .bc_free_rqst = xprt_rdma_bc_free_rqst,
760 .bc_destroy = xprt_rdma_bc_destroy,
761 #endif
762 };
763
764 static struct xprt_class xprt_rdma = {
765 .list = LIST_HEAD_INIT(xprt_rdma.list),
766 .name = "rdma",
767 .owner = THIS_MODULE,
768 .ident = XPRT_TRANSPORT_RDMA,
769 .setup = xprt_setup_rdma,
770 .netid = { "rdma", "rdma6", "" },
771 };
772
xprt_rdma_cleanup(void)773 void xprt_rdma_cleanup(void)
774 {
775 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
776 if (sunrpc_table_header) {
777 unregister_sysctl_table(sunrpc_table_header);
778 sunrpc_table_header = NULL;
779 }
780 #endif
781
782 xprt_unregister_transport(&xprt_rdma);
783 xprt_unregister_transport(&xprt_rdma_bc);
784 }
785
xprt_rdma_init(void)786 int xprt_rdma_init(void)
787 {
788 int rc;
789
790 rc = xprt_register_transport(&xprt_rdma);
791 if (rc)
792 return rc;
793
794 rc = xprt_register_transport(&xprt_rdma_bc);
795 if (rc) {
796 xprt_unregister_transport(&xprt_rdma);
797 return rc;
798 }
799
800 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
801 if (!sunrpc_table_header)
802 sunrpc_table_header = register_sysctl_table(sunrpc_table);
803 #endif
804 return 0;
805 }
806