SUNRPC: Move the test for XPRT_CONNECTING into xprt_connect()
[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/slab.h>
53 #include <linux/seq_file.h>
54
55 #include "xprt_rdma.h"
56
57 #ifdef RPC_DEBUG
58 # define RPCDBG_FACILITY        RPCDBG_TRANS
59 #endif
60
61 MODULE_LICENSE("Dual BSD/GPL");
62
63 MODULE_DESCRIPTION("RPC/RDMA Transport for Linux kernel NFS");
64 MODULE_AUTHOR("Network Appliance, Inc.");
65
66 /*
67  * tunables
68  */
69
70 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
71 static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
72 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
73 static unsigned int xprt_rdma_inline_write_padding;
74 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
75                 int xprt_rdma_pad_optimize = 0;
76
77 #ifdef RPC_DEBUG
78
79 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
80 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
81 static unsigned int zero;
82 static unsigned int max_padding = PAGE_SIZE;
83 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
84 static unsigned int max_memreg = RPCRDMA_LAST - 1;
85
86 static struct ctl_table_header *sunrpc_table_header;
87
88 static ctl_table xr_tunables_table[] = {
89         {
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                 .extra1         = &min_slot_table_size,
96                 .extra2         = &max_slot_table_size
97         },
98         {
99                 .procname       = "rdma_max_inline_read",
100                 .data           = &xprt_rdma_max_inline_read,
101                 .maxlen         = sizeof(unsigned int),
102                 .mode           = 0644,
103                 .proc_handler   = proc_dointvec,
104         },
105         {
106                 .procname       = "rdma_max_inline_write",
107                 .data           = &xprt_rdma_max_inline_write,
108                 .maxlen         = sizeof(unsigned int),
109                 .mode           = 0644,
110                 .proc_handler   = proc_dointvec,
111         },
112         {
113                 .procname       = "rdma_inline_write_padding",
114                 .data           = &xprt_rdma_inline_write_padding,
115                 .maxlen         = sizeof(unsigned int),
116                 .mode           = 0644,
117                 .proc_handler   = proc_dointvec_minmax,
118                 .extra1         = &zero,
119                 .extra2         = &max_padding,
120         },
121         {
122                 .procname       = "rdma_memreg_strategy",
123                 .data           = &xprt_rdma_memreg_strategy,
124                 .maxlen         = sizeof(unsigned int),
125                 .mode           = 0644,
126                 .proc_handler   = proc_dointvec_minmax,
127                 .extra1         = &min_memreg,
128                 .extra2         = &max_memreg,
129         },
130         {
131                 .procname       = "rdma_pad_optimize",
132                 .data           = &xprt_rdma_pad_optimize,
133                 .maxlen         = sizeof(unsigned int),
134                 .mode           = 0644,
135                 .proc_handler   = proc_dointvec,
136         },
137         { },
138 };
139
140 static ctl_table sunrpc_table[] = {
141         {
142                 .procname       = "sunrpc",
143                 .mode           = 0555,
144                 .child          = xr_tunables_table
145         },
146         { },
147 };
148
149 #endif
150
151 static struct rpc_xprt_ops xprt_rdma_procs;     /* forward reference */
152
153 static void
154 xprt_rdma_format_addresses(struct rpc_xprt *xprt)
155 {
156         struct sockaddr *sap = (struct sockaddr *)
157                                         &rpcx_to_rdmad(xprt).addr;
158         struct sockaddr_in *sin = (struct sockaddr_in *)sap;
159         char buf[64];
160
161         (void)rpc_ntop(sap, buf, sizeof(buf));
162         xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
163
164         snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
165         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
166
167         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
168
169         snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
170         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
171
172         snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
173         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
174
175         /* netid */
176         xprt->address_strings[RPC_DISPLAY_NETID] = "rdma";
177 }
178
179 static void
180 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
181 {
182         unsigned int i;
183
184         for (i = 0; i < RPC_DISPLAY_MAX; i++)
185                 switch (i) {
186                 case RPC_DISPLAY_PROTO:
187                 case RPC_DISPLAY_NETID:
188                         continue;
189                 default:
190                         kfree(xprt->address_strings[i]);
191                 }
192 }
193
194 static void
195 xprt_rdma_connect_worker(struct work_struct *work)
196 {
197         struct rpcrdma_xprt *r_xprt =
198                 container_of(work, struct rpcrdma_xprt, rdma_connect.work);
199         struct rpc_xprt *xprt = &r_xprt->xprt;
200         int rc = 0;
201
202         if (!xprt->shutdown) {
203                 xprt_clear_connected(xprt);
204
205                 dprintk("RPC:       %s: %sconnect\n", __func__,
206                                 r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
207                 rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
208                 if (rc)
209                         goto out;
210         }
211         goto out_clear;
212
213 out:
214         xprt_wake_pending_tasks(xprt, rc);
215
216 out_clear:
217         dprintk("RPC:       %s: exit\n", __func__);
218         xprt_clear_connecting(xprt);
219 }
220
221 /*
222  * xprt_rdma_destroy
223  *
224  * Destroy the xprt.
225  * Free all memory associated with the object, including its own.
226  * NOTE: none of the *destroy methods free memory for their top-level
227  * objects, even though they may have allocated it (they do free
228  * private memory). It's up to the caller to handle it. In this
229  * case (RDMA transport), all structure memory is inlined with the
230  * struct rpcrdma_xprt.
231  */
232 static void
233 xprt_rdma_destroy(struct rpc_xprt *xprt)
234 {
235         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
236         int rc;
237
238         dprintk("RPC:       %s: called\n", __func__);
239
240         cancel_delayed_work(&r_xprt->rdma_connect);
241         flush_scheduled_work();
242
243         xprt_clear_connected(xprt);
244
245         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
246         rc = rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
247         if (rc)
248                 dprintk("RPC:       %s: rpcrdma_ep_destroy returned %i\n",
249                         __func__, rc);
250         rpcrdma_ia_close(&r_xprt->rx_ia);
251
252         xprt_rdma_free_addresses(xprt);
253
254         kfree(xprt->slot);
255         xprt->slot = NULL;
256         kfree(xprt);
257
258         dprintk("RPC:       %s: returning\n", __func__);
259
260         module_put(THIS_MODULE);
261 }
262
263 static const struct rpc_timeout xprt_rdma_default_timeout = {
264         .to_initval = 60 * HZ,
265         .to_maxval = 60 * HZ,
266 };
267
268 /**
269  * xprt_setup_rdma - Set up transport to use RDMA
270  *
271  * @args: rpc transport arguments
272  */
273 static struct rpc_xprt *
274 xprt_setup_rdma(struct xprt_create *args)
275 {
276         struct rpcrdma_create_data_internal cdata;
277         struct rpc_xprt *xprt;
278         struct rpcrdma_xprt *new_xprt;
279         struct rpcrdma_ep *new_ep;
280         struct sockaddr_in *sin;
281         int rc;
282
283         if (args->addrlen > sizeof(xprt->addr)) {
284                 dprintk("RPC:       %s: address too large\n", __func__);
285                 return ERR_PTR(-EBADF);
286         }
287
288         xprt = kzalloc(sizeof(struct rpcrdma_xprt), GFP_KERNEL);
289         if (xprt == NULL) {
290                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
291                         __func__);
292                 return ERR_PTR(-ENOMEM);
293         }
294
295         xprt->max_reqs = xprt_rdma_slot_table_entries;
296         xprt->slot = kcalloc(xprt->max_reqs,
297                                 sizeof(struct rpc_rqst), GFP_KERNEL);
298         if (xprt->slot == NULL) {
299                 dprintk("RPC:       %s: couldn't allocate %d slots\n",
300                         __func__, xprt->max_reqs);
301                 kfree(xprt);
302                 return ERR_PTR(-ENOMEM);
303         }
304
305         /* 60 second timeout, no retries */
306         xprt->timeout = &xprt_rdma_default_timeout;
307         xprt->bind_timeout = (60U * HZ);
308         xprt->connect_timeout = (60U * HZ);
309         xprt->reestablish_timeout = (5U * HZ);
310         xprt->idle_timeout = (5U * 60 * HZ);
311
312         xprt->resvport = 0;             /* privileged port not needed */
313         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
314         xprt->max_payload = RPCRDMA_MAX_DATA_SEGS * PAGE_SIZE;
315         xprt->ops = &xprt_rdma_procs;
316
317         /*
318          * Set up RDMA-specific connect data.
319          */
320
321         /* Put server RDMA address in local cdata */
322         memcpy(&cdata.addr, args->dstaddr, args->addrlen);
323
324         /* Ensure xprt->addr holds valid server TCP (not RDMA)
325          * address, for any side protocols which peek at it */
326         xprt->prot = IPPROTO_TCP;
327         xprt->addrlen = args->addrlen;
328         memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
329
330         sin = (struct sockaddr_in *)&cdata.addr;
331         if (ntohs(sin->sin_port) != 0)
332                 xprt_set_bound(xprt);
333
334         dprintk("RPC:       %s: %pI4:%u\n",
335                 __func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port));
336
337         /* Set max requests */
338         cdata.max_requests = xprt->max_reqs;
339
340         /* Set some length limits */
341         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
342         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
343
344         cdata.inline_wsize = xprt_rdma_max_inline_write;
345         if (cdata.inline_wsize > cdata.wsize)
346                 cdata.inline_wsize = cdata.wsize;
347
348         cdata.inline_rsize = xprt_rdma_max_inline_read;
349         if (cdata.inline_rsize > cdata.rsize)
350                 cdata.inline_rsize = cdata.rsize;
351
352         cdata.padding = xprt_rdma_inline_write_padding;
353
354         /*
355          * Create new transport instance, which includes initialized
356          *  o ia
357          *  o endpoint
358          *  o buffers
359          */
360
361         new_xprt = rpcx_to_rdmax(xprt);
362
363         rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
364                                 xprt_rdma_memreg_strategy);
365         if (rc)
366                 goto out1;
367
368         /*
369          * initialize and create ep
370          */
371         new_xprt->rx_data = cdata;
372         new_ep = &new_xprt->rx_ep;
373         new_ep->rep_remote_addr = cdata.addr;
374
375         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
376                                 &new_xprt->rx_ia, &new_xprt->rx_data);
377         if (rc)
378                 goto out2;
379
380         /*
381          * Allocate pre-registered send and receive buffers for headers and
382          * any inline data. Also specify any padding which will be provided
383          * from a preregistered zero buffer.
384          */
385         rc = rpcrdma_buffer_create(&new_xprt->rx_buf, new_ep, &new_xprt->rx_ia,
386                                 &new_xprt->rx_data);
387         if (rc)
388                 goto out3;
389
390         /*
391          * Register a callback for connection events. This is necessary because
392          * connection loss notification is async. We also catch connection loss
393          * when reaping receives.
394          */
395         INIT_DELAYED_WORK(&new_xprt->rdma_connect, xprt_rdma_connect_worker);
396         new_ep->rep_func = rpcrdma_conn_func;
397         new_ep->rep_xprt = xprt;
398
399         xprt_rdma_format_addresses(xprt);
400
401         if (!try_module_get(THIS_MODULE))
402                 goto out4;
403
404         return xprt;
405
406 out4:
407         xprt_rdma_free_addresses(xprt);
408         rc = -EINVAL;
409 out3:
410         (void) rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
411 out2:
412         rpcrdma_ia_close(&new_xprt->rx_ia);
413 out1:
414         kfree(xprt->slot);
415         kfree(xprt);
416         return ERR_PTR(rc);
417 }
418
419 /*
420  * Close a connection, during shutdown or timeout/reconnect
421  */
422 static void
423 xprt_rdma_close(struct rpc_xprt *xprt)
424 {
425         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
426
427         dprintk("RPC:       %s: closing\n", __func__);
428         if (r_xprt->rx_ep.rep_connected > 0)
429                 xprt->reestablish_timeout = 0;
430         xprt_disconnect_done(xprt);
431         (void) rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
432 }
433
434 static void
435 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
436 {
437         struct sockaddr_in *sap;
438
439         sap = (struct sockaddr_in *)&xprt->addr;
440         sap->sin_port = htons(port);
441         sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
442         sap->sin_port = htons(port);
443         dprintk("RPC:       %s: %u\n", __func__, port);
444 }
445
446 static void
447 xprt_rdma_connect(struct rpc_task *task)
448 {
449         struct rpc_xprt *xprt = (struct rpc_xprt *)task->tk_xprt;
450         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
451
452         if (r_xprt->rx_ep.rep_connected != 0) {
453                 /* Reconnect */
454                 schedule_delayed_work(&r_xprt->rdma_connect,
455                         xprt->reestablish_timeout);
456                 xprt->reestablish_timeout <<= 1;
457                 if (xprt->reestablish_timeout > (30 * HZ))
458                         xprt->reestablish_timeout = (30 * HZ);
459                 else if (xprt->reestablish_timeout < (5 * HZ))
460                         xprt->reestablish_timeout = (5 * HZ);
461         } else {
462                 schedule_delayed_work(&r_xprt->rdma_connect, 0);
463                 if (!RPC_IS_ASYNC(task))
464                         flush_scheduled_work();
465         }
466 }
467
468 static int
469 xprt_rdma_reserve_xprt(struct rpc_task *task)
470 {
471         struct rpc_xprt *xprt = task->tk_xprt;
472         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
473         int credits = atomic_read(&r_xprt->rx_buf.rb_credits);
474
475         /* == RPC_CWNDSCALE @ init, but *after* setup */
476         if (r_xprt->rx_buf.rb_cwndscale == 0UL) {
477                 r_xprt->rx_buf.rb_cwndscale = xprt->cwnd;
478                 dprintk("RPC:       %s: cwndscale %lu\n", __func__,
479                         r_xprt->rx_buf.rb_cwndscale);
480                 BUG_ON(r_xprt->rx_buf.rb_cwndscale <= 0);
481         }
482         xprt->cwnd = credits * r_xprt->rx_buf.rb_cwndscale;
483         return xprt_reserve_xprt_cong(task);
484 }
485
486 /*
487  * The RDMA allocate/free functions need the task structure as a place
488  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
489  * sequence. For this reason, the recv buffers are attached to send
490  * buffers for portions of the RPC. Note that the RPC layer allocates
491  * both send and receive buffers in the same call. We may register
492  * the receive buffer portion when using reply chunks.
493  */
494 static void *
495 xprt_rdma_allocate(struct rpc_task *task, size_t size)
496 {
497         struct rpc_xprt *xprt = task->tk_xprt;
498         struct rpcrdma_req *req, *nreq;
499
500         req = rpcrdma_buffer_get(&rpcx_to_rdmax(xprt)->rx_buf);
501         BUG_ON(NULL == req);
502
503         if (size > req->rl_size) {
504                 dprintk("RPC:       %s: size %zd too large for buffer[%zd]: "
505                         "prog %d vers %d proc %d\n",
506                         __func__, size, req->rl_size,
507                         task->tk_client->cl_prog, task->tk_client->cl_vers,
508                         task->tk_msg.rpc_proc->p_proc);
509                 /*
510                  * Outgoing length shortage. Our inline write max must have
511                  * been configured to perform direct i/o.
512                  *
513                  * This is therefore a large metadata operation, and the
514                  * allocate call was made on the maximum possible message,
515                  * e.g. containing long filename(s) or symlink data. In
516                  * fact, while these metadata operations *might* carry
517                  * large outgoing payloads, they rarely *do*. However, we
518                  * have to commit to the request here, so reallocate and
519                  * register it now. The data path will never require this
520                  * reallocation.
521                  *
522                  * If the allocation or registration fails, the RPC framework
523                  * will (doggedly) retry.
524                  */
525                 if (rpcx_to_rdmax(xprt)->rx_ia.ri_memreg_strategy ==
526                                 RPCRDMA_BOUNCEBUFFERS) {
527                         /* forced to "pure inline" */
528                         dprintk("RPC:       %s: too much data (%zd) for inline "
529                                         "(r/w max %d/%d)\n", __func__, size,
530                                         rpcx_to_rdmad(xprt).inline_rsize,
531                                         rpcx_to_rdmad(xprt).inline_wsize);
532                         size = req->rl_size;
533                         rpc_exit(task, -EIO);           /* fail the operation */
534                         rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
535                         goto out;
536                 }
537                 if (task->tk_flags & RPC_TASK_SWAPPER)
538                         nreq = kmalloc(sizeof *req + size, GFP_ATOMIC);
539                 else
540                         nreq = kmalloc(sizeof *req + size, GFP_NOFS);
541                 if (nreq == NULL)
542                         goto outfail;
543
544                 if (rpcrdma_register_internal(&rpcx_to_rdmax(xprt)->rx_ia,
545                                 nreq->rl_base, size + sizeof(struct rpcrdma_req)
546                                 - offsetof(struct rpcrdma_req, rl_base),
547                                 &nreq->rl_handle, &nreq->rl_iov)) {
548                         kfree(nreq);
549                         goto outfail;
550                 }
551                 rpcx_to_rdmax(xprt)->rx_stats.hardway_register_count += size;
552                 nreq->rl_size = size;
553                 nreq->rl_niovs = 0;
554                 nreq->rl_nchunks = 0;
555                 nreq->rl_buffer = (struct rpcrdma_buffer *)req;
556                 nreq->rl_reply = req->rl_reply;
557                 memcpy(nreq->rl_segments,
558                         req->rl_segments, sizeof nreq->rl_segments);
559                 /* flag the swap with an unused field */
560                 nreq->rl_iov.length = 0;
561                 req->rl_reply = NULL;
562                 req = nreq;
563         }
564         dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
565 out:
566         req->rl_connect_cookie = 0;     /* our reserved value */
567         return req->rl_xdr_buf;
568
569 outfail:
570         rpcrdma_buffer_put(req);
571         rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
572         return NULL;
573 }
574
575 /*
576  * This function returns all RDMA resources to the pool.
577  */
578 static void
579 xprt_rdma_free(void *buffer)
580 {
581         struct rpcrdma_req *req;
582         struct rpcrdma_xprt *r_xprt;
583         struct rpcrdma_rep *rep;
584         int i;
585
586         if (buffer == NULL)
587                 return;
588
589         req = container_of(buffer, struct rpcrdma_req, rl_xdr_buf[0]);
590         if (req->rl_iov.length == 0) {  /* see allocate above */
591                 r_xprt = container_of(((struct rpcrdma_req *) req->rl_buffer)->rl_buffer,
592                                       struct rpcrdma_xprt, rx_buf);
593         } else
594                 r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
595         rep = req->rl_reply;
596
597         dprintk("RPC:       %s: called on 0x%p%s\n",
598                 __func__, rep, (rep && rep->rr_func) ? " (with waiter)" : "");
599
600         /*
601          * Finish the deregistration. When using mw bind, this was
602          * begun in rpcrdma_reply_handler(). In all other modes, we
603          * do it here, in thread context. The process is considered
604          * complete when the rr_func vector becomes NULL - this
605          * was put in place during rpcrdma_reply_handler() - the wait
606          * call below will not block if the dereg is "done". If
607          * interrupted, our framework will clean up.
608          */
609         for (i = 0; req->rl_nchunks;) {
610                 --req->rl_nchunks;
611                 i += rpcrdma_deregister_external(
612                         &req->rl_segments[i], r_xprt, NULL);
613         }
614
615         if (rep && wait_event_interruptible(rep->rr_unbind, !rep->rr_func)) {
616                 rep->rr_func = NULL;    /* abandon the callback */
617                 req->rl_reply = NULL;
618         }
619
620         if (req->rl_iov.length == 0) {  /* see allocate above */
621                 struct rpcrdma_req *oreq = (struct rpcrdma_req *)req->rl_buffer;
622                 oreq->rl_reply = req->rl_reply;
623                 (void) rpcrdma_deregister_internal(&r_xprt->rx_ia,
624                                                    req->rl_handle,
625                                                    &req->rl_iov);
626                 kfree(req);
627                 req = oreq;
628         }
629
630         /* Put back request+reply buffers */
631         rpcrdma_buffer_put(req);
632 }
633
634 /*
635  * send_request invokes the meat of RPC RDMA. It must do the following:
636  *  1.  Marshal the RPC request into an RPC RDMA request, which means
637  *      putting a header in front of data, and creating IOVs for RDMA
638  *      from those in the request.
639  *  2.  In marshaling, detect opportunities for RDMA, and use them.
640  *  3.  Post a recv message to set up asynch completion, then send
641  *      the request (rpcrdma_ep_post).
642  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
643  */
644
645 static int
646 xprt_rdma_send_request(struct rpc_task *task)
647 {
648         struct rpc_rqst *rqst = task->tk_rqstp;
649         struct rpc_xprt *xprt = task->tk_xprt;
650         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
651         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
652
653         /* marshal the send itself */
654         if (req->rl_niovs == 0 && rpcrdma_marshal_req(rqst) != 0) {
655                 r_xprt->rx_stats.failed_marshal_count++;
656                 dprintk("RPC:       %s: rpcrdma_marshal_req failed\n",
657                         __func__);
658                 return -EIO;
659         }
660
661         if (req->rl_reply == NULL)              /* e.g. reconnection */
662                 rpcrdma_recv_buffer_get(req);
663
664         if (req->rl_reply) {
665                 req->rl_reply->rr_func = rpcrdma_reply_handler;
666                 /* this need only be done once, but... */
667                 req->rl_reply->rr_xprt = xprt;
668         }
669
670         /* Must suppress retransmit to maintain credits */
671         if (req->rl_connect_cookie == xprt->connect_cookie)
672                 goto drop_connection;
673         req->rl_connect_cookie = xprt->connect_cookie;
674
675         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
676                 goto drop_connection;
677
678         task->tk_bytes_sent += rqst->rq_snd_buf.len;
679         rqst->rq_bytes_sent = 0;
680         return 0;
681
682 drop_connection:
683         xprt_disconnect_done(xprt);
684         return -ENOTCONN;       /* implies disconnect */
685 }
686
687 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
688 {
689         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
690         long idle_time = 0;
691
692         if (xprt_connected(xprt))
693                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
694
695         seq_printf(seq,
696           "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
697           "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
698
699            0,   /* need a local port? */
700            xprt->stat.bind_count,
701            xprt->stat.connect_count,
702            xprt->stat.connect_time,
703            idle_time,
704            xprt->stat.sends,
705            xprt->stat.recvs,
706            xprt->stat.bad_xids,
707            xprt->stat.req_u,
708            xprt->stat.bklog_u,
709
710            r_xprt->rx_stats.read_chunk_count,
711            r_xprt->rx_stats.write_chunk_count,
712            r_xprt->rx_stats.reply_chunk_count,
713            r_xprt->rx_stats.total_rdma_request,
714            r_xprt->rx_stats.total_rdma_reply,
715            r_xprt->rx_stats.pullup_copy_count,
716            r_xprt->rx_stats.fixup_copy_count,
717            r_xprt->rx_stats.hardway_register_count,
718            r_xprt->rx_stats.failed_marshal_count,
719            r_xprt->rx_stats.bad_reply_count);
720 }
721
722 /*
723  * Plumbing for rpc transport switch and kernel module
724  */
725
726 static struct rpc_xprt_ops xprt_rdma_procs = {
727         .reserve_xprt           = xprt_rdma_reserve_xprt,
728         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
729         .release_request        = xprt_release_rqst_cong,       /* ditto */
730         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
731         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
732         .set_port               = xprt_rdma_set_port,
733         .connect                = xprt_rdma_connect,
734         .buf_alloc              = xprt_rdma_allocate,
735         .buf_free               = xprt_rdma_free,
736         .send_request           = xprt_rdma_send_request,
737         .close                  = xprt_rdma_close,
738         .destroy                = xprt_rdma_destroy,
739         .print_stats            = xprt_rdma_print_stats
740 };
741
742 static struct xprt_class xprt_rdma = {
743         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
744         .name                   = "rdma",
745         .owner                  = THIS_MODULE,
746         .ident                  = XPRT_TRANSPORT_RDMA,
747         .setup                  = xprt_setup_rdma,
748 };
749
750 static void __exit xprt_rdma_cleanup(void)
751 {
752         int rc;
753
754         dprintk(KERN_INFO "RPCRDMA Module Removed, deregister RPC RDMA transport\n");
755 #ifdef RPC_DEBUG
756         if (sunrpc_table_header) {
757                 unregister_sysctl_table(sunrpc_table_header);
758                 sunrpc_table_header = NULL;
759         }
760 #endif
761         rc = xprt_unregister_transport(&xprt_rdma);
762         if (rc)
763                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
764                         __func__, rc);
765 }
766
767 static int __init xprt_rdma_init(void)
768 {
769         int rc;
770
771         rc = xprt_register_transport(&xprt_rdma);
772
773         if (rc)
774                 return rc;
775
776         dprintk(KERN_INFO "RPCRDMA Module Init, register RPC RDMA transport\n");
777
778         dprintk(KERN_INFO "Defaults:\n");
779         dprintk(KERN_INFO "\tSlots %d\n"
780                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
781                 xprt_rdma_slot_table_entries,
782                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
783         dprintk(KERN_INFO "\tPadding %d\n\tMemreg %d\n",
784                 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
785
786 #ifdef RPC_DEBUG
787         if (!sunrpc_table_header)
788                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
789 #endif
790         return 0;
791 }
792
793 module_init(xprt_rdma_init);
794 module_exit(xprt_rdma_cleanup);