Linux-libre 4.19.20-gnu
[librecmc/linux-libre.git] / net / sunrpc / xprtrdma / transport.c
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
71 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
72 unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
73 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
74 unsigned int xprt_rdma_memreg_strategy          = RPCRDMA_FRWR;
75 int xprt_rdma_pad_optimize;
76
77 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
78
79 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
80 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
81 static unsigned int min_inline_size = RPCRDMA_MIN_INLINE;
82 static unsigned int max_inline_size = RPCRDMA_MAX_INLINE;
83 static unsigned int zero;
84 static unsigned int max_padding = PAGE_SIZE;
85 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
86 static unsigned int max_memreg = RPCRDMA_LAST - 1;
87 static unsigned int dummy;
88
89 static struct ctl_table_header *sunrpc_table_header;
90
91 static struct ctl_table xr_tunables_table[] = {
92         {
93                 .procname       = "rdma_slot_table_entries",
94                 .data           = &xprt_rdma_slot_table_entries,
95                 .maxlen         = sizeof(unsigned int),
96                 .mode           = 0644,
97                 .proc_handler   = proc_dointvec_minmax,
98                 .extra1         = &min_slot_table_size,
99                 .extra2         = &max_slot_table_size
100         },
101         {
102                 .procname       = "rdma_max_inline_read",
103                 .data           = &xprt_rdma_max_inline_read,
104                 .maxlen         = sizeof(unsigned int),
105                 .mode           = 0644,
106                 .proc_handler   = proc_dointvec_minmax,
107                 .extra1         = &min_inline_size,
108                 .extra2         = &max_inline_size,
109         },
110         {
111                 .procname       = "rdma_max_inline_write",
112                 .data           = &xprt_rdma_max_inline_write,
113                 .maxlen         = sizeof(unsigned int),
114                 .mode           = 0644,
115                 .proc_handler   = proc_dointvec_minmax,
116                 .extra1         = &min_inline_size,
117                 .extra2         = &max_inline_size,
118         },
119         {
120                 .procname       = "rdma_inline_write_padding",
121                 .data           = &dummy,
122                 .maxlen         = sizeof(unsigned int),
123                 .mode           = 0644,
124                 .proc_handler   = proc_dointvec_minmax,
125                 .extra1         = &zero,
126                 .extra2         = &max_padding,
127         },
128         {
129                 .procname       = "rdma_memreg_strategy",
130                 .data           = &xprt_rdma_memreg_strategy,
131                 .maxlen         = sizeof(unsigned int),
132                 .mode           = 0644,
133                 .proc_handler   = proc_dointvec_minmax,
134                 .extra1         = &min_memreg,
135                 .extra2         = &max_memreg,
136         },
137         {
138                 .procname       = "rdma_pad_optimize",
139                 .data           = &xprt_rdma_pad_optimize,
140                 .maxlen         = sizeof(unsigned int),
141                 .mode           = 0644,
142                 .proc_handler   = proc_dointvec,
143         },
144         { },
145 };
146
147 static struct ctl_table sunrpc_table[] = {
148         {
149                 .procname       = "sunrpc",
150                 .mode           = 0555,
151                 .child          = xr_tunables_table
152         },
153         { },
154 };
155
156 #endif
157
158 static const struct rpc_xprt_ops xprt_rdma_procs;
159
160 static void
161 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
162 {
163         struct sockaddr_in *sin = (struct sockaddr_in *)sap;
164         char buf[20];
165
166         snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
167         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
168
169         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
170 }
171
172 static void
173 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
174 {
175         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
176         char buf[40];
177
178         snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
179         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
180
181         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
182 }
183
184 void
185 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
186 {
187         char buf[128];
188
189         switch (sap->sa_family) {
190         case AF_INET:
191                 xprt_rdma_format_addresses4(xprt, sap);
192                 break;
193         case AF_INET6:
194                 xprt_rdma_format_addresses6(xprt, sap);
195                 break;
196         default:
197                 pr_err("rpcrdma: Unrecognized address family\n");
198                 return;
199         }
200
201         (void)rpc_ntop(sap, buf, sizeof(buf));
202         xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
203
204         snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
205         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
206
207         snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
208         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
209
210         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
211 }
212
213 void
214 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
215 {
216         unsigned int i;
217
218         for (i = 0; i < RPC_DISPLAY_MAX; i++)
219                 switch (i) {
220                 case RPC_DISPLAY_PROTO:
221                 case RPC_DISPLAY_NETID:
222                         continue;
223                 default:
224                         kfree(xprt->address_strings[i]);
225                 }
226 }
227
228 void
229 rpcrdma_conn_func(struct rpcrdma_ep *ep)
230 {
231         schedule_delayed_work(&ep->rep_connect_worker, 0);
232 }
233
234 void
235 rpcrdma_connect_worker(struct work_struct *work)
236 {
237         struct rpcrdma_ep *ep =
238                 container_of(work, struct rpcrdma_ep, rep_connect_worker.work);
239         struct rpcrdma_xprt *r_xprt =
240                 container_of(ep, struct rpcrdma_xprt, rx_ep);
241         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
242
243         spin_lock_bh(&xprt->transport_lock);
244         if (ep->rep_connected > 0) {
245                 if (!xprt_test_and_set_connected(xprt))
246                         xprt_wake_pending_tasks(xprt, 0);
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
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         rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
265         if (rc)
266                 xprt_wake_pending_tasks(xprt, rc);
267
268         xprt_clear_connecting(xprt);
269 }
270
271 static void
272 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
273 {
274         struct rpcrdma_xprt *r_xprt = container_of(xprt, struct rpcrdma_xprt,
275                                                    rx_xprt);
276
277         trace_xprtrdma_inject_dsc(r_xprt);
278         rdma_disconnect(r_xprt->rx_ia.ri_id);
279 }
280
281 /*
282  * xprt_rdma_destroy
283  *
284  * Destroy the xprt.
285  * Free all memory associated with the object, including its own.
286  * NOTE: none of the *destroy methods free memory for their top-level
287  * objects, even though they may have allocated it (they do free
288  * private memory). It's up to the caller to handle it. In this
289  * case (RDMA transport), all structure memory is inlined with the
290  * struct rpcrdma_xprt.
291  */
292 static void
293 xprt_rdma_destroy(struct rpc_xprt *xprt)
294 {
295         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
296
297         trace_xprtrdma_destroy(r_xprt);
298
299         cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
300
301         xprt_clear_connected(xprt);
302
303         rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
304         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
305         rpcrdma_ia_close(&r_xprt->rx_ia);
306
307         xprt_rdma_free_addresses(xprt);
308         xprt_free(xprt);
309
310         module_put(THIS_MODULE);
311 }
312
313 static const struct rpc_timeout xprt_rdma_default_timeout = {
314         .to_initval = 60 * HZ,
315         .to_maxval = 60 * HZ,
316 };
317
318 /**
319  * xprt_setup_rdma - Set up transport to use RDMA
320  *
321  * @args: rpc transport arguments
322  */
323 static struct rpc_xprt *
324 xprt_setup_rdma(struct xprt_create *args)
325 {
326         struct rpcrdma_create_data_internal cdata;
327         struct rpc_xprt *xprt;
328         struct rpcrdma_xprt *new_xprt;
329         struct rpcrdma_ep *new_ep;
330         struct sockaddr *sap;
331         int rc;
332
333         if (args->addrlen > sizeof(xprt->addr)) {
334                 dprintk("RPC:       %s: address too large\n", __func__);
335                 return ERR_PTR(-EBADF);
336         }
337
338         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt), 0, 0);
339         if (xprt == NULL) {
340                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
341                         __func__);
342                 return ERR_PTR(-ENOMEM);
343         }
344
345         /* 60 second timeout, no retries */
346         xprt->timeout = &xprt_rdma_default_timeout;
347         xprt->bind_timeout = RPCRDMA_BIND_TO;
348         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
349         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
350
351         xprt->resvport = 0;             /* privileged port not needed */
352         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
353         xprt->ops = &xprt_rdma_procs;
354
355         /*
356          * Set up RDMA-specific connect data.
357          */
358         sap = args->dstaddr;
359
360         /* Ensure xprt->addr holds valid server TCP (not RDMA)
361          * address, for any side protocols which peek at it */
362         xprt->prot = IPPROTO_TCP;
363         xprt->addrlen = args->addrlen;
364         memcpy(&xprt->addr, sap, xprt->addrlen);
365
366         if (rpc_get_port(sap))
367                 xprt_set_bound(xprt);
368         xprt_rdma_format_addresses(xprt, sap);
369
370         cdata.max_requests = xprt_rdma_slot_table_entries;
371
372         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
373         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
374
375         cdata.inline_wsize = xprt_rdma_max_inline_write;
376         if (cdata.inline_wsize > cdata.wsize)
377                 cdata.inline_wsize = cdata.wsize;
378
379         cdata.inline_rsize = xprt_rdma_max_inline_read;
380         if (cdata.inline_rsize > cdata.rsize)
381                 cdata.inline_rsize = cdata.rsize;
382
383         /*
384          * Create new transport instance, which includes initialized
385          *  o ia
386          *  o endpoint
387          *  o buffers
388          */
389
390         new_xprt = rpcx_to_rdmax(xprt);
391
392         rc = rpcrdma_ia_open(new_xprt);
393         if (rc)
394                 goto out1;
395
396         /*
397          * initialize and create ep
398          */
399         new_xprt->rx_data = cdata;
400         new_ep = &new_xprt->rx_ep;
401
402         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
403                                 &new_xprt->rx_ia, &new_xprt->rx_data);
404         if (rc)
405                 goto out2;
406
407         rc = rpcrdma_buffer_create(new_xprt);
408         if (rc)
409                 goto out3;
410
411         INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
412                           xprt_rdma_connect_worker);
413
414         xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
415         if (xprt->max_payload == 0)
416                 goto out4;
417         xprt->max_payload <<= PAGE_SHIFT;
418         dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
419                 __func__, xprt->max_payload);
420
421         if (!try_module_get(THIS_MODULE))
422                 goto out4;
423
424         dprintk("RPC:       %s: %s:%s\n", __func__,
425                 xprt->address_strings[RPC_DISPLAY_ADDR],
426                 xprt->address_strings[RPC_DISPLAY_PORT]);
427         trace_xprtrdma_create(new_xprt);
428         return xprt;
429
430 out4:
431         rpcrdma_buffer_destroy(&new_xprt->rx_buf);
432         rc = -ENODEV;
433 out3:
434         rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
435 out2:
436         rpcrdma_ia_close(&new_xprt->rx_ia);
437 out1:
438         trace_xprtrdma_destroy(new_xprt);
439         xprt_rdma_free_addresses(xprt);
440         xprt_free(xprt);
441         return ERR_PTR(rc);
442 }
443
444 /**
445  * xprt_rdma_close - Close down RDMA connection
446  * @xprt: generic transport to be closed
447  *
448  * Called during transport shutdown reconnect, or device
449  * removal. Caller holds the transport's write lock.
450  */
451 static void
452 xprt_rdma_close(struct rpc_xprt *xprt)
453 {
454         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
455         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
456         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
457
458         dprintk("RPC:       %s: closing xprt %p\n", __func__, xprt);
459
460         if (test_and_clear_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags)) {
461                 xprt_clear_connected(xprt);
462                 rpcrdma_ia_remove(ia);
463                 return;
464         }
465         if (ep->rep_connected == -ENODEV)
466                 return;
467         if (ep->rep_connected > 0)
468                 xprt->reestablish_timeout = 0;
469         xprt_disconnect_done(xprt);
470         rpcrdma_ep_disconnect(ep, ia);
471
472         /* Prepare @xprt for the next connection by reinitializing
473          * its credit grant to one (see RFC 8166, Section 3.3.3).
474          */
475         r_xprt->rx_buf.rb_credits = 1;
476         xprt->cwnd = RPC_CWNDSHIFT;
477 }
478
479 /**
480  * xprt_rdma_set_port - update server port with rpcbind result
481  * @xprt: controlling RPC transport
482  * @port: new port value
483  *
484  * Transport connect status is unchanged.
485  */
486 static void
487 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
488 {
489         struct sockaddr *sap = (struct sockaddr *)&xprt->addr;
490         char buf[8];
491
492         dprintk("RPC:       %s: setting port for xprt %p (%s:%s) to %u\n",
493                 __func__, xprt,
494                 xprt->address_strings[RPC_DISPLAY_ADDR],
495                 xprt->address_strings[RPC_DISPLAY_PORT],
496                 port);
497
498         rpc_set_port(sap, port);
499
500         kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
501         snprintf(buf, sizeof(buf), "%u", port);
502         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
503
504         kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]);
505         snprintf(buf, sizeof(buf), "%4hx", port);
506         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
507 }
508
509 /**
510  * xprt_rdma_timer - invoked when an RPC times out
511  * @xprt: controlling RPC transport
512  * @task: RPC task that timed out
513  *
514  * Invoked when the transport is still connected, but an RPC
515  * retransmit timeout occurs.
516  *
517  * Since RDMA connections don't have a keep-alive, forcibly
518  * disconnect and retry to connect. This drives full
519  * detection of the network path, and retransmissions of
520  * all pending RPCs.
521  */
522 static void
523 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
524 {
525         xprt_force_disconnect(xprt);
526 }
527
528 static void
529 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
530 {
531         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
532
533         if (r_xprt->rx_ep.rep_connected != 0) {
534                 /* Reconnect */
535                 schedule_delayed_work(&r_xprt->rx_connect_worker,
536                                       xprt->reestablish_timeout);
537                 xprt->reestablish_timeout <<= 1;
538                 if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
539                         xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
540                 else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
541                         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
542         } else {
543                 schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
544                 if (!RPC_IS_ASYNC(task))
545                         flush_delayed_work(&r_xprt->rx_connect_worker);
546         }
547 }
548
549 /**
550  * xprt_rdma_alloc_slot - allocate an rpc_rqst
551  * @xprt: controlling RPC transport
552  * @task: RPC task requesting a fresh rpc_rqst
553  *
554  * tk_status values:
555  *      %0 if task->tk_rqstp points to a fresh rpc_rqst
556  *      %-EAGAIN if no rpc_rqst is available; queued on backlog
557  */
558 static void
559 xprt_rdma_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
560 {
561         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
562         struct rpcrdma_req *req;
563
564         req = rpcrdma_buffer_get(&r_xprt->rx_buf);
565         if (!req)
566                 goto out_sleep;
567         task->tk_rqstp = &req->rl_slot;
568         task->tk_status = 0;
569         return;
570
571 out_sleep:
572         rpc_sleep_on(&xprt->backlog, task, NULL);
573         task->tk_status = -EAGAIN;
574 }
575
576 /**
577  * xprt_rdma_free_slot - release an rpc_rqst
578  * @xprt: controlling RPC transport
579  * @rqst: rpc_rqst to release
580  *
581  */
582 static void
583 xprt_rdma_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *rqst)
584 {
585         memset(rqst, 0, sizeof(*rqst));
586         rpcrdma_buffer_put(rpcr_to_rdmar(rqst));
587         rpc_wake_up_next(&xprt->backlog);
588 }
589
590 static bool
591 rpcrdma_get_sendbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
592                     size_t size, gfp_t flags)
593 {
594         struct rpcrdma_regbuf *rb;
595
596         if (req->rl_sendbuf && rdmab_length(req->rl_sendbuf) >= size)
597                 return true;
598
599         rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
600         if (IS_ERR(rb))
601                 return false;
602
603         rpcrdma_free_regbuf(req->rl_sendbuf);
604         r_xprt->rx_stats.hardway_register_count += size;
605         req->rl_sendbuf = rb;
606         return true;
607 }
608
609 /* The rq_rcv_buf is used only if a Reply chunk is necessary.
610  * The decision to use a Reply chunk is made later in
611  * rpcrdma_marshal_req. This buffer is registered at that time.
612  *
613  * Otherwise, the associated RPC Reply arrives in a separate
614  * Receive buffer, arbitrarily chosen by the HCA. The buffer
615  * allocated here for the RPC Reply is not utilized in that
616  * case. See rpcrdma_inline_fixup.
617  *
618  * A regbuf is used here to remember the buffer size.
619  */
620 static bool
621 rpcrdma_get_recvbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
622                     size_t size, gfp_t flags)
623 {
624         struct rpcrdma_regbuf *rb;
625
626         if (req->rl_recvbuf && rdmab_length(req->rl_recvbuf) >= size)
627                 return true;
628
629         rb = rpcrdma_alloc_regbuf(size, DMA_NONE, flags);
630         if (IS_ERR(rb))
631                 return false;
632
633         rpcrdma_free_regbuf(req->rl_recvbuf);
634         r_xprt->rx_stats.hardway_register_count += size;
635         req->rl_recvbuf = rb;
636         return true;
637 }
638
639 /**
640  * xprt_rdma_allocate - allocate transport resources for an RPC
641  * @task: RPC task
642  *
643  * Return values:
644  *        0:    Success; rq_buffer points to RPC buffer to use
645  *   ENOMEM:    Out of memory, call again later
646  *      EIO:    A permanent error occurred, do not retry
647  *
648  * The RDMA allocate/free functions need the task structure as a place
649  * to hide the struct rpcrdma_req, which is necessary for the actual
650  * send/recv sequence.
651  *
652  * xprt_rdma_allocate provides buffers that are already mapped for
653  * DMA, and a local DMA lkey is provided for each.
654  */
655 static int
656 xprt_rdma_allocate(struct rpc_task *task)
657 {
658         struct rpc_rqst *rqst = task->tk_rqstp;
659         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
660         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
661         gfp_t flags;
662
663         flags = RPCRDMA_DEF_GFP;
664         if (RPC_IS_SWAPPER(task))
665                 flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
666
667         if (!rpcrdma_get_sendbuf(r_xprt, req, rqst->rq_callsize, flags))
668                 goto out_fail;
669         if (!rpcrdma_get_recvbuf(r_xprt, req, rqst->rq_rcvsize, flags))
670                 goto out_fail;
671
672         rqst->rq_buffer = req->rl_sendbuf->rg_base;
673         rqst->rq_rbuffer = req->rl_recvbuf->rg_base;
674         trace_xprtrdma_allocate(task, req);
675         return 0;
676
677 out_fail:
678         trace_xprtrdma_allocate(task, NULL);
679         return -ENOMEM;
680 }
681
682 /**
683  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
684  * @task: RPC task
685  *
686  * Caller guarantees rqst->rq_buffer is non-NULL.
687  */
688 static void
689 xprt_rdma_free(struct rpc_task *task)
690 {
691         struct rpc_rqst *rqst = task->tk_rqstp;
692         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
693         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
694
695         if (test_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags))
696                 rpcrdma_release_rqst(r_xprt, req);
697         trace_xprtrdma_rpc_done(task, req);
698 }
699
700 /**
701  * xprt_rdma_send_request - marshal and send an RPC request
702  * @task: RPC task with an RPC message in rq_snd_buf
703  *
704  * Caller holds the transport's write lock.
705  *
706  * Returns:
707  *      %0 if the RPC message has been sent
708  *      %-ENOTCONN if the caller should reconnect and call again
709  *      %-EAGAIN if the caller should call again
710  *      %-ENOBUFS if the caller should call again after a delay
711  *      %-EIO if a permanent error occurred and the request was not
712  *              sent. Do not try to send this message again.
713  */
714 static int
715 xprt_rdma_send_request(struct rpc_task *task)
716 {
717         struct rpc_rqst *rqst = task->tk_rqstp;
718         struct rpc_xprt *xprt = rqst->rq_xprt;
719         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
720         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
721         int rc = 0;
722
723 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
724         if (unlikely(!rqst->rq_buffer))
725                 return xprt_rdma_bc_send_reply(rqst);
726 #endif  /* CONFIG_SUNRPC_BACKCHANNEL */
727
728         if (!xprt_connected(xprt))
729                 goto drop_connection;
730
731         rc = rpcrdma_marshal_req(r_xprt, rqst);
732         if (rc < 0)
733                 goto failed_marshal;
734
735         /* Must suppress retransmit to maintain credits */
736         if (rqst->rq_connect_cookie == xprt->connect_cookie)
737                 goto drop_connection;
738         rqst->rq_xtime = ktime_get();
739
740         __set_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags);
741         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
742                 goto drop_connection;
743
744         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
745         rqst->rq_bytes_sent = 0;
746
747         /* An RPC with no reply will throw off credit accounting,
748          * so drop the connection to reset the credit grant.
749          */
750         if (!rpc_reply_expected(task))
751                 goto drop_connection;
752         return 0;
753
754 failed_marshal:
755         if (rc != -ENOTCONN)
756                 return rc;
757 drop_connection:
758         xprt_disconnect_done(xprt);
759         return -ENOTCONN;       /* implies disconnect */
760 }
761
762 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
763 {
764         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
765         long idle_time = 0;
766
767         if (xprt_connected(xprt))
768                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
769
770         seq_puts(seq, "\txprt:\trdma ");
771         seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
772                    0,   /* need a local port? */
773                    xprt->stat.bind_count,
774                    xprt->stat.connect_count,
775                    xprt->stat.connect_time,
776                    idle_time,
777                    xprt->stat.sends,
778                    xprt->stat.recvs,
779                    xprt->stat.bad_xids,
780                    xprt->stat.req_u,
781                    xprt->stat.bklog_u);
782         seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
783                    r_xprt->rx_stats.read_chunk_count,
784                    r_xprt->rx_stats.write_chunk_count,
785                    r_xprt->rx_stats.reply_chunk_count,
786                    r_xprt->rx_stats.total_rdma_request,
787                    r_xprt->rx_stats.total_rdma_reply,
788                    r_xprt->rx_stats.pullup_copy_count,
789                    r_xprt->rx_stats.fixup_copy_count,
790                    r_xprt->rx_stats.hardway_register_count,
791                    r_xprt->rx_stats.failed_marshal_count,
792                    r_xprt->rx_stats.bad_reply_count,
793                    r_xprt->rx_stats.nomsg_call_count);
794         seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n",
795                    r_xprt->rx_stats.mrs_recovered,
796                    r_xprt->rx_stats.mrs_orphaned,
797                    r_xprt->rx_stats.mrs_allocated,
798                    r_xprt->rx_stats.local_inv_needed,
799                    r_xprt->rx_stats.empty_sendctx_q,
800                    r_xprt->rx_stats.reply_waits_for_send);
801 }
802
803 static int
804 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
805 {
806         return 0;
807 }
808
809 static void
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_rdma_alloc_slot,
822         .free_slot              = xprt_rdma_free_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 0\n\tMemreg %d\n", 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 }