Merge tag 'nfsd-4.9' of git://linux-nfs.org/~bfields/linux
[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 min_inline_size = RPCRDMA_MIN_INLINE;
77 static unsigned int max_inline_size = RPCRDMA_MAX_INLINE;
78 static unsigned int zero;
79 static unsigned int max_padding = PAGE_SIZE;
80 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
81 static unsigned int max_memreg = RPCRDMA_LAST - 1;
82
83 static struct ctl_table_header *sunrpc_table_header;
84
85 static struct ctl_table xr_tunables_table[] = {
86         {
87                 .procname       = "rdma_slot_table_entries",
88                 .data           = &xprt_rdma_slot_table_entries,
89                 .maxlen         = sizeof(unsigned int),
90                 .mode           = 0644,
91                 .proc_handler   = proc_dointvec_minmax,
92                 .extra1         = &min_slot_table_size,
93                 .extra2         = &max_slot_table_size
94         },
95         {
96                 .procname       = "rdma_max_inline_read",
97                 .data           = &xprt_rdma_max_inline_read,
98                 .maxlen         = sizeof(unsigned int),
99                 .mode           = 0644,
100                 .proc_handler   = proc_dointvec,
101                 .extra1         = &min_inline_size,
102                 .extra2         = &max_inline_size,
103         },
104         {
105                 .procname       = "rdma_max_inline_write",
106                 .data           = &xprt_rdma_max_inline_write,
107                 .maxlen         = sizeof(unsigned int),
108                 .mode           = 0644,
109                 .proc_handler   = proc_dointvec,
110                 .extra1         = &min_inline_size,
111                 .extra2         = &max_inline_size,
112         },
113         {
114                 .procname       = "rdma_inline_write_padding",
115                 .data           = &xprt_rdma_inline_write_padding,
116                 .maxlen         = sizeof(unsigned int),
117                 .mode           = 0644,
118                 .proc_handler   = proc_dointvec_minmax,
119                 .extra1         = &zero,
120                 .extra2         = &max_padding,
121         },
122         {
123                 .procname       = "rdma_memreg_strategy",
124                 .data           = &xprt_rdma_memreg_strategy,
125                 .maxlen         = sizeof(unsigned int),
126                 .mode           = 0644,
127                 .proc_handler   = proc_dointvec_minmax,
128                 .extra1         = &min_memreg,
129                 .extra2         = &max_memreg,
130         },
131         {
132                 .procname       = "rdma_pad_optimize",
133                 .data           = &xprt_rdma_pad_optimize,
134                 .maxlen         = sizeof(unsigned int),
135                 .mode           = 0644,
136                 .proc_handler   = proc_dointvec,
137         },
138         { },
139 };
140
141 static struct ctl_table sunrpc_table[] = {
142         {
143                 .procname       = "sunrpc",
144                 .mode           = 0555,
145                 .child          = xr_tunables_table
146         },
147         { },
148 };
149
150 #endif
151
152 static struct rpc_xprt_ops xprt_rdma_procs;     /*forward reference */
153
154 static void
155 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
156 {
157         struct sockaddr_in *sin = (struct sockaddr_in *)sap;
158         char buf[20];
159
160         snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
161         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
162
163         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
164 }
165
166 static void
167 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
168 {
169         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
170         char buf[40];
171
172         snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
173         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
174
175         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
176 }
177
178 void
179 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
180 {
181         char buf[128];
182
183         switch (sap->sa_family) {
184         case AF_INET:
185                 xprt_rdma_format_addresses4(xprt, sap);
186                 break;
187         case AF_INET6:
188                 xprt_rdma_format_addresses6(xprt, sap);
189                 break;
190         default:
191                 pr_err("rpcrdma: Unrecognized address family\n");
192                 return;
193         }
194
195         (void)rpc_ntop(sap, buf, sizeof(buf));
196         xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
197
198         snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
199         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
200
201         snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
202         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
203
204         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
205 }
206
207 void
208 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
209 {
210         unsigned int i;
211
212         for (i = 0; i < RPC_DISPLAY_MAX; i++)
213                 switch (i) {
214                 case RPC_DISPLAY_PROTO:
215                 case RPC_DISPLAY_NETID:
216                         continue;
217                 default:
218                         kfree(xprt->address_strings[i]);
219                 }
220 }
221
222 static void
223 xprt_rdma_connect_worker(struct work_struct *work)
224 {
225         struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
226                                                    rx_connect_worker.work);
227         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
228         int rc = 0;
229
230         xprt_clear_connected(xprt);
231
232         dprintk("RPC:       %s: %sconnect\n", __func__,
233                         r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
234         rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
235         if (rc)
236                 xprt_wake_pending_tasks(xprt, rc);
237
238         dprintk("RPC:       %s: exit\n", __func__);
239         xprt_clear_connecting(xprt);
240 }
241
242 static void
243 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
244 {
245         struct rpcrdma_xprt *r_xprt = container_of(xprt, struct rpcrdma_xprt,
246                                                    rx_xprt);
247
248         pr_info("rpcrdma: injecting transport disconnect on xprt=%p\n", xprt);
249         rdma_disconnect(r_xprt->rx_ia.ri_id);
250 }
251
252 /*
253  * xprt_rdma_destroy
254  *
255  * Destroy the xprt.
256  * Free all memory associated with the object, including its own.
257  * NOTE: none of the *destroy methods free memory for their top-level
258  * objects, even though they may have allocated it (they do free
259  * private memory). It's up to the caller to handle it. In this
260  * case (RDMA transport), all structure memory is inlined with the
261  * struct rpcrdma_xprt.
262  */
263 static void
264 xprt_rdma_destroy(struct rpc_xprt *xprt)
265 {
266         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
267
268         dprintk("RPC:       %s: called\n", __func__);
269
270         cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
271
272         xprt_clear_connected(xprt);
273
274         rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
275         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
276         rpcrdma_ia_close(&r_xprt->rx_ia);
277
278         xprt_rdma_free_addresses(xprt);
279
280         xprt_free(xprt);
281
282         dprintk("RPC:       %s: returning\n", __func__);
283
284         module_put(THIS_MODULE);
285 }
286
287 static const struct rpc_timeout xprt_rdma_default_timeout = {
288         .to_initval = 60 * HZ,
289         .to_maxval = 60 * HZ,
290 };
291
292 /**
293  * xprt_setup_rdma - Set up transport to use RDMA
294  *
295  * @args: rpc transport arguments
296  */
297 static struct rpc_xprt *
298 xprt_setup_rdma(struct xprt_create *args)
299 {
300         struct rpcrdma_create_data_internal cdata;
301         struct rpc_xprt *xprt;
302         struct rpcrdma_xprt *new_xprt;
303         struct rpcrdma_ep *new_ep;
304         struct sockaddr *sap;
305         int rc;
306
307         if (args->addrlen > sizeof(xprt->addr)) {
308                 dprintk("RPC:       %s: address too large\n", __func__);
309                 return ERR_PTR(-EBADF);
310         }
311
312         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
313                         xprt_rdma_slot_table_entries,
314                         xprt_rdma_slot_table_entries);
315         if (xprt == NULL) {
316                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
317                         __func__);
318                 return ERR_PTR(-ENOMEM);
319         }
320
321         /* 60 second timeout, no retries */
322         xprt->timeout = &xprt_rdma_default_timeout;
323         xprt->bind_timeout = RPCRDMA_BIND_TO;
324         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
325         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
326
327         xprt->resvport = 0;             /* privileged port not needed */
328         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
329         xprt->ops = &xprt_rdma_procs;
330
331         /*
332          * Set up RDMA-specific connect data.
333          */
334
335         sap = (struct sockaddr *)&cdata.addr;
336         memcpy(sap, args->dstaddr, args->addrlen);
337
338         /* Ensure xprt->addr holds valid server TCP (not RDMA)
339          * address, for any side protocols which peek at it */
340         xprt->prot = IPPROTO_TCP;
341         xprt->addrlen = args->addrlen;
342         memcpy(&xprt->addr, sap, xprt->addrlen);
343
344         if (rpc_get_port(sap))
345                 xprt_set_bound(xprt);
346
347         cdata.max_requests = xprt->max_reqs;
348
349         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
350         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
351
352         cdata.inline_wsize = xprt_rdma_max_inline_write;
353         if (cdata.inline_wsize > cdata.wsize)
354                 cdata.inline_wsize = cdata.wsize;
355
356         cdata.inline_rsize = xprt_rdma_max_inline_read;
357         if (cdata.inline_rsize > cdata.rsize)
358                 cdata.inline_rsize = cdata.rsize;
359
360         cdata.padding = xprt_rdma_inline_write_padding;
361
362         /*
363          * Create new transport instance, which includes initialized
364          *  o ia
365          *  o endpoint
366          *  o buffers
367          */
368
369         new_xprt = rpcx_to_rdmax(xprt);
370
371         rc = rpcrdma_ia_open(new_xprt, sap, xprt_rdma_memreg_strategy);
372         if (rc)
373                 goto out1;
374
375         /*
376          * initialize and create ep
377          */
378         new_xprt->rx_data = cdata;
379         new_ep = &new_xprt->rx_ep;
380         new_ep->rep_remote_addr = cdata.addr;
381
382         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
383                                 &new_xprt->rx_ia, &new_xprt->rx_data);
384         if (rc)
385                 goto out2;
386
387         /*
388          * Allocate pre-registered send and receive buffers for headers and
389          * any inline data. Also specify any padding which will be provided
390          * from a preregistered zero buffer.
391          */
392         rc = rpcrdma_buffer_create(new_xprt);
393         if (rc)
394                 goto out3;
395
396         /*
397          * Register a callback for connection events. This is necessary because
398          * connection loss notification is async. We also catch connection loss
399          * when reaping receives.
400          */
401         INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
402                           xprt_rdma_connect_worker);
403
404         xprt_rdma_format_addresses(xprt, sap);
405         xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
406         if (xprt->max_payload == 0)
407                 goto out4;
408         xprt->max_payload <<= PAGE_SHIFT;
409         dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
410                 __func__, xprt->max_payload);
411
412         if (!try_module_get(THIS_MODULE))
413                 goto out4;
414
415         dprintk("RPC:       %s: %s:%s\n", __func__,
416                 xprt->address_strings[RPC_DISPLAY_ADDR],
417                 xprt->address_strings[RPC_DISPLAY_PORT]);
418         return xprt;
419
420 out4:
421         xprt_rdma_free_addresses(xprt);
422         rc = -EINVAL;
423 out3:
424         rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
425 out2:
426         rpcrdma_ia_close(&new_xprt->rx_ia);
427 out1:
428         xprt_free(xprt);
429         return ERR_PTR(rc);
430 }
431
432 /*
433  * Close a connection, during shutdown or timeout/reconnect
434  */
435 static void
436 xprt_rdma_close(struct rpc_xprt *xprt)
437 {
438         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
439
440         dprintk("RPC:       %s: closing\n", __func__);
441         if (r_xprt->rx_ep.rep_connected > 0)
442                 xprt->reestablish_timeout = 0;
443         xprt_disconnect_done(xprt);
444         rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
445 }
446
447 static void
448 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
449 {
450         struct sockaddr_in *sap;
451
452         sap = (struct sockaddr_in *)&xprt->addr;
453         sap->sin_port = htons(port);
454         sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
455         sap->sin_port = htons(port);
456         dprintk("RPC:       %s: %u\n", __func__, port);
457 }
458
459 static void
460 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
461 {
462         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
463
464         if (r_xprt->rx_ep.rep_connected != 0) {
465                 /* Reconnect */
466                 schedule_delayed_work(&r_xprt->rx_connect_worker,
467                                       xprt->reestablish_timeout);
468                 xprt->reestablish_timeout <<= 1;
469                 if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
470                         xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
471                 else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
472                         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
473         } else {
474                 schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
475                 if (!RPC_IS_ASYNC(task))
476                         flush_delayed_work(&r_xprt->rx_connect_worker);
477         }
478 }
479
480 /*
481  * The RDMA allocate/free functions need the task structure as a place
482  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
483  * sequence.
484  *
485  * The RPC layer allocates both send and receive buffers in the same call
486  * (rq_send_buf and rq_rcv_buf are both part of a single contiguous buffer).
487  * We may register rq_rcv_buf when using reply chunks.
488  */
489 static void *
490 xprt_rdma_allocate(struct rpc_task *task, size_t size)
491 {
492         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
493         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
494         struct rpcrdma_regbuf *rb;
495         struct rpcrdma_req *req;
496         size_t min_size;
497         gfp_t flags;
498
499         req = rpcrdma_buffer_get(&r_xprt->rx_buf);
500         if (req == NULL)
501                 return NULL;
502
503         flags = RPCRDMA_DEF_GFP;
504         if (RPC_IS_SWAPPER(task))
505                 flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
506
507         if (req->rl_rdmabuf == NULL)
508                 goto out_rdmabuf;
509         if (req->rl_sendbuf == NULL)
510                 goto out_sendbuf;
511         if (size > req->rl_sendbuf->rg_size)
512                 goto out_sendbuf;
513
514 out:
515         dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
516         req->rl_connect_cookie = 0;     /* our reserved value */
517         req->rl_task = task;
518         return req->rl_sendbuf->rg_base;
519
520 out_rdmabuf:
521         min_size = RPCRDMA_INLINE_WRITE_THRESHOLD(task->tk_rqstp);
522         rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, min_size, flags);
523         if (IS_ERR(rb))
524                 goto out_fail;
525         req->rl_rdmabuf = rb;
526
527 out_sendbuf:
528         /* XDR encoding and RPC/RDMA marshaling of this request has not
529          * yet occurred. Thus a lower bound is needed to prevent buffer
530          * overrun during marshaling.
531          *
532          * RPC/RDMA marshaling may choose to send payload bearing ops
533          * inline, if the result is smaller than the inline threshold.
534          * The value of the "size" argument accounts for header
535          * requirements but not for the payload in these cases.
536          *
537          * Likewise, allocate enough space to receive a reply up to the
538          * size of the inline threshold.
539          *
540          * It's unlikely that both the send header and the received
541          * reply will be large, but slush is provided here to allow
542          * flexibility when marshaling.
543          */
544         min_size = RPCRDMA_INLINE_READ_THRESHOLD(task->tk_rqstp);
545         min_size += RPCRDMA_INLINE_WRITE_THRESHOLD(task->tk_rqstp);
546         if (size < min_size)
547                 size = min_size;
548
549         rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, size, flags);
550         if (IS_ERR(rb))
551                 goto out_fail;
552         rb->rg_owner = req;
553
554         r_xprt->rx_stats.hardway_register_count += size;
555         rpcrdma_free_regbuf(&r_xprt->rx_ia, req->rl_sendbuf);
556         req->rl_sendbuf = rb;
557         goto out;
558
559 out_fail:
560         rpcrdma_buffer_put(req);
561         return NULL;
562 }
563
564 /*
565  * This function returns all RDMA resources to the pool.
566  */
567 static void
568 xprt_rdma_free(void *buffer)
569 {
570         struct rpcrdma_req *req;
571         struct rpcrdma_xprt *r_xprt;
572         struct rpcrdma_regbuf *rb;
573
574         if (buffer == NULL)
575                 return;
576
577         rb = container_of(buffer, struct rpcrdma_regbuf, rg_base[0]);
578         req = rb->rg_owner;
579         if (req->rl_backchannel)
580                 return;
581
582         r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
583
584         dprintk("RPC:       %s: called on 0x%p\n", __func__, req->rl_reply);
585
586         r_xprt->rx_ia.ri_ops->ro_unmap_safe(r_xprt, req,
587                                             !RPC_IS_ASYNC(req->rl_task));
588
589         rpcrdma_buffer_put(req);
590 }
591
592 /**
593  * xprt_rdma_send_request - marshal and send an RPC request
594  * @task: RPC task with an RPC message in rq_snd_buf
595  *
596  * Return values:
597  *        0:    The request has been sent
598  * ENOTCONN:    Caller needs to invoke connect logic then call again
599  *  ENOBUFS:    Call again later to send the request
600  *      EIO:    A permanent error occurred. The request was not sent,
601  *              and don't try it again
602  *
603  * send_request invokes the meat of RPC RDMA. It must do the following:
604  *
605  *  1.  Marshal the RPC request into an RPC RDMA request, which means
606  *      putting a header in front of data, and creating IOVs for RDMA
607  *      from those in the request.
608  *  2.  In marshaling, detect opportunities for RDMA, and use them.
609  *  3.  Post a recv message to set up asynch completion, then send
610  *      the request (rpcrdma_ep_post).
611  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
612  */
613 static int
614 xprt_rdma_send_request(struct rpc_task *task)
615 {
616         struct rpc_rqst *rqst = task->tk_rqstp;
617         struct rpc_xprt *xprt = rqst->rq_xprt;
618         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
619         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
620         int rc = 0;
621
622         /* On retransmit, remove any previously registered chunks */
623         r_xprt->rx_ia.ri_ops->ro_unmap_safe(r_xprt, req, false);
624
625         rc = rpcrdma_marshal_req(rqst);
626         if (rc < 0)
627                 goto failed_marshal;
628
629         if (req->rl_reply == NULL)              /* e.g. reconnection */
630                 rpcrdma_recv_buffer_get(req);
631
632         /* Must suppress retransmit to maintain credits */
633         if (req->rl_connect_cookie == xprt->connect_cookie)
634                 goto drop_connection;
635         req->rl_connect_cookie = xprt->connect_cookie;
636
637         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
638                 goto drop_connection;
639
640         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
641         rqst->rq_bytes_sent = 0;
642         return 0;
643
644 failed_marshal:
645         dprintk("RPC:       %s: rpcrdma_marshal_req failed, status %i\n",
646                 __func__, rc);
647         if (rc == -EIO)
648                 r_xprt->rx_stats.failed_marshal_count++;
649         if (rc != -ENOTCONN)
650                 return rc;
651 drop_connection:
652         xprt_disconnect_done(xprt);
653         return -ENOTCONN;       /* implies disconnect */
654 }
655
656 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
657 {
658         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
659         long idle_time = 0;
660
661         if (xprt_connected(xprt))
662                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
663
664         seq_puts(seq, "\txprt:\trdma ");
665         seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
666                    0,   /* need a local port? */
667                    xprt->stat.bind_count,
668                    xprt->stat.connect_count,
669                    xprt->stat.connect_time,
670                    idle_time,
671                    xprt->stat.sends,
672                    xprt->stat.recvs,
673                    xprt->stat.bad_xids,
674                    xprt->stat.req_u,
675                    xprt->stat.bklog_u);
676         seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
677                    r_xprt->rx_stats.read_chunk_count,
678                    r_xprt->rx_stats.write_chunk_count,
679                    r_xprt->rx_stats.reply_chunk_count,
680                    r_xprt->rx_stats.total_rdma_request,
681                    r_xprt->rx_stats.total_rdma_reply,
682                    r_xprt->rx_stats.pullup_copy_count,
683                    r_xprt->rx_stats.fixup_copy_count,
684                    r_xprt->rx_stats.hardway_register_count,
685                    r_xprt->rx_stats.failed_marshal_count,
686                    r_xprt->rx_stats.bad_reply_count,
687                    r_xprt->rx_stats.nomsg_call_count);
688         seq_printf(seq, "%lu %lu %lu\n",
689                    r_xprt->rx_stats.mrs_recovered,
690                    r_xprt->rx_stats.mrs_orphaned,
691                    r_xprt->rx_stats.mrs_allocated);
692 }
693
694 static int
695 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
696 {
697         return 0;
698 }
699
700 static void
701 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
702 {
703 }
704
705 /*
706  * Plumbing for rpc transport switch and kernel module
707  */
708
709 static struct rpc_xprt_ops xprt_rdma_procs = {
710         .reserve_xprt           = xprt_reserve_xprt_cong,
711         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
712         .alloc_slot             = xprt_alloc_slot,
713         .release_request        = xprt_release_rqst_cong,       /* ditto */
714         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
715         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
716         .set_port               = xprt_rdma_set_port,
717         .connect                = xprt_rdma_connect,
718         .buf_alloc              = xprt_rdma_allocate,
719         .buf_free               = xprt_rdma_free,
720         .send_request           = xprt_rdma_send_request,
721         .close                  = xprt_rdma_close,
722         .destroy                = xprt_rdma_destroy,
723         .print_stats            = xprt_rdma_print_stats,
724         .enable_swap            = xprt_rdma_enable_swap,
725         .disable_swap           = xprt_rdma_disable_swap,
726         .inject_disconnect      = xprt_rdma_inject_disconnect,
727 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
728         .bc_setup               = xprt_rdma_bc_setup,
729         .bc_up                  = xprt_rdma_bc_up,
730         .bc_maxpayload          = xprt_rdma_bc_maxpayload,
731         .bc_free_rqst           = xprt_rdma_bc_free_rqst,
732         .bc_destroy             = xprt_rdma_bc_destroy,
733 #endif
734 };
735
736 static struct xprt_class xprt_rdma = {
737         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
738         .name                   = "rdma",
739         .owner                  = THIS_MODULE,
740         .ident                  = XPRT_TRANSPORT_RDMA,
741         .setup                  = xprt_setup_rdma,
742 };
743
744 void xprt_rdma_cleanup(void)
745 {
746         int rc;
747
748         dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
749 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
750         if (sunrpc_table_header) {
751                 unregister_sysctl_table(sunrpc_table_header);
752                 sunrpc_table_header = NULL;
753         }
754 #endif
755         rc = xprt_unregister_transport(&xprt_rdma);
756         if (rc)
757                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
758                         __func__, rc);
759
760         rpcrdma_destroy_wq();
761
762         rc = xprt_unregister_transport(&xprt_rdma_bc);
763         if (rc)
764                 dprintk("RPC:       %s: xprt_unregister(bc) returned %i\n",
765                         __func__, rc);
766 }
767
768 int xprt_rdma_init(void)
769 {
770         int rc;
771
772         rc = rpcrdma_alloc_wq();
773         if (rc)
774                 return rc;
775
776         rc = xprt_register_transport(&xprt_rdma);
777         if (rc) {
778                 rpcrdma_destroy_wq();
779                 return rc;
780         }
781
782         rc = xprt_register_transport(&xprt_rdma_bc);
783         if (rc) {
784                 xprt_unregister_transport(&xprt_rdma);
785                 rpcrdma_destroy_wq();
786                 return rc;
787         }
788
789         dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
790
791         dprintk("Defaults:\n");
792         dprintk("\tSlots %d\n"
793                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
794                 xprt_rdma_slot_table_entries,
795                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
796         dprintk("\tPadding %d\n\tMemreg %d\n",
797                 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
798
799 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
800         if (!sunrpc_table_header)
801                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
802 #endif
803         return 0;
804 }