b1b009f10ea375a3f63148973e58e486df1aa282
[cascardo/linux.git] / net / sunrpc / xprtrdma / transport.c
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 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
70                 int xprt_rdma_pad_optimize = 1;
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 zero;
77 static unsigned int max_padding = PAGE_SIZE;
78 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
79 static unsigned int max_memreg = RPCRDMA_LAST - 1;
80
81 static struct ctl_table_header *sunrpc_table_header;
82
83 static struct ctl_table xr_tunables_table[] = {
84         {
85                 .procname       = "rdma_slot_table_entries",
86                 .data           = &xprt_rdma_slot_table_entries,
87                 .maxlen         = sizeof(unsigned int),
88                 .mode           = 0644,
89                 .proc_handler   = proc_dointvec_minmax,
90                 .extra1         = &min_slot_table_size,
91                 .extra2         = &max_slot_table_size
92         },
93         {
94                 .procname       = "rdma_max_inline_read",
95                 .data           = &xprt_rdma_max_inline_read,
96                 .maxlen         = sizeof(unsigned int),
97                 .mode           = 0644,
98                 .proc_handler   = proc_dointvec,
99         },
100         {
101                 .procname       = "rdma_max_inline_write",
102                 .data           = &xprt_rdma_max_inline_write,
103                 .maxlen         = sizeof(unsigned int),
104                 .mode           = 0644,
105                 .proc_handler   = proc_dointvec,
106         },
107         {
108                 .procname       = "rdma_inline_write_padding",
109                 .data           = &xprt_rdma_inline_write_padding,
110                 .maxlen         = sizeof(unsigned int),
111                 .mode           = 0644,
112                 .proc_handler   = proc_dointvec_minmax,
113                 .extra1         = &zero,
114                 .extra2         = &max_padding,
115         },
116         {
117                 .procname       = "rdma_memreg_strategy",
118                 .data           = &xprt_rdma_memreg_strategy,
119                 .maxlen         = sizeof(unsigned int),
120                 .mode           = 0644,
121                 .proc_handler   = proc_dointvec_minmax,
122                 .extra1         = &min_memreg,
123                 .extra2         = &max_memreg,
124         },
125         {
126                 .procname       = "rdma_pad_optimize",
127                 .data           = &xprt_rdma_pad_optimize,
128                 .maxlen         = sizeof(unsigned int),
129                 .mode           = 0644,
130                 .proc_handler   = proc_dointvec,
131         },
132         { },
133 };
134
135 static struct ctl_table sunrpc_table[] = {
136         {
137                 .procname       = "sunrpc",
138                 .mode           = 0555,
139                 .child          = xr_tunables_table
140         },
141         { },
142 };
143
144 #endif
145
146 static struct rpc_xprt_ops xprt_rdma_procs;     /*forward reference */
147
148 static void
149 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
150 {
151         struct sockaddr_in *sin = (struct sockaddr_in *)sap;
152         char buf[20];
153
154         snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
155         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
156
157         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
158 }
159
160 static void
161 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
162 {
163         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
164         char buf[40];
165
166         snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
167         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
168
169         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
170 }
171
172 void
173 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
174 {
175         char buf[128];
176
177         switch (sap->sa_family) {
178         case AF_INET:
179                 xprt_rdma_format_addresses4(xprt, sap);
180                 break;
181         case AF_INET6:
182                 xprt_rdma_format_addresses6(xprt, sap);
183                 break;
184         default:
185                 pr_err("rpcrdma: Unrecognized address family\n");
186                 return;
187         }
188
189         (void)rpc_ntop(sap, buf, sizeof(buf));
190         xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
191
192         snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
193         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
194
195         snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
196         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
197
198         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
199 }
200
201 void
202 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
203 {
204         unsigned int i;
205
206         for (i = 0; i < RPC_DISPLAY_MAX; i++)
207                 switch (i) {
208                 case RPC_DISPLAY_PROTO:
209                 case RPC_DISPLAY_NETID:
210                         continue;
211                 default:
212                         kfree(xprt->address_strings[i]);
213                 }
214 }
215
216 static void
217 xprt_rdma_connect_worker(struct work_struct *work)
218 {
219         struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
220                                                    rx_connect_worker.work);
221         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
222         int rc = 0;
223
224         xprt_clear_connected(xprt);
225
226         dprintk("RPC:       %s: %sconnect\n", __func__,
227                         r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
228         rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
229         if (rc)
230                 xprt_wake_pending_tasks(xprt, rc);
231
232         dprintk("RPC:       %s: exit\n", __func__);
233         xprt_clear_connecting(xprt);
234 }
235
236 static void
237 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
238 {
239         struct rpcrdma_xprt *r_xprt = container_of(xprt, struct rpcrdma_xprt,
240                                                    rx_xprt);
241
242         pr_info("rpcrdma: injecting transport disconnect on xprt=%p\n", xprt);
243         rdma_disconnect(r_xprt->rx_ia.ri_id);
244 }
245
246 /*
247  * xprt_rdma_destroy
248  *
249  * Destroy the xprt.
250  * Free all memory associated with the object, including its own.
251  * NOTE: none of the *destroy methods free memory for their top-level
252  * objects, even though they may have allocated it (they do free
253  * private memory). It's up to the caller to handle it. In this
254  * case (RDMA transport), all structure memory is inlined with the
255  * struct rpcrdma_xprt.
256  */
257 static void
258 xprt_rdma_destroy(struct rpc_xprt *xprt)
259 {
260         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
261
262         dprintk("RPC:       %s: called\n", __func__);
263
264         cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
265
266         xprt_clear_connected(xprt);
267
268         rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
269         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
270         rpcrdma_ia_close(&r_xprt->rx_ia);
271
272         xprt_rdma_free_addresses(xprt);
273
274         xprt_free(xprt);
275
276         dprintk("RPC:       %s: returning\n", __func__);
277
278         module_put(THIS_MODULE);
279 }
280
281 static const struct rpc_timeout xprt_rdma_default_timeout = {
282         .to_initval = 60 * HZ,
283         .to_maxval = 60 * HZ,
284 };
285
286 /**
287  * xprt_setup_rdma - Set up transport to use RDMA
288  *
289  * @args: rpc transport arguments
290  */
291 static struct rpc_xprt *
292 xprt_setup_rdma(struct xprt_create *args)
293 {
294         struct rpcrdma_create_data_internal cdata;
295         struct rpc_xprt *xprt;
296         struct rpcrdma_xprt *new_xprt;
297         struct rpcrdma_ep *new_ep;
298         struct sockaddr *sap;
299         int rc;
300
301         if (args->addrlen > sizeof(xprt->addr)) {
302                 dprintk("RPC:       %s: address too large\n", __func__);
303                 return ERR_PTR(-EBADF);
304         }
305
306         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
307                         xprt_rdma_slot_table_entries,
308                         xprt_rdma_slot_table_entries);
309         if (xprt == NULL) {
310                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
311                         __func__);
312                 return ERR_PTR(-ENOMEM);
313         }
314
315         /* 60 second timeout, no retries */
316         xprt->timeout = &xprt_rdma_default_timeout;
317         xprt->bind_timeout = RPCRDMA_BIND_TO;
318         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
319         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
320
321         xprt->resvport = 0;             /* privileged port not needed */
322         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
323         xprt->ops = &xprt_rdma_procs;
324
325         /*
326          * Set up RDMA-specific connect data.
327          */
328
329         sap = (struct sockaddr *)&cdata.addr;
330         memcpy(sap, args->dstaddr, args->addrlen);
331
332         /* Ensure xprt->addr holds valid server TCP (not RDMA)
333          * address, for any side protocols which peek at it */
334         xprt->prot = IPPROTO_TCP;
335         xprt->addrlen = args->addrlen;
336         memcpy(&xprt->addr, sap, xprt->addrlen);
337
338         if (rpc_get_port(sap))
339                 xprt_set_bound(xprt);
340
341         cdata.max_requests = xprt->max_reqs;
342
343         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
344         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
345
346         cdata.inline_wsize = xprt_rdma_max_inline_write;
347         if (cdata.inline_wsize > cdata.wsize)
348                 cdata.inline_wsize = cdata.wsize;
349
350         cdata.inline_rsize = xprt_rdma_max_inline_read;
351         if (cdata.inline_rsize > cdata.rsize)
352                 cdata.inline_rsize = cdata.rsize;
353
354         cdata.padding = xprt_rdma_inline_write_padding;
355
356         /*
357          * Create new transport instance, which includes initialized
358          *  o ia
359          *  o endpoint
360          *  o buffers
361          */
362
363         new_xprt = rpcx_to_rdmax(xprt);
364
365         rc = rpcrdma_ia_open(new_xprt, sap, xprt_rdma_memreg_strategy);
366         if (rc)
367                 goto out1;
368
369         /*
370          * initialize and create ep
371          */
372         new_xprt->rx_data = cdata;
373         new_ep = &new_xprt->rx_ep;
374         new_ep->rep_remote_addr = cdata.addr;
375
376         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
377                                 &new_xprt->rx_ia, &new_xprt->rx_data);
378         if (rc)
379                 goto out2;
380
381         /*
382          * Allocate pre-registered send and receive buffers for headers and
383          * any inline data. Also specify any padding which will be provided
384          * from a preregistered zero buffer.
385          */
386         rc = rpcrdma_buffer_create(new_xprt);
387         if (rc)
388                 goto out3;
389
390         /*
391          * Register a callback for connection events. This is necessary because
392          * connection loss notification is async. We also catch connection loss
393          * when reaping receives.
394          */
395         INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
396                           xprt_rdma_connect_worker);
397
398         xprt_rdma_format_addresses(xprt, sap);
399         xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
400         if (xprt->max_payload == 0)
401                 goto out4;
402         xprt->max_payload <<= PAGE_SHIFT;
403         dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
404                 __func__, xprt->max_payload);
405
406         if (!try_module_get(THIS_MODULE))
407                 goto out4;
408
409         dprintk("RPC:       %s: %s:%s\n", __func__,
410                 xprt->address_strings[RPC_DISPLAY_ADDR],
411                 xprt->address_strings[RPC_DISPLAY_PORT]);
412         return xprt;
413
414 out4:
415         xprt_rdma_free_addresses(xprt);
416         rc = -EINVAL;
417 out3:
418         rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
419 out2:
420         rpcrdma_ia_close(&new_xprt->rx_ia);
421 out1:
422         xprt_free(xprt);
423         return ERR_PTR(rc);
424 }
425
426 /*
427  * Close a connection, during shutdown or timeout/reconnect
428  */
429 static void
430 xprt_rdma_close(struct rpc_xprt *xprt)
431 {
432         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
433
434         dprintk("RPC:       %s: closing\n", __func__);
435         if (r_xprt->rx_ep.rep_connected > 0)
436                 xprt->reestablish_timeout = 0;
437         xprt_disconnect_done(xprt);
438         rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
439 }
440
441 static void
442 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
443 {
444         struct sockaddr_in *sap;
445
446         sap = (struct sockaddr_in *)&xprt->addr;
447         sap->sin_port = htons(port);
448         sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
449         sap->sin_port = htons(port);
450         dprintk("RPC:       %s: %u\n", __func__, port);
451 }
452
453 static void
454 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
455 {
456         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
457
458         if (r_xprt->rx_ep.rep_connected != 0) {
459                 /* Reconnect */
460                 schedule_delayed_work(&r_xprt->rx_connect_worker,
461                                       xprt->reestablish_timeout);
462                 xprt->reestablish_timeout <<= 1;
463                 if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
464                         xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
465                 else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
466                         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
467         } else {
468                 schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
469                 if (!RPC_IS_ASYNC(task))
470                         flush_delayed_work(&r_xprt->rx_connect_worker);
471         }
472 }
473
474 /*
475  * The RDMA allocate/free functions need the task structure as a place
476  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
477  * sequence.
478  *
479  * The RPC layer allocates both send and receive buffers in the same call
480  * (rq_send_buf and rq_rcv_buf are both part of a single contiguous buffer).
481  * We may register rq_rcv_buf when using reply chunks.
482  */
483 static void *
484 xprt_rdma_allocate(struct rpc_task *task, size_t size)
485 {
486         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
487         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
488         struct rpcrdma_regbuf *rb;
489         struct rpcrdma_req *req;
490         size_t min_size;
491         gfp_t flags;
492
493         req = rpcrdma_buffer_get(&r_xprt->rx_buf);
494         if (req == NULL)
495                 return NULL;
496
497         flags = RPCRDMA_DEF_GFP;
498         if (RPC_IS_SWAPPER(task))
499                 flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
500
501         if (req->rl_rdmabuf == NULL)
502                 goto out_rdmabuf;
503         if (req->rl_sendbuf == NULL)
504                 goto out_sendbuf;
505         if (size > req->rl_sendbuf->rg_size)
506                 goto out_sendbuf;
507
508 out:
509         dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
510         req->rl_connect_cookie = 0;     /* our reserved value */
511         return req->rl_sendbuf->rg_base;
512
513 out_rdmabuf:
514         min_size = RPCRDMA_INLINE_WRITE_THRESHOLD(task->tk_rqstp);
515         rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, min_size, flags);
516         if (IS_ERR(rb))
517                 goto out_fail;
518         req->rl_rdmabuf = rb;
519
520 out_sendbuf:
521         /* XDR encoding and RPC/RDMA marshaling of this request has not
522          * yet occurred. Thus a lower bound is needed to prevent buffer
523          * overrun during marshaling.
524          *
525          * RPC/RDMA marshaling may choose to send payload bearing ops
526          * inline, if the result is smaller than the inline threshold.
527          * The value of the "size" argument accounts for header
528          * requirements but not for the payload in these cases.
529          *
530          * Likewise, allocate enough space to receive a reply up to the
531          * size of the inline threshold.
532          *
533          * It's unlikely that both the send header and the received
534          * reply will be large, but slush is provided here to allow
535          * flexibility when marshaling.
536          */
537         min_size = RPCRDMA_INLINE_READ_THRESHOLD(task->tk_rqstp);
538         min_size += RPCRDMA_INLINE_WRITE_THRESHOLD(task->tk_rqstp);
539         if (size < min_size)
540                 size = min_size;
541
542         rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, size, flags);
543         if (IS_ERR(rb))
544                 goto out_fail;
545         rb->rg_owner = req;
546
547         r_xprt->rx_stats.hardway_register_count += size;
548         rpcrdma_free_regbuf(&r_xprt->rx_ia, req->rl_sendbuf);
549         req->rl_sendbuf = rb;
550         goto out;
551
552 out_fail:
553         rpcrdma_buffer_put(req);
554         r_xprt->rx_stats.failed_marshal_count++;
555         return NULL;
556 }
557
558 /*
559  * This function returns all RDMA resources to the pool.
560  */
561 static void
562 xprt_rdma_free(void *buffer)
563 {
564         struct rpcrdma_req *req;
565         struct rpcrdma_xprt *r_xprt;
566         struct rpcrdma_regbuf *rb;
567         int i;
568
569         if (buffer == NULL)
570                 return;
571
572         rb = container_of(buffer, struct rpcrdma_regbuf, rg_base[0]);
573         req = rb->rg_owner;
574         if (req->rl_backchannel)
575                 return;
576
577         r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
578
579         dprintk("RPC:       %s: called on 0x%p\n", __func__, req->rl_reply);
580
581         for (i = 0; req->rl_nchunks;) {
582                 --req->rl_nchunks;
583                 i += r_xprt->rx_ia.ri_ops->ro_unmap(r_xprt,
584                                                     &req->rl_segments[i]);
585         }
586
587         rpcrdma_buffer_put(req);
588 }
589
590 /*
591  * send_request invokes the meat of RPC RDMA. It must do the following:
592  *  1.  Marshal the RPC request into an RPC RDMA request, which means
593  *      putting a header in front of data, and creating IOVs for RDMA
594  *      from those in the request.
595  *  2.  In marshaling, detect opportunities for RDMA, and use them.
596  *  3.  Post a recv message to set up asynch completion, then send
597  *      the request (rpcrdma_ep_post).
598  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
599  */
600
601 static int
602 xprt_rdma_send_request(struct rpc_task *task)
603 {
604         struct rpc_rqst *rqst = task->tk_rqstp;
605         struct rpc_xprt *xprt = rqst->rq_xprt;
606         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
607         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
608         int rc = 0;
609
610         rc = rpcrdma_marshal_req(rqst);
611         if (rc < 0)
612                 goto failed_marshal;
613
614         if (req->rl_reply == NULL)              /* e.g. reconnection */
615                 rpcrdma_recv_buffer_get(req);
616
617         /* Must suppress retransmit to maintain credits */
618         if (req->rl_connect_cookie == xprt->connect_cookie)
619                 goto drop_connection;
620         req->rl_connect_cookie = xprt->connect_cookie;
621
622         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
623                 goto drop_connection;
624
625         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
626         rqst->rq_bytes_sent = 0;
627         return 0;
628
629 failed_marshal:
630         r_xprt->rx_stats.failed_marshal_count++;
631         dprintk("RPC:       %s: rpcrdma_marshal_req failed, status %i\n",
632                 __func__, rc);
633         if (rc == -EIO)
634                 return -EIO;
635 drop_connection:
636         xprt_disconnect_done(xprt);
637         return -ENOTCONN;       /* implies disconnect */
638 }
639
640 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
641 {
642         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
643         long idle_time = 0;
644
645         if (xprt_connected(xprt))
646                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
647
648         seq_puts(seq, "\txprt:\trdma ");
649         seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
650                    0,   /* need a local port? */
651                    xprt->stat.bind_count,
652                    xprt->stat.connect_count,
653                    xprt->stat.connect_time,
654                    idle_time,
655                    xprt->stat.sends,
656                    xprt->stat.recvs,
657                    xprt->stat.bad_xids,
658                    xprt->stat.req_u,
659                    xprt->stat.bklog_u);
660         seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu\n",
661                    r_xprt->rx_stats.read_chunk_count,
662                    r_xprt->rx_stats.write_chunk_count,
663                    r_xprt->rx_stats.reply_chunk_count,
664                    r_xprt->rx_stats.total_rdma_request,
665                    r_xprt->rx_stats.total_rdma_reply,
666                    r_xprt->rx_stats.pullup_copy_count,
667                    r_xprt->rx_stats.fixup_copy_count,
668                    r_xprt->rx_stats.hardway_register_count,
669                    r_xprt->rx_stats.failed_marshal_count,
670                    r_xprt->rx_stats.bad_reply_count,
671                    r_xprt->rx_stats.nomsg_call_count);
672 }
673
674 static int
675 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
676 {
677         return 0;
678 }
679
680 static void
681 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
682 {
683 }
684
685 /*
686  * Plumbing for rpc transport switch and kernel module
687  */
688
689 static struct rpc_xprt_ops xprt_rdma_procs = {
690         .reserve_xprt           = xprt_reserve_xprt_cong,
691         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
692         .alloc_slot             = xprt_alloc_slot,
693         .release_request        = xprt_release_rqst_cong,       /* ditto */
694         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
695         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
696         .set_port               = xprt_rdma_set_port,
697         .connect                = xprt_rdma_connect,
698         .buf_alloc              = xprt_rdma_allocate,
699         .buf_free               = xprt_rdma_free,
700         .send_request           = xprt_rdma_send_request,
701         .close                  = xprt_rdma_close,
702         .destroy                = xprt_rdma_destroy,
703         .print_stats            = xprt_rdma_print_stats,
704         .enable_swap            = xprt_rdma_enable_swap,
705         .disable_swap           = xprt_rdma_disable_swap,
706         .inject_disconnect      = xprt_rdma_inject_disconnect,
707 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
708         .bc_setup               = xprt_rdma_bc_setup,
709         .bc_up                  = xprt_rdma_bc_up,
710         .bc_free_rqst           = xprt_rdma_bc_free_rqst,
711         .bc_destroy             = xprt_rdma_bc_destroy,
712 #endif
713 };
714
715 static struct xprt_class xprt_rdma = {
716         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
717         .name                   = "rdma",
718         .owner                  = THIS_MODULE,
719         .ident                  = XPRT_TRANSPORT_RDMA,
720         .setup                  = xprt_setup_rdma,
721 };
722
723 void xprt_rdma_cleanup(void)
724 {
725         int rc;
726
727         dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
728 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
729         if (sunrpc_table_header) {
730                 unregister_sysctl_table(sunrpc_table_header);
731                 sunrpc_table_header = NULL;
732         }
733 #endif
734         rc = xprt_unregister_transport(&xprt_rdma);
735         if (rc)
736                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
737                         __func__, rc);
738
739         rpcrdma_destroy_wq();
740         frwr_destroy_recovery_wq();
741
742         rc = xprt_unregister_transport(&xprt_rdma_bc);
743         if (rc)
744                 dprintk("RPC:       %s: xprt_unregister(bc) returned %i\n",
745                         __func__, rc);
746 }
747
748 int xprt_rdma_init(void)
749 {
750         int rc;
751
752         rc = frwr_alloc_recovery_wq();
753         if (rc)
754                 return rc;
755
756         rc = rpcrdma_alloc_wq();
757         if (rc) {
758                 frwr_destroy_recovery_wq();
759                 return rc;
760         }
761
762         rc = xprt_register_transport(&xprt_rdma);
763         if (rc) {
764                 rpcrdma_destroy_wq();
765                 frwr_destroy_recovery_wq();
766                 return rc;
767         }
768
769         rc = xprt_register_transport(&xprt_rdma_bc);
770         if (rc) {
771                 xprt_unregister_transport(&xprt_rdma);
772                 rpcrdma_destroy_wq();
773                 frwr_destroy_recovery_wq();
774                 return rc;
775         }
776
777         dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
778
779         dprintk("Defaults:\n");
780         dprintk("\tSlots %d\n"
781                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
782                 xprt_rdma_slot_table_entries,
783                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
784         dprintk("\tPadding %d\n\tMemreg %d\n",
785                 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
786
787 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
788         if (!sunrpc_table_header)
789                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
790 #endif
791         return 0;
792 }