1dd6123070e94b52023a2f6596462edb3928e860
[linux-2.6.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/seq_file.h>
53
54 #include "xprt_rdma.h"
55
56 #ifdef RPC_DEBUG
57 # define RPCDBG_FACILITY        RPCDBG_TRANS
58 #endif
59
60 MODULE_LICENSE("Dual BSD/GPL");
61
62 MODULE_DESCRIPTION("RPC/RDMA Transport for Linux kernel NFS");
63 MODULE_AUTHOR("Network Appliance, Inc.");
64
65 /*
66  * tunables
67  */
68
69 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
70 static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
71 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
72 static unsigned int xprt_rdma_inline_write_padding;
73 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
74                 int xprt_rdma_pad_optimize = 0;
75
76 #ifdef RPC_DEBUG
77
78 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
79 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
80 static unsigned int zero;
81 static unsigned int max_padding = PAGE_SIZE;
82 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
83 static unsigned int max_memreg = RPCRDMA_LAST - 1;
84
85 static struct ctl_table_header *sunrpc_table_header;
86
87 static ctl_table xr_tunables_table[] = {
88         {
89                 .ctl_name       = CTL_UNNUMBERED,
90                 .procname       = "rdma_slot_table_entries",
91                 .data           = &xprt_rdma_slot_table_entries,
92                 .maxlen         = sizeof(unsigned int),
93                 .mode           = 0644,
94                 .proc_handler   = &proc_dointvec_minmax,
95                 .strategy       = &sysctl_intvec,
96                 .extra1         = &min_slot_table_size,
97                 .extra2         = &max_slot_table_size
98         },
99         {
100                 .ctl_name       = CTL_UNNUMBERED,
101                 .procname       = "rdma_max_inline_read",
102                 .data           = &xprt_rdma_max_inline_read,
103                 .maxlen         = sizeof(unsigned int),
104                 .mode           = 0644,
105                 .proc_handler   = &proc_dointvec,
106                 .strategy       = &sysctl_intvec,
107         },
108         {
109                 .ctl_name       = CTL_UNNUMBERED,
110                 .procname       = "rdma_max_inline_write",
111                 .data           = &xprt_rdma_max_inline_write,
112                 .maxlen         = sizeof(unsigned int),
113                 .mode           = 0644,
114                 .proc_handler   = &proc_dointvec,
115                 .strategy       = &sysctl_intvec,
116         },
117         {
118                 .ctl_name       = CTL_UNNUMBERED,
119                 .procname       = "rdma_inline_write_padding",
120                 .data           = &xprt_rdma_inline_write_padding,
121                 .maxlen         = sizeof(unsigned int),
122                 .mode           = 0644,
123                 .proc_handler   = &proc_dointvec_minmax,
124                 .strategy       = &sysctl_intvec,
125                 .extra1         = &zero,
126                 .extra2         = &max_padding,
127         },
128         {
129                 .ctl_name       = CTL_UNNUMBERED,
130                 .procname       = "rdma_memreg_strategy",
131                 .data           = &xprt_rdma_memreg_strategy,
132                 .maxlen         = sizeof(unsigned int),
133                 .mode           = 0644,
134                 .proc_handler   = &proc_dointvec_minmax,
135                 .strategy       = &sysctl_intvec,
136                 .extra1         = &min_memreg,
137                 .extra2         = &max_memreg,
138         },
139         {
140                 .ctl_name       = CTL_UNNUMBERED,
141                 .procname       = "rdma_pad_optimize",
142                 .data           = &xprt_rdma_pad_optimize,
143                 .maxlen         = sizeof(unsigned int),
144                 .mode           = 0644,
145                 .proc_handler   = &proc_dointvec,
146         },
147         {
148                 .ctl_name = 0,
149         },
150 };
151
152 static ctl_table sunrpc_table[] = {
153         {
154                 .ctl_name       = CTL_SUNRPC,
155                 .procname       = "sunrpc",
156                 .mode           = 0555,
157                 .child          = xr_tunables_table
158         },
159         {
160                 .ctl_name = 0,
161         },
162 };
163
164 #endif
165
166 static struct rpc_xprt_ops xprt_rdma_procs;     /* forward reference */
167
168 static void
169 xprt_rdma_format_addresses(struct rpc_xprt *xprt)
170 {
171         struct sockaddr_in *addr = (struct sockaddr_in *)
172                                         &rpcx_to_rdmad(xprt).addr;
173         char *buf;
174
175         buf = kzalloc(20, GFP_KERNEL);
176         if (buf)
177                 snprintf(buf, 20, "%pI4", &addr->sin_addr.s_addr);
178         xprt->address_strings[RPC_DISPLAY_ADDR] = buf;
179
180         buf = kzalloc(8, GFP_KERNEL);
181         if (buf)
182                 snprintf(buf, 8, "%u", ntohs(addr->sin_port));
183         xprt->address_strings[RPC_DISPLAY_PORT] = buf;
184
185         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
186
187         buf = kzalloc(48, GFP_KERNEL);
188         if (buf)
189                 snprintf(buf, 48, "addr=%pI4 port=%u proto=%s",
190                         &addr->sin_addr.s_addr,
191                         ntohs(addr->sin_port), "rdma");
192         xprt->address_strings[RPC_DISPLAY_ALL] = buf;
193
194         buf = kzalloc(10, GFP_KERNEL);
195         if (buf)
196                 snprintf(buf, 10, "%02x%02x%02x%02x",
197                         NIPQUAD(addr->sin_addr.s_addr));
198         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = buf;
199
200         buf = kzalloc(8, GFP_KERNEL);
201         if (buf)
202                 snprintf(buf, 8, "%4hx", ntohs(addr->sin_port));
203         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = buf;
204
205         buf = kzalloc(30, GFP_KERNEL);
206         if (buf)
207                 snprintf(buf, 30, "%pI4.%u.%u",
208                         &addr->sin_addr.s_addr,
209                         ntohs(addr->sin_port) >> 8,
210                         ntohs(addr->sin_port) & 0xff);
211         xprt->address_strings[RPC_DISPLAY_UNIVERSAL_ADDR] = buf;
212
213         /* netid */
214         xprt->address_strings[RPC_DISPLAY_NETID] = "rdma";
215 }
216
217 static void
218 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
219 {
220         unsigned int i;
221
222         for (i = 0; i < RPC_DISPLAY_MAX; i++)
223                 switch (i) {
224                 case RPC_DISPLAY_PROTO:
225                 case RPC_DISPLAY_NETID:
226                         continue;
227                 default:
228                         kfree(xprt->address_strings[i]);
229                 }
230 }
231
232 static void
233 xprt_rdma_connect_worker(struct work_struct *work)
234 {
235         struct rpcrdma_xprt *r_xprt =
236                 container_of(work, struct rpcrdma_xprt, rdma_connect.work);
237         struct rpc_xprt *xprt = &r_xprt->xprt;
238         int rc = 0;
239
240         if (!xprt->shutdown) {
241                 xprt_clear_connected(xprt);
242
243                 dprintk("RPC:       %s: %sconnect\n", __func__,
244                                 r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
245                 rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
246                 if (rc)
247                         goto out;
248         }
249         goto out_clear;
250
251 out:
252         xprt_wake_pending_tasks(xprt, rc);
253
254 out_clear:
255         dprintk("RPC:       %s: exit\n", __func__);
256         xprt_clear_connecting(xprt);
257 }
258
259 /*
260  * xprt_rdma_destroy
261  *
262  * Destroy the xprt.
263  * Free all memory associated with the object, including its own.
264  * NOTE: none of the *destroy methods free memory for their top-level
265  * objects, even though they may have allocated it (they do free
266  * private memory). It's up to the caller to handle it. In this
267  * case (RDMA transport), all structure memory is inlined with the
268  * struct rpcrdma_xprt.
269  */
270 static void
271 xprt_rdma_destroy(struct rpc_xprt *xprt)
272 {
273         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
274         int rc;
275
276         dprintk("RPC:       %s: called\n", __func__);
277
278         cancel_delayed_work(&r_xprt->rdma_connect);
279         flush_scheduled_work();
280
281         xprt_clear_connected(xprt);
282
283         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
284         rc = rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
285         if (rc)
286                 dprintk("RPC:       %s: rpcrdma_ep_destroy returned %i\n",
287                         __func__, rc);
288         rpcrdma_ia_close(&r_xprt->rx_ia);
289
290         xprt_rdma_free_addresses(xprt);
291
292         kfree(xprt->slot);
293         xprt->slot = NULL;
294         kfree(xprt);
295
296         dprintk("RPC:       %s: returning\n", __func__);
297
298         module_put(THIS_MODULE);
299 }
300
301 static const struct rpc_timeout xprt_rdma_default_timeout = {
302         .to_initval = 60 * HZ,
303         .to_maxval = 60 * HZ,
304 };
305
306 /**
307  * xprt_setup_rdma - Set up transport to use RDMA
308  *
309  * @args: rpc transport arguments
310  */
311 static struct rpc_xprt *
312 xprt_setup_rdma(struct xprt_create *args)
313 {
314         struct rpcrdma_create_data_internal cdata;
315         struct rpc_xprt *xprt;
316         struct rpcrdma_xprt *new_xprt;
317         struct rpcrdma_ep *new_ep;
318         struct sockaddr_in *sin;
319         int rc;
320
321         if (args->addrlen > sizeof(xprt->addr)) {
322                 dprintk("RPC:       %s: address too large\n", __func__);
323                 return ERR_PTR(-EBADF);
324         }
325
326         xprt = kzalloc(sizeof(struct rpcrdma_xprt), GFP_KERNEL);
327         if (xprt == NULL) {
328                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
329                         __func__);
330                 return ERR_PTR(-ENOMEM);
331         }
332
333         xprt->max_reqs = xprt_rdma_slot_table_entries;
334         xprt->slot = kcalloc(xprt->max_reqs,
335                                 sizeof(struct rpc_rqst), GFP_KERNEL);
336         if (xprt->slot == NULL) {
337                 dprintk("RPC:       %s: couldn't allocate %d slots\n",
338                         __func__, xprt->max_reqs);
339                 kfree(xprt);
340                 return ERR_PTR(-ENOMEM);
341         }
342
343         /* 60 second timeout, no retries */
344         xprt->timeout = &xprt_rdma_default_timeout;
345         xprt->bind_timeout = (60U * HZ);
346         xprt->connect_timeout = (60U * HZ);
347         xprt->reestablish_timeout = (5U * HZ);
348         xprt->idle_timeout = (5U * 60 * HZ);
349
350         xprt->resvport = 0;             /* privileged port not needed */
351         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
352         xprt->max_payload = RPCRDMA_MAX_DATA_SEGS * PAGE_SIZE;
353         xprt->ops = &xprt_rdma_procs;
354
355         /*
356          * Set up RDMA-specific connect data.
357          */
358
359         /* Put server RDMA address in local cdata */
360         memcpy(&cdata.addr, args->dstaddr, args->addrlen);
361
362         /* Ensure xprt->addr holds valid server TCP (not RDMA)
363          * address, for any side protocols which peek at it */
364         xprt->prot = IPPROTO_TCP;
365         xprt->addrlen = args->addrlen;
366         memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
367
368         sin = (struct sockaddr_in *)&cdata.addr;
369         if (ntohs(sin->sin_port) != 0)
370                 xprt_set_bound(xprt);
371
372         dprintk("RPC:       %s: %pI4:%u\n",
373                 __func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port));
374
375         /* Set max requests */
376         cdata.max_requests = xprt->max_reqs;
377
378         /* Set some length limits */
379         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
380         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
381
382         cdata.inline_wsize = xprt_rdma_max_inline_write;
383         if (cdata.inline_wsize > cdata.wsize)
384                 cdata.inline_wsize = cdata.wsize;
385
386         cdata.inline_rsize = xprt_rdma_max_inline_read;
387         if (cdata.inline_rsize > cdata.rsize)
388                 cdata.inline_rsize = cdata.rsize;
389
390         cdata.padding = xprt_rdma_inline_write_padding;
391
392         /*
393          * Create new transport instance, which includes initialized
394          *  o ia
395          *  o endpoint
396          *  o buffers
397          */
398
399         new_xprt = rpcx_to_rdmax(xprt);
400
401         rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
402                                 xprt_rdma_memreg_strategy);
403         if (rc)
404                 goto out1;
405
406         /*
407          * initialize and create ep
408          */
409         new_xprt->rx_data = cdata;
410         new_ep = &new_xprt->rx_ep;
411         new_ep->rep_remote_addr = cdata.addr;
412
413         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
414                                 &new_xprt->rx_ia, &new_xprt->rx_data);
415         if (rc)
416                 goto out2;
417
418         /*
419          * Allocate pre-registered send and receive buffers for headers and
420          * any inline data. Also specify any padding which will be provided
421          * from a preregistered zero buffer.
422          */
423         rc = rpcrdma_buffer_create(&new_xprt->rx_buf, new_ep, &new_xprt->rx_ia,
424                                 &new_xprt->rx_data);
425         if (rc)
426                 goto out3;
427
428         /*
429          * Register a callback for connection events. This is necessary because
430          * connection loss notification is async. We also catch connection loss
431          * when reaping receives.
432          */
433         INIT_DELAYED_WORK(&new_xprt->rdma_connect, xprt_rdma_connect_worker);
434         new_ep->rep_func = rpcrdma_conn_func;
435         new_ep->rep_xprt = xprt;
436
437         xprt_rdma_format_addresses(xprt);
438
439         if (!try_module_get(THIS_MODULE))
440                 goto out4;
441
442         return xprt;
443
444 out4:
445         xprt_rdma_free_addresses(xprt);
446         rc = -EINVAL;
447 out3:
448         (void) rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
449 out2:
450         rpcrdma_ia_close(&new_xprt->rx_ia);
451 out1:
452         kfree(xprt->slot);
453         kfree(xprt);
454         return ERR_PTR(rc);
455 }
456
457 /*
458  * Close a connection, during shutdown or timeout/reconnect
459  */
460 static void
461 xprt_rdma_close(struct rpc_xprt *xprt)
462 {
463         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
464
465         dprintk("RPC:       %s: closing\n", __func__);
466         if (r_xprt->rx_ep.rep_connected > 0)
467                 xprt->reestablish_timeout = 0;
468         xprt_disconnect_done(xprt);
469         (void) rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
470 }
471
472 static void
473 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
474 {
475         struct sockaddr_in *sap;
476
477         sap = (struct sockaddr_in *)&xprt->addr;
478         sap->sin_port = htons(port);
479         sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
480         sap->sin_port = htons(port);
481         dprintk("RPC:       %s: %u\n", __func__, port);
482 }
483
484 static void
485 xprt_rdma_connect(struct rpc_task *task)
486 {
487         struct rpc_xprt *xprt = (struct rpc_xprt *)task->tk_xprt;
488         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
489
490         if (!xprt_test_and_set_connecting(xprt)) {
491                 if (r_xprt->rx_ep.rep_connected != 0) {
492                         /* Reconnect */
493                         schedule_delayed_work(&r_xprt->rdma_connect,
494                                 xprt->reestablish_timeout);
495                         xprt->reestablish_timeout <<= 1;
496                         if (xprt->reestablish_timeout > (30 * HZ))
497                                 xprt->reestablish_timeout = (30 * HZ);
498                         else if (xprt->reestablish_timeout < (5 * HZ))
499                                 xprt->reestablish_timeout = (5 * HZ);
500                 } else {
501                         schedule_delayed_work(&r_xprt->rdma_connect, 0);
502                         if (!RPC_IS_ASYNC(task))
503                                 flush_scheduled_work();
504                 }
505         }
506 }
507
508 static int
509 xprt_rdma_reserve_xprt(struct rpc_task *task)
510 {
511         struct rpc_xprt *xprt = task->tk_xprt;
512         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
513         int credits = atomic_read(&r_xprt->rx_buf.rb_credits);
514
515         /* == RPC_CWNDSCALE @ init, but *after* setup */
516         if (r_xprt->rx_buf.rb_cwndscale == 0UL) {
517                 r_xprt->rx_buf.rb_cwndscale = xprt->cwnd;
518                 dprintk("RPC:       %s: cwndscale %lu\n", __func__,
519                         r_xprt->rx_buf.rb_cwndscale);
520                 BUG_ON(r_xprt->rx_buf.rb_cwndscale <= 0);
521         }
522         xprt->cwnd = credits * r_xprt->rx_buf.rb_cwndscale;
523         return xprt_reserve_xprt_cong(task);
524 }
525
526 /*
527  * The RDMA allocate/free functions need the task structure as a place
528  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
529  * sequence. For this reason, the recv buffers are attached to send
530  * buffers for portions of the RPC. Note that the RPC layer allocates
531  * both send and receive buffers in the same call. We may register
532  * the receive buffer portion when using reply chunks.
533  */
534 static void *
535 xprt_rdma_allocate(struct rpc_task *task, size_t size)
536 {
537         struct rpc_xprt *xprt = task->tk_xprt;
538         struct rpcrdma_req *req, *nreq;
539
540         req = rpcrdma_buffer_get(&rpcx_to_rdmax(xprt)->rx_buf);
541         BUG_ON(NULL == req);
542
543         if (size > req->rl_size) {
544                 dprintk("RPC:       %s: size %zd too large for buffer[%zd]: "
545                         "prog %d vers %d proc %d\n",
546                         __func__, size, req->rl_size,
547                         task->tk_client->cl_prog, task->tk_client->cl_vers,
548                         task->tk_msg.rpc_proc->p_proc);
549                 /*
550                  * Outgoing length shortage. Our inline write max must have
551                  * been configured to perform direct i/o.
552                  *
553                  * This is therefore a large metadata operation, and the
554                  * allocate call was made on the maximum possible message,
555                  * e.g. containing long filename(s) or symlink data. In
556                  * fact, while these metadata operations *might* carry
557                  * large outgoing payloads, they rarely *do*. However, we
558                  * have to commit to the request here, so reallocate and
559                  * register it now. The data path will never require this
560                  * reallocation.
561                  *
562                  * If the allocation or registration fails, the RPC framework
563                  * will (doggedly) retry.
564                  */
565                 if (rpcx_to_rdmax(xprt)->rx_ia.ri_memreg_strategy ==
566                                 RPCRDMA_BOUNCEBUFFERS) {
567                         /* forced to "pure inline" */
568                         dprintk("RPC:       %s: too much data (%zd) for inline "
569                                         "(r/w max %d/%d)\n", __func__, size,
570                                         rpcx_to_rdmad(xprt).inline_rsize,
571                                         rpcx_to_rdmad(xprt).inline_wsize);
572                         size = req->rl_size;
573                         rpc_exit(task, -EIO);           /* fail the operation */
574                         rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
575                         goto out;
576                 }
577                 if (task->tk_flags & RPC_TASK_SWAPPER)
578                         nreq = kmalloc(sizeof *req + size, GFP_ATOMIC);
579                 else
580                         nreq = kmalloc(sizeof *req + size, GFP_NOFS);
581                 if (nreq == NULL)
582                         goto outfail;
583
584                 if (rpcrdma_register_internal(&rpcx_to_rdmax(xprt)->rx_ia,
585                                 nreq->rl_base, size + sizeof(struct rpcrdma_req)
586                                 - offsetof(struct rpcrdma_req, rl_base),
587                                 &nreq->rl_handle, &nreq->rl_iov)) {
588                         kfree(nreq);
589                         goto outfail;
590                 }
591                 rpcx_to_rdmax(xprt)->rx_stats.hardway_register_count += size;
592                 nreq->rl_size = size;
593                 nreq->rl_niovs = 0;
594                 nreq->rl_nchunks = 0;
595                 nreq->rl_buffer = (struct rpcrdma_buffer *)req;
596                 nreq->rl_reply = req->rl_reply;
597                 memcpy(nreq->rl_segments,
598                         req->rl_segments, sizeof nreq->rl_segments);
599                 /* flag the swap with an unused field */
600                 nreq->rl_iov.length = 0;
601                 req->rl_reply = NULL;
602                 req = nreq;
603         }
604         dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
605 out:
606         req->rl_connect_cookie = 0;     /* our reserved value */
607         return req->rl_xdr_buf;
608
609 outfail:
610         rpcrdma_buffer_put(req);
611         rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
612         return NULL;
613 }
614
615 /*
616  * This function returns all RDMA resources to the pool.
617  */
618 static void
619 xprt_rdma_free(void *buffer)
620 {
621         struct rpcrdma_req *req;
622         struct rpcrdma_xprt *r_xprt;
623         struct rpcrdma_rep *rep;
624         int i;
625
626         if (buffer == NULL)
627                 return;
628
629         req = container_of(buffer, struct rpcrdma_req, rl_xdr_buf[0]);
630         if (req->rl_iov.length == 0) {  /* see allocate above */
631                 r_xprt = container_of(((struct rpcrdma_req *) req->rl_buffer)->rl_buffer,
632                                       struct rpcrdma_xprt, rx_buf);
633         } else
634                 r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
635         rep = req->rl_reply;
636
637         dprintk("RPC:       %s: called on 0x%p%s\n",
638                 __func__, rep, (rep && rep->rr_func) ? " (with waiter)" : "");
639
640         /*
641          * Finish the deregistration. When using mw bind, this was
642          * begun in rpcrdma_reply_handler(). In all other modes, we
643          * do it here, in thread context. The process is considered
644          * complete when the rr_func vector becomes NULL - this
645          * was put in place during rpcrdma_reply_handler() - the wait
646          * call below will not block if the dereg is "done". If
647          * interrupted, our framework will clean up.
648          */
649         for (i = 0; req->rl_nchunks;) {
650                 --req->rl_nchunks;
651                 i += rpcrdma_deregister_external(
652                         &req->rl_segments[i], r_xprt, NULL);
653         }
654
655         if (rep && wait_event_interruptible(rep->rr_unbind, !rep->rr_func)) {
656                 rep->rr_func = NULL;    /* abandon the callback */
657                 req->rl_reply = NULL;
658         }
659
660         if (req->rl_iov.length == 0) {  /* see allocate above */
661                 struct rpcrdma_req *oreq = (struct rpcrdma_req *)req->rl_buffer;
662                 oreq->rl_reply = req->rl_reply;
663                 (void) rpcrdma_deregister_internal(&r_xprt->rx_ia,
664                                                    req->rl_handle,
665                                                    &req->rl_iov);
666                 kfree(req);
667                 req = oreq;
668         }
669
670         /* Put back request+reply buffers */
671         rpcrdma_buffer_put(req);
672 }
673
674 /*
675  * send_request invokes the meat of RPC RDMA. It must do the following:
676  *  1.  Marshal the RPC request into an RPC RDMA request, which means
677  *      putting a header in front of data, and creating IOVs for RDMA
678  *      from those in the request.
679  *  2.  In marshaling, detect opportunities for RDMA, and use them.
680  *  3.  Post a recv message to set up asynch completion, then send
681  *      the request (rpcrdma_ep_post).
682  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
683  */
684
685 static int
686 xprt_rdma_send_request(struct rpc_task *task)
687 {
688         struct rpc_rqst *rqst = task->tk_rqstp;
689         struct rpc_xprt *xprt = task->tk_xprt;
690         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
691         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
692
693         /* marshal the send itself */
694         if (req->rl_niovs == 0 && rpcrdma_marshal_req(rqst) != 0) {
695                 r_xprt->rx_stats.failed_marshal_count++;
696                 dprintk("RPC:       %s: rpcrdma_marshal_req failed\n",
697                         __func__);
698                 return -EIO;
699         }
700
701         if (req->rl_reply == NULL)              /* e.g. reconnection */
702                 rpcrdma_recv_buffer_get(req);
703
704         if (req->rl_reply) {
705                 req->rl_reply->rr_func = rpcrdma_reply_handler;
706                 /* this need only be done once, but... */
707                 req->rl_reply->rr_xprt = xprt;
708         }
709
710         /* Must suppress retransmit to maintain credits */
711         if (req->rl_connect_cookie == xprt->connect_cookie)
712                 goto drop_connection;
713         req->rl_connect_cookie = xprt->connect_cookie;
714
715         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
716                 goto drop_connection;
717
718         task->tk_bytes_sent += rqst->rq_snd_buf.len;
719         rqst->rq_bytes_sent = 0;
720         return 0;
721
722 drop_connection:
723         xprt_disconnect_done(xprt);
724         return -ENOTCONN;       /* implies disconnect */
725 }
726
727 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
728 {
729         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
730         long idle_time = 0;
731
732         if (xprt_connected(xprt))
733                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
734
735         seq_printf(seq,
736           "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
737           "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
738
739            0,   /* need a local port? */
740            xprt->stat.bind_count,
741            xprt->stat.connect_count,
742            xprt->stat.connect_time,
743            idle_time,
744            xprt->stat.sends,
745            xprt->stat.recvs,
746            xprt->stat.bad_xids,
747            xprt->stat.req_u,
748            xprt->stat.bklog_u,
749
750            r_xprt->rx_stats.read_chunk_count,
751            r_xprt->rx_stats.write_chunk_count,
752            r_xprt->rx_stats.reply_chunk_count,
753            r_xprt->rx_stats.total_rdma_request,
754            r_xprt->rx_stats.total_rdma_reply,
755            r_xprt->rx_stats.pullup_copy_count,
756            r_xprt->rx_stats.fixup_copy_count,
757            r_xprt->rx_stats.hardway_register_count,
758            r_xprt->rx_stats.failed_marshal_count,
759            r_xprt->rx_stats.bad_reply_count);
760 }
761
762 /*
763  * Plumbing for rpc transport switch and kernel module
764  */
765
766 static struct rpc_xprt_ops xprt_rdma_procs = {
767         .reserve_xprt           = xprt_rdma_reserve_xprt,
768         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
769         .release_request        = xprt_release_rqst_cong,       /* ditto */
770         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
771         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
772         .set_port               = xprt_rdma_set_port,
773         .connect                = xprt_rdma_connect,
774         .buf_alloc              = xprt_rdma_allocate,
775         .buf_free               = xprt_rdma_free,
776         .send_request           = xprt_rdma_send_request,
777         .close                  = xprt_rdma_close,
778         .destroy                = xprt_rdma_destroy,
779         .print_stats            = xprt_rdma_print_stats
780 };
781
782 static struct xprt_class xprt_rdma = {
783         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
784         .name                   = "rdma",
785         .owner                  = THIS_MODULE,
786         .ident                  = XPRT_TRANSPORT_RDMA,
787         .setup                  = xprt_setup_rdma,
788 };
789
790 static void __exit xprt_rdma_cleanup(void)
791 {
792         int rc;
793
794         dprintk(KERN_INFO "RPCRDMA Module Removed, deregister RPC RDMA transport\n");
795 #ifdef RPC_DEBUG
796         if (sunrpc_table_header) {
797                 unregister_sysctl_table(sunrpc_table_header);
798                 sunrpc_table_header = NULL;
799         }
800 #endif
801         rc = xprt_unregister_transport(&xprt_rdma);
802         if (rc)
803                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
804                         __func__, rc);
805 }
806
807 static int __init xprt_rdma_init(void)
808 {
809         int rc;
810
811         rc = xprt_register_transport(&xprt_rdma);
812
813         if (rc)
814                 return rc;
815
816         dprintk(KERN_INFO "RPCRDMA Module Init, register RPC RDMA transport\n");
817
818         dprintk(KERN_INFO "Defaults:\n");
819         dprintk(KERN_INFO "\tSlots %d\n"
820                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
821                 xprt_rdma_slot_table_entries,
822                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
823         dprintk(KERN_INFO "\tPadding %d\n\tMemreg %d\n",
824                 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
825
826 #ifdef RPC_DEBUG
827         if (!sunrpc_table_header)
828                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
829 #endif
830         return 0;
831 }
832
833 module_init(xprt_rdma_init);
834 module_exit(xprt_rdma_cleanup);