]> git.karo-electronics.de Git - karo-tx-linux.git/blob - net/sunrpc/xprtrdma/transport.c
xprtrdma: Perform a full marshal on retransmit
[karo-tx-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/init.h>
52 #include <linux/slab.h>
53 #include <linux/seq_file.h>
54 #include <linux/sunrpc/addr.h>
55
56 #include "xprt_rdma.h"
57
58 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
59 # define RPCDBG_FACILITY        RPCDBG_TRANS
60 #endif
61
62 MODULE_LICENSE("Dual BSD/GPL");
63
64 MODULE_DESCRIPTION("RPC/RDMA Transport for Linux kernel NFS");
65 MODULE_AUTHOR("Network Appliance, Inc.");
66
67 /*
68  * tunables
69  */
70
71 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
72 static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
73 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
74 static unsigned int xprt_rdma_inline_write_padding;
75 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
76                 int xprt_rdma_pad_optimize = 1;
77
78 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
79
80 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
81 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
82 static unsigned int zero;
83 static unsigned int max_padding = PAGE_SIZE;
84 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
85 static unsigned int max_memreg = RPCRDMA_LAST - 1;
86
87 static struct ctl_table_header *sunrpc_table_header;
88
89 static struct ctl_table xr_tunables_table[] = {
90         {
91                 .procname       = "rdma_slot_table_entries",
92                 .data           = &xprt_rdma_slot_table_entries,
93                 .maxlen         = sizeof(unsigned int),
94                 .mode           = 0644,
95                 .proc_handler   = proc_dointvec_minmax,
96                 .extra1         = &min_slot_table_size,
97                 .extra2         = &max_slot_table_size
98         },
99         {
100                 .procname       = "rdma_max_inline_read",
101                 .data           = &xprt_rdma_max_inline_read,
102                 .maxlen         = sizeof(unsigned int),
103                 .mode           = 0644,
104                 .proc_handler   = proc_dointvec,
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,
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 #define RPCRDMA_BIND_TO         (60U * HZ)
153 #define RPCRDMA_INIT_REEST_TO   (5U * HZ)
154 #define RPCRDMA_MAX_REEST_TO    (30U * HZ)
155 #define RPCRDMA_IDLE_DISC_TO    (5U * 60 * HZ)
156
157 static struct rpc_xprt_ops xprt_rdma_procs;     /* forward reference */
158
159 static void
160 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
161 {
162         struct sockaddr_in *sin = (struct sockaddr_in *)sap;
163         char buf[20];
164
165         snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
166         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
167
168         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
169 }
170
171 static void
172 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
173 {
174         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
175         char buf[40];
176
177         snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
178         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
179
180         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
181 }
182
183 static void
184 xprt_rdma_format_addresses(struct rpc_xprt *xprt)
185 {
186         struct sockaddr *sap = (struct sockaddr *)
187                                         &rpcx_to_rdmad(xprt).addr;
188         char buf[128];
189
190         switch (sap->sa_family) {
191         case AF_INET:
192                 xprt_rdma_format_addresses4(xprt, sap);
193                 break;
194         case AF_INET6:
195                 xprt_rdma_format_addresses6(xprt, sap);
196                 break;
197         default:
198                 pr_err("rpcrdma: Unrecognized address family\n");
199                 return;
200         }
201
202         (void)rpc_ntop(sap, buf, sizeof(buf));
203         xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
204
205         snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
206         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
207
208         snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
209         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
210
211         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
212 }
213
214 static void
215 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
216 {
217         unsigned int i;
218
219         for (i = 0; i < RPC_DISPLAY_MAX; i++)
220                 switch (i) {
221                 case RPC_DISPLAY_PROTO:
222                 case RPC_DISPLAY_NETID:
223                         continue;
224                 default:
225                         kfree(xprt->address_strings[i]);
226                 }
227 }
228
229 static void
230 xprt_rdma_connect_worker(struct work_struct *work)
231 {
232         struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
233                                                    rx_connect_worker.work);
234         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
235         int rc = 0;
236
237         xprt_clear_connected(xprt);
238
239         dprintk("RPC:       %s: %sconnect\n", __func__,
240                         r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
241         rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
242         if (rc)
243                 xprt_wake_pending_tasks(xprt, rc);
244
245         dprintk("RPC:       %s: exit\n", __func__);
246         xprt_clear_connecting(xprt);
247 }
248
249 /*
250  * xprt_rdma_destroy
251  *
252  * Destroy the xprt.
253  * Free all memory associated with the object, including its own.
254  * NOTE: none of the *destroy methods free memory for their top-level
255  * objects, even though they may have allocated it (they do free
256  * private memory). It's up to the caller to handle it. In this
257  * case (RDMA transport), all structure memory is inlined with the
258  * struct rpcrdma_xprt.
259  */
260 static void
261 xprt_rdma_destroy(struct rpc_xprt *xprt)
262 {
263         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
264
265         dprintk("RPC:       %s: called\n", __func__);
266
267         cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
268
269         xprt_clear_connected(xprt);
270
271         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
272         rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
273         rpcrdma_ia_close(&r_xprt->rx_ia);
274
275         xprt_rdma_free_addresses(xprt);
276
277         xprt_free(xprt);
278
279         dprintk("RPC:       %s: returning\n", __func__);
280
281         module_put(THIS_MODULE);
282 }
283
284 static const struct rpc_timeout xprt_rdma_default_timeout = {
285         .to_initval = 60 * HZ,
286         .to_maxval = 60 * HZ,
287 };
288
289 /**
290  * xprt_setup_rdma - Set up transport to use RDMA
291  *
292  * @args: rpc transport arguments
293  */
294 static struct rpc_xprt *
295 xprt_setup_rdma(struct xprt_create *args)
296 {
297         struct rpcrdma_create_data_internal cdata;
298         struct rpc_xprt *xprt;
299         struct rpcrdma_xprt *new_xprt;
300         struct rpcrdma_ep *new_ep;
301         struct sockaddr_in *sin;
302         int rc;
303
304         if (args->addrlen > sizeof(xprt->addr)) {
305                 dprintk("RPC:       %s: address too large\n", __func__);
306                 return ERR_PTR(-EBADF);
307         }
308
309         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
310                         xprt_rdma_slot_table_entries,
311                         xprt_rdma_slot_table_entries);
312         if (xprt == NULL) {
313                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
314                         __func__);
315                 return ERR_PTR(-ENOMEM);
316         }
317
318         /* 60 second timeout, no retries */
319         xprt->timeout = &xprt_rdma_default_timeout;
320         xprt->bind_timeout = RPCRDMA_BIND_TO;
321         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
322         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
323
324         xprt->resvport = 0;             /* privileged port not needed */
325         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
326         xprt->ops = &xprt_rdma_procs;
327
328         /*
329          * Set up RDMA-specific connect data.
330          */
331
332         /* Put server RDMA address in local cdata */
333         memcpy(&cdata.addr, args->dstaddr, args->addrlen);
334
335         /* Ensure xprt->addr holds valid server TCP (not RDMA)
336          * address, for any side protocols which peek at it */
337         xprt->prot = IPPROTO_TCP;
338         xprt->addrlen = args->addrlen;
339         memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
340
341         sin = (struct sockaddr_in *)&cdata.addr;
342         if (ntohs(sin->sin_port) != 0)
343                 xprt_set_bound(xprt);
344
345         dprintk("RPC:       %s: %pI4:%u\n",
346                 __func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port));
347
348         /* Set max requests */
349         cdata.max_requests = xprt->max_reqs;
350
351         /* Set some length limits */
352         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
353         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
354
355         cdata.inline_wsize = xprt_rdma_max_inline_write;
356         if (cdata.inline_wsize > cdata.wsize)
357                 cdata.inline_wsize = cdata.wsize;
358
359         cdata.inline_rsize = xprt_rdma_max_inline_read;
360         if (cdata.inline_rsize > cdata.rsize)
361                 cdata.inline_rsize = cdata.rsize;
362
363         cdata.padding = xprt_rdma_inline_write_padding;
364
365         /*
366          * Create new transport instance, which includes initialized
367          *  o ia
368          *  o endpoint
369          *  o buffers
370          */
371
372         new_xprt = rpcx_to_rdmax(xprt);
373
374         rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
375                                 xprt_rdma_memreg_strategy);
376         if (rc)
377                 goto out1;
378
379         /*
380          * initialize and create ep
381          */
382         new_xprt->rx_data = cdata;
383         new_ep = &new_xprt->rx_ep;
384         new_ep->rep_remote_addr = cdata.addr;
385
386         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
387                                 &new_xprt->rx_ia, &new_xprt->rx_data);
388         if (rc)
389                 goto out2;
390
391         /*
392          * Allocate pre-registered send and receive buffers for headers and
393          * any inline data. Also specify any padding which will be provided
394          * from a preregistered zero buffer.
395          */
396         rc = rpcrdma_buffer_create(new_xprt);
397         if (rc)
398                 goto out3;
399
400         /*
401          * Register a callback for connection events. This is necessary because
402          * connection loss notification is async. We also catch connection loss
403          * when reaping receives.
404          */
405         INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
406                           xprt_rdma_connect_worker);
407
408         xprt_rdma_format_addresses(xprt);
409         xprt->max_payload = rpcrdma_max_payload(new_xprt);
410         dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
411                 __func__, xprt->max_payload);
412
413         if (!try_module_get(THIS_MODULE))
414                 goto out4;
415
416         return xprt;
417
418 out4:
419         xprt_rdma_free_addresses(xprt);
420         rc = -EINVAL;
421 out3:
422         rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
423 out2:
424         rpcrdma_ia_close(&new_xprt->rx_ia);
425 out1:
426         xprt_free(xprt);
427         return ERR_PTR(rc);
428 }
429
430 /*
431  * Close a connection, during shutdown or timeout/reconnect
432  */
433 static void
434 xprt_rdma_close(struct rpc_xprt *xprt)
435 {
436         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
437
438         dprintk("RPC:       %s: closing\n", __func__);
439         if (r_xprt->rx_ep.rep_connected > 0)
440                 xprt->reestablish_timeout = 0;
441         xprt_disconnect_done(xprt);
442         rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
443 }
444
445 static void
446 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
447 {
448         struct sockaddr_in *sap;
449
450         sap = (struct sockaddr_in *)&xprt->addr;
451         sap->sin_port = htons(port);
452         sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
453         sap->sin_port = htons(port);
454         dprintk("RPC:       %s: %u\n", __func__, port);
455 }
456
457 static void
458 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
459 {
460         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
461
462         if (r_xprt->rx_ep.rep_connected != 0) {
463                 /* Reconnect */
464                 schedule_delayed_work(&r_xprt->rx_connect_worker,
465                                       xprt->reestablish_timeout);
466                 xprt->reestablish_timeout <<= 1;
467                 if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
468                         xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
469                 else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
470                         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
471         } else {
472                 schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
473                 if (!RPC_IS_ASYNC(task))
474                         flush_delayed_work(&r_xprt->rx_connect_worker);
475         }
476 }
477
478 /*
479  * The RDMA allocate/free functions need the task structure as a place
480  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
481  * sequence.
482  *
483  * The RPC layer allocates both send and receive buffers in the same call
484  * (rq_send_buf and rq_rcv_buf are both part of a single contiguous buffer).
485  * We may register rq_rcv_buf when using reply chunks.
486  */
487 static void *
488 xprt_rdma_allocate(struct rpc_task *task, size_t size)
489 {
490         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
491         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
492         struct rpcrdma_regbuf *rb;
493         struct rpcrdma_req *req;
494         size_t min_size;
495         gfp_t flags;
496
497         req = rpcrdma_buffer_get(&r_xprt->rx_buf);
498         if (req == NULL)
499                 return NULL;
500
501         flags = GFP_NOIO | __GFP_NOWARN;
502         if (RPC_IS_SWAPPER(task))
503                 flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
504
505         if (req->rl_rdmabuf == NULL)
506                 goto out_rdmabuf;
507         if (req->rl_sendbuf == NULL)
508                 goto out_sendbuf;
509         if (size > req->rl_sendbuf->rg_size)
510                 goto out_sendbuf;
511
512 out:
513         dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
514         req->rl_connect_cookie = 0;     /* our reserved value */
515         return req->rl_sendbuf->rg_base;
516
517 out_rdmabuf:
518         min_size = RPCRDMA_INLINE_WRITE_THRESHOLD(task->tk_rqstp);
519         rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, min_size, flags);
520         if (IS_ERR(rb))
521                 goto out_fail;
522         req->rl_rdmabuf = rb;
523
524 out_sendbuf:
525         /* XDR encoding and RPC/RDMA marshaling of this request has not
526          * yet occurred. Thus a lower bound is needed to prevent buffer
527          * overrun during marshaling.
528          *
529          * RPC/RDMA marshaling may choose to send payload bearing ops
530          * inline, if the result is smaller than the inline threshold.
531          * The value of the "size" argument accounts for header
532          * requirements but not for the payload in these cases.
533          *
534          * Likewise, allocate enough space to receive a reply up to the
535          * size of the inline threshold.
536          *
537          * It's unlikely that both the send header and the received
538          * reply will be large, but slush is provided here to allow
539          * flexibility when marshaling.
540          */
541         min_size = RPCRDMA_INLINE_READ_THRESHOLD(task->tk_rqstp);
542         min_size += RPCRDMA_INLINE_WRITE_THRESHOLD(task->tk_rqstp);
543         if (size < min_size)
544                 size = min_size;
545
546         rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, size, flags);
547         if (IS_ERR(rb))
548                 goto out_fail;
549         rb->rg_owner = req;
550
551         r_xprt->rx_stats.hardway_register_count += size;
552         rpcrdma_free_regbuf(&r_xprt->rx_ia, req->rl_sendbuf);
553         req->rl_sendbuf = rb;
554         goto out;
555
556 out_fail:
557         rpcrdma_buffer_put(req);
558         r_xprt->rx_stats.failed_marshal_count++;
559         return NULL;
560 }
561
562 /*
563  * This function returns all RDMA resources to the pool.
564  */
565 static void
566 xprt_rdma_free(void *buffer)
567 {
568         struct rpcrdma_req *req;
569         struct rpcrdma_xprt *r_xprt;
570         struct rpcrdma_regbuf *rb;
571         int i;
572
573         if (buffer == NULL)
574                 return;
575
576         rb = container_of(buffer, struct rpcrdma_regbuf, rg_base[0]);
577         req = rb->rg_owner;
578         r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
579
580         dprintk("RPC:       %s: called on 0x%p\n", __func__, req->rl_reply);
581
582         for (i = 0; req->rl_nchunks;) {
583                 --req->rl_nchunks;
584                 i += rpcrdma_deregister_external(
585                         &req->rl_segments[i], r_xprt);
586         }
587
588         rpcrdma_buffer_put(req);
589 }
590
591 /*
592  * send_request invokes the meat of RPC RDMA. It must do the following:
593  *  1.  Marshal the RPC request into an RPC RDMA request, which means
594  *      putting a header in front of data, and creating IOVs for RDMA
595  *      from those in the request.
596  *  2.  In marshaling, detect opportunities for RDMA, and use them.
597  *  3.  Post a recv message to set up asynch completion, then send
598  *      the request (rpcrdma_ep_post).
599  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
600  */
601
602 static int
603 xprt_rdma_send_request(struct rpc_task *task)
604 {
605         struct rpc_rqst *rqst = task->tk_rqstp;
606         struct rpc_xprt *xprt = rqst->rq_xprt;
607         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
608         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
609         int rc = 0;
610
611         rc = rpcrdma_marshal_req(rqst);
612         if (rc < 0)
613                 goto failed_marshal;
614
615         if (req->rl_reply == NULL)              /* e.g. reconnection */
616                 rpcrdma_recv_buffer_get(req);
617
618         if (req->rl_reply) {
619                 req->rl_reply->rr_func = rpcrdma_reply_handler;
620                 /* this need only be done once, but... */
621                 req->rl_reply->rr_xprt = xprt;
622         }
623
624         /* Must suppress retransmit to maintain credits */
625         if (req->rl_connect_cookie == xprt->connect_cookie)
626                 goto drop_connection;
627         req->rl_connect_cookie = xprt->connect_cookie;
628
629         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
630                 goto drop_connection;
631
632         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
633         rqst->rq_bytes_sent = 0;
634         return 0;
635
636 failed_marshal:
637         r_xprt->rx_stats.failed_marshal_count++;
638         dprintk("RPC:       %s: rpcrdma_marshal_req failed, status %i\n",
639                 __func__, rc);
640         if (rc == -EIO)
641                 return -EIO;
642 drop_connection:
643         xprt_disconnect_done(xprt);
644         return -ENOTCONN;       /* implies disconnect */
645 }
646
647 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
648 {
649         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
650         long idle_time = 0;
651
652         if (xprt_connected(xprt))
653                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
654
655         seq_printf(seq,
656           "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
657           "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
658
659            0,   /* need a local port? */
660            xprt->stat.bind_count,
661            xprt->stat.connect_count,
662            xprt->stat.connect_time,
663            idle_time,
664            xprt->stat.sends,
665            xprt->stat.recvs,
666            xprt->stat.bad_xids,
667            xprt->stat.req_u,
668            xprt->stat.bklog_u,
669
670            r_xprt->rx_stats.read_chunk_count,
671            r_xprt->rx_stats.write_chunk_count,
672            r_xprt->rx_stats.reply_chunk_count,
673            r_xprt->rx_stats.total_rdma_request,
674            r_xprt->rx_stats.total_rdma_reply,
675            r_xprt->rx_stats.pullup_copy_count,
676            r_xprt->rx_stats.fixup_copy_count,
677            r_xprt->rx_stats.hardway_register_count,
678            r_xprt->rx_stats.failed_marshal_count,
679            r_xprt->rx_stats.bad_reply_count);
680 }
681
682 /*
683  * Plumbing for rpc transport switch and kernel module
684  */
685
686 static struct rpc_xprt_ops xprt_rdma_procs = {
687         .reserve_xprt           = xprt_reserve_xprt_cong,
688         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
689         .alloc_slot             = xprt_alloc_slot,
690         .release_request        = xprt_release_rqst_cong,       /* ditto */
691         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
692         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
693         .set_port               = xprt_rdma_set_port,
694         .connect                = xprt_rdma_connect,
695         .buf_alloc              = xprt_rdma_allocate,
696         .buf_free               = xprt_rdma_free,
697         .send_request           = xprt_rdma_send_request,
698         .close                  = xprt_rdma_close,
699         .destroy                = xprt_rdma_destroy,
700         .print_stats            = xprt_rdma_print_stats
701 };
702
703 static struct xprt_class xprt_rdma = {
704         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
705         .name                   = "rdma",
706         .owner                  = THIS_MODULE,
707         .ident                  = XPRT_TRANSPORT_RDMA,
708         .setup                  = xprt_setup_rdma,
709 };
710
711 static void __exit xprt_rdma_cleanup(void)
712 {
713         int rc;
714
715         dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
716 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
717         if (sunrpc_table_header) {
718                 unregister_sysctl_table(sunrpc_table_header);
719                 sunrpc_table_header = NULL;
720         }
721 #endif
722         rc = xprt_unregister_transport(&xprt_rdma);
723         if (rc)
724                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
725                         __func__, rc);
726 }
727
728 static int __init xprt_rdma_init(void)
729 {
730         int rc;
731
732         rc = xprt_register_transport(&xprt_rdma);
733
734         if (rc)
735                 return rc;
736
737         dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
738
739         dprintk("Defaults:\n");
740         dprintk("\tSlots %d\n"
741                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
742                 xprt_rdma_slot_table_entries,
743                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
744         dprintk("\tPadding %d\n\tMemreg %d\n",
745                 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
746
747 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
748         if (!sunrpc_table_header)
749                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
750 #endif
751         return 0;
752 }
753
754 module_init(xprt_rdma_init);
755 module_exit(xprt_rdma_cleanup);