Merge tag 'nfs-for-4.21-1' of git://git.linux-nfs.org/projects/anna/linux-nfs
[sfrench/cifs-2.6.git] / net / sunrpc / xprtrdma / transport.c
1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
3  * Copyright (c) 2014-2017 Oracle.  All rights reserved.
4  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the BSD-type
10  * license below:
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  *
16  *      Redistributions of source code must retain the above copyright
17  *      notice, this list of conditions and the following disclaimer.
18  *
19  *      Redistributions in binary form must reproduce the above
20  *      copyright notice, this list of conditions and the following
21  *      disclaimer in the documentation and/or other materials provided
22  *      with the distribution.
23  *
24  *      Neither the name of the Network Appliance, Inc. nor the names of
25  *      its contributors may be used to endorse or promote products
26  *      derived from this software without specific prior written
27  *      permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40  */
41
42 /*
43  * transport.c
44  *
45  * This file contains the top-level implementation of an RPC RDMA
46  * transport.
47  *
48  * Naming convention: functions beginning with xprt_ are part of the
49  * transport switch. All others are RPC RDMA internal.
50  */
51
52 #include <linux/module.h>
53 #include <linux/slab.h>
54 #include <linux/seq_file.h>
55 #include <linux/smp.h>
56
57 #include <linux/sunrpc/addr.h>
58 #include <linux/sunrpc/svc_rdma.h>
59
60 #include "xprt_rdma.h"
61 #include <trace/events/rpcrdma.h>
62
63 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
64 # define RPCDBG_FACILITY        RPCDBG_TRANS
65 #endif
66
67 /*
68  * tunables
69  */
70
71 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
72 unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
73 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
74 unsigned int xprt_rdma_memreg_strategy          = RPCRDMA_FRWR;
75 int xprt_rdma_pad_optimize;
76
77 #if IS_ENABLED(CONFIG_SUNRPC_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 min_inline_size = RPCRDMA_MIN_INLINE;
82 static unsigned int max_inline_size = RPCRDMA_MAX_INLINE;
83 static unsigned int zero;
84 static unsigned int max_padding = PAGE_SIZE;
85 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
86 static unsigned int max_memreg = RPCRDMA_LAST - 1;
87 static unsigned int dummy;
88
89 static struct ctl_table_header *sunrpc_table_header;
90
91 static struct ctl_table xr_tunables_table[] = {
92         {
93                 .procname       = "rdma_slot_table_entries",
94                 .data           = &xprt_rdma_slot_table_entries,
95                 .maxlen         = sizeof(unsigned int),
96                 .mode           = 0644,
97                 .proc_handler   = proc_dointvec_minmax,
98                 .extra1         = &min_slot_table_size,
99                 .extra2         = &max_slot_table_size
100         },
101         {
102                 .procname       = "rdma_max_inline_read",
103                 .data           = &xprt_rdma_max_inline_read,
104                 .maxlen         = sizeof(unsigned int),
105                 .mode           = 0644,
106                 .proc_handler   = proc_dointvec_minmax,
107                 .extra1         = &min_inline_size,
108                 .extra2         = &max_inline_size,
109         },
110         {
111                 .procname       = "rdma_max_inline_write",
112                 .data           = &xprt_rdma_max_inline_write,
113                 .maxlen         = sizeof(unsigned int),
114                 .mode           = 0644,
115                 .proc_handler   = proc_dointvec_minmax,
116                 .extra1         = &min_inline_size,
117                 .extra2         = &max_inline_size,
118         },
119         {
120                 .procname       = "rdma_inline_write_padding",
121                 .data           = &dummy,
122                 .maxlen         = sizeof(unsigned int),
123                 .mode           = 0644,
124                 .proc_handler   = proc_dointvec_minmax,
125                 .extra1         = &zero,
126                 .extra2         = &max_padding,
127         },
128         {
129                 .procname       = "rdma_memreg_strategy",
130                 .data           = &xprt_rdma_memreg_strategy,
131                 .maxlen         = sizeof(unsigned int),
132                 .mode           = 0644,
133                 .proc_handler   = proc_dointvec_minmax,
134                 .extra1         = &min_memreg,
135                 .extra2         = &max_memreg,
136         },
137         {
138                 .procname       = "rdma_pad_optimize",
139                 .data           = &xprt_rdma_pad_optimize,
140                 .maxlen         = sizeof(unsigned int),
141                 .mode           = 0644,
142                 .proc_handler   = proc_dointvec,
143         },
144         { },
145 };
146
147 static struct ctl_table sunrpc_table[] = {
148         {
149                 .procname       = "sunrpc",
150                 .mode           = 0555,
151                 .child          = xr_tunables_table
152         },
153         { },
154 };
155
156 #endif
157
158 static const struct rpc_xprt_ops xprt_rdma_procs;
159
160 static void
161 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
162 {
163         struct sockaddr_in *sin = (struct sockaddr_in *)sap;
164         char buf[20];
165
166         snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
167         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
168
169         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
170 }
171
172 static void
173 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
174 {
175         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
176         char buf[40];
177
178         snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
179         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
180
181         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
182 }
183
184 void
185 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
186 {
187         char buf[128];
188
189         switch (sap->sa_family) {
190         case AF_INET:
191                 xprt_rdma_format_addresses4(xprt, sap);
192                 break;
193         case AF_INET6:
194                 xprt_rdma_format_addresses6(xprt, sap);
195                 break;
196         default:
197                 pr_err("rpcrdma: Unrecognized address family\n");
198                 return;
199         }
200
201         (void)rpc_ntop(sap, buf, sizeof(buf));
202         xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
203
204         snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
205         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
206
207         snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
208         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
209
210         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
211 }
212
213 void
214 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
215 {
216         unsigned int i;
217
218         for (i = 0; i < RPC_DISPLAY_MAX; i++)
219                 switch (i) {
220                 case RPC_DISPLAY_PROTO:
221                 case RPC_DISPLAY_NETID:
222                         continue;
223                 default:
224                         kfree(xprt->address_strings[i]);
225                 }
226 }
227
228 /**
229  * xprt_rdma_connect_worker - establish connection in the background
230  * @work: worker thread context
231  *
232  * Requester holds the xprt's send lock to prevent activity on this
233  * transport while a fresh connection is being established. RPC tasks
234  * sleep on the xprt's pending queue waiting for connect to complete.
235  */
236 static void
237 xprt_rdma_connect_worker(struct work_struct *work)
238 {
239         struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
240                                                    rx_connect_worker.work);
241         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
242         int rc;
243
244         rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
245         xprt_clear_connecting(xprt);
246         if (r_xprt->rx_ep.rep_connected > 0) {
247                 if (!xprt_test_and_set_connected(xprt)) {
248                         xprt->stat.connect_count++;
249                         xprt->stat.connect_time += (long)jiffies -
250                                                    xprt->stat.connect_start;
251                         xprt_wake_pending_tasks(xprt, -EAGAIN);
252                 }
253         } else {
254                 if (xprt_test_and_clear_connected(xprt))
255                         xprt_wake_pending_tasks(xprt, rc);
256         }
257 }
258
259 /**
260  * xprt_rdma_inject_disconnect - inject a connection fault
261  * @xprt: transport context
262  *
263  * If @xprt is connected, disconnect it to simulate spurious connection
264  * loss.
265  */
266 static void
267 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
268 {
269         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
270
271         trace_xprtrdma_op_inject_dsc(r_xprt);
272         rdma_disconnect(r_xprt->rx_ia.ri_id);
273 }
274
275 /**
276  * xprt_rdma_destroy - Full tear down of transport
277  * @xprt: doomed transport context
278  *
279  * Caller guarantees there will be no more calls to us with
280  * this @xprt.
281  */
282 static void
283 xprt_rdma_destroy(struct rpc_xprt *xprt)
284 {
285         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
286
287         trace_xprtrdma_op_destroy(r_xprt);
288
289         cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
290
291         rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
292         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
293         rpcrdma_ia_close(&r_xprt->rx_ia);
294
295         xprt_rdma_free_addresses(xprt);
296         xprt_free(xprt);
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 *sap;
319         int rc;
320
321         if (args->addrlen > sizeof(xprt->addr))
322                 return ERR_PTR(-EBADF);
323
324         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt), 0, 0);
325         if (!xprt)
326                 return ERR_PTR(-ENOMEM);
327
328         /* 60 second timeout, no retries */
329         xprt->timeout = &xprt_rdma_default_timeout;
330         xprt->bind_timeout = RPCRDMA_BIND_TO;
331         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
332         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
333
334         xprt->resvport = 0;             /* privileged port not needed */
335         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
336         xprt->ops = &xprt_rdma_procs;
337
338         /*
339          * Set up RDMA-specific connect data.
340          */
341         sap = args->dstaddr;
342
343         /* Ensure xprt->addr holds valid server TCP (not RDMA)
344          * address, for any side protocols which peek at it */
345         xprt->prot = IPPROTO_TCP;
346         xprt->addrlen = args->addrlen;
347         memcpy(&xprt->addr, sap, xprt->addrlen);
348
349         if (rpc_get_port(sap))
350                 xprt_set_bound(xprt);
351         xprt_rdma_format_addresses(xprt, sap);
352
353         cdata.max_requests = xprt_rdma_slot_table_entries;
354
355         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
356         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
357
358         cdata.inline_wsize = xprt_rdma_max_inline_write;
359         if (cdata.inline_wsize > cdata.wsize)
360                 cdata.inline_wsize = cdata.wsize;
361
362         cdata.inline_rsize = xprt_rdma_max_inline_read;
363         if (cdata.inline_rsize > cdata.rsize)
364                 cdata.inline_rsize = cdata.rsize;
365
366         /*
367          * Create new transport instance, which includes initialized
368          *  o ia
369          *  o endpoint
370          *  o buffers
371          */
372
373         new_xprt = rpcx_to_rdmax(xprt);
374
375         rc = rpcrdma_ia_open(new_xprt);
376         if (rc)
377                 goto out1;
378
379         /*
380          * initialize and create ep
381          */
382         new_xprt->rx_data = cdata;
383         new_ep = &new_xprt->rx_ep;
384
385         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
386                                 &new_xprt->rx_ia, &new_xprt->rx_data);
387         if (rc)
388                 goto out2;
389
390         rc = rpcrdma_buffer_create(new_xprt);
391         if (rc)
392                 goto out3;
393
394         INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
395                           xprt_rdma_connect_worker);
396
397         xprt->max_payload = frwr_maxpages(new_xprt);
398         if (xprt->max_payload == 0)
399                 goto out4;
400         xprt->max_payload <<= PAGE_SHIFT;
401         dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
402                 __func__, xprt->max_payload);
403
404         if (!try_module_get(THIS_MODULE))
405                 goto out4;
406
407         dprintk("RPC:       %s: %s:%s\n", __func__,
408                 xprt->address_strings[RPC_DISPLAY_ADDR],
409                 xprt->address_strings[RPC_DISPLAY_PORT]);
410         trace_xprtrdma_create(new_xprt);
411         return xprt;
412
413 out4:
414         rpcrdma_buffer_destroy(&new_xprt->rx_buf);
415         rc = -ENODEV;
416 out3:
417         rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
418 out2:
419         rpcrdma_ia_close(&new_xprt->rx_ia);
420 out1:
421         trace_xprtrdma_op_destroy(new_xprt);
422         xprt_rdma_free_addresses(xprt);
423         xprt_free(xprt);
424         return ERR_PTR(rc);
425 }
426
427 /**
428  * xprt_rdma_close - close a transport connection
429  * @xprt: transport context
430  *
431  * Called during autoclose or device removal.
432  *
433  * Caller holds @xprt's send lock to prevent activity on this
434  * transport while the connection is torn down.
435  */
436 void xprt_rdma_close(struct rpc_xprt *xprt)
437 {
438         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
439         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
440         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
441
442         might_sleep();
443
444         trace_xprtrdma_op_close(r_xprt);
445
446         /* Prevent marshaling and sending of new requests */
447         xprt_clear_connected(xprt);
448
449         if (test_and_clear_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags)) {
450                 rpcrdma_ia_remove(ia);
451                 goto out;
452         }
453
454         if (ep->rep_connected == -ENODEV)
455                 return;
456         if (ep->rep_connected > 0)
457                 xprt->reestablish_timeout = 0;
458         rpcrdma_ep_disconnect(ep, ia);
459
460         /* Prepare @xprt for the next connection by reinitializing
461          * its credit grant to one (see RFC 8166, Section 3.3.3).
462          */
463         r_xprt->rx_buf.rb_credits = 1;
464         xprt->cwnd = RPC_CWNDSHIFT;
465
466 out:
467         ++xprt->connect_cookie;
468         xprt_disconnect_done(xprt);
469 }
470
471 /**
472  * xprt_rdma_set_port - update server port with rpcbind result
473  * @xprt: controlling RPC transport
474  * @port: new port value
475  *
476  * Transport connect status is unchanged.
477  */
478 static void
479 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
480 {
481         struct sockaddr *sap = (struct sockaddr *)&xprt->addr;
482         char buf[8];
483
484         dprintk("RPC:       %s: setting port for xprt %p (%s:%s) to %u\n",
485                 __func__, xprt,
486                 xprt->address_strings[RPC_DISPLAY_ADDR],
487                 xprt->address_strings[RPC_DISPLAY_PORT],
488                 port);
489
490         rpc_set_port(sap, port);
491
492         kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
493         snprintf(buf, sizeof(buf), "%u", port);
494         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
495
496         kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]);
497         snprintf(buf, sizeof(buf), "%4hx", port);
498         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
499 }
500
501 /**
502  * xprt_rdma_timer - invoked when an RPC times out
503  * @xprt: controlling RPC transport
504  * @task: RPC task that timed out
505  *
506  * Invoked when the transport is still connected, but an RPC
507  * retransmit timeout occurs.
508  *
509  * Since RDMA connections don't have a keep-alive, forcibly
510  * disconnect and retry to connect. This drives full
511  * detection of the network path, and retransmissions of
512  * all pending RPCs.
513  */
514 static void
515 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
516 {
517         xprt_force_disconnect(xprt);
518 }
519
520 /**
521  * xprt_rdma_connect - try to establish a transport connection
522  * @xprt: transport state
523  * @task: RPC scheduler context
524  *
525  */
526 static void
527 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
528 {
529         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
530
531         trace_xprtrdma_op_connect(r_xprt);
532         if (r_xprt->rx_ep.rep_connected != 0) {
533                 /* Reconnect */
534                 schedule_delayed_work(&r_xprt->rx_connect_worker,
535                                       xprt->reestablish_timeout);
536                 xprt->reestablish_timeout <<= 1;
537                 if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
538                         xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
539                 else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
540                         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
541         } else {
542                 schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
543                 if (!RPC_IS_ASYNC(task))
544                         flush_delayed_work(&r_xprt->rx_connect_worker);
545         }
546 }
547
548 /**
549  * xprt_rdma_alloc_slot - allocate an rpc_rqst
550  * @xprt: controlling RPC transport
551  * @task: RPC task requesting a fresh rpc_rqst
552  *
553  * tk_status values:
554  *      %0 if task->tk_rqstp points to a fresh rpc_rqst
555  *      %-EAGAIN if no rpc_rqst is available; queued on backlog
556  */
557 static void
558 xprt_rdma_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
559 {
560         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
561         struct rpcrdma_req *req;
562
563         req = rpcrdma_buffer_get(&r_xprt->rx_buf);
564         if (!req)
565                 goto out_sleep;
566         task->tk_rqstp = &req->rl_slot;
567         task->tk_status = 0;
568         return;
569
570 out_sleep:
571         rpc_sleep_on(&xprt->backlog, task, NULL);
572         task->tk_status = -EAGAIN;
573 }
574
575 /**
576  * xprt_rdma_free_slot - release an rpc_rqst
577  * @xprt: controlling RPC transport
578  * @rqst: rpc_rqst to release
579  *
580  */
581 static void
582 xprt_rdma_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *rqst)
583 {
584         memset(rqst, 0, sizeof(*rqst));
585         rpcrdma_buffer_put(rpcr_to_rdmar(rqst));
586         rpc_wake_up_next(&xprt->backlog);
587 }
588
589 static bool
590 rpcrdma_get_sendbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
591                     size_t size, gfp_t flags)
592 {
593         struct rpcrdma_regbuf *rb;
594
595         if (req->rl_sendbuf && rdmab_length(req->rl_sendbuf) >= size)
596                 return true;
597
598         rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
599         if (IS_ERR(rb))
600                 return false;
601
602         rpcrdma_free_regbuf(req->rl_sendbuf);
603         r_xprt->rx_stats.hardway_register_count += size;
604         req->rl_sendbuf = rb;
605         return true;
606 }
607
608 /* The rq_rcv_buf is used only if a Reply chunk is necessary.
609  * The decision to use a Reply chunk is made later in
610  * rpcrdma_marshal_req. This buffer is registered at that time.
611  *
612  * Otherwise, the associated RPC Reply arrives in a separate
613  * Receive buffer, arbitrarily chosen by the HCA. The buffer
614  * allocated here for the RPC Reply is not utilized in that
615  * case. See rpcrdma_inline_fixup.
616  *
617  * A regbuf is used here to remember the buffer size.
618  */
619 static bool
620 rpcrdma_get_recvbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
621                     size_t size, gfp_t flags)
622 {
623         struct rpcrdma_regbuf *rb;
624
625         if (req->rl_recvbuf && rdmab_length(req->rl_recvbuf) >= size)
626                 return true;
627
628         rb = rpcrdma_alloc_regbuf(size, DMA_NONE, flags);
629         if (IS_ERR(rb))
630                 return false;
631
632         rpcrdma_free_regbuf(req->rl_recvbuf);
633         r_xprt->rx_stats.hardway_register_count += size;
634         req->rl_recvbuf = rb;
635         return true;
636 }
637
638 /**
639  * xprt_rdma_allocate - allocate transport resources for an RPC
640  * @task: RPC task
641  *
642  * Return values:
643  *        0:    Success; rq_buffer points to RPC buffer to use
644  *   ENOMEM:    Out of memory, call again later
645  *      EIO:    A permanent error occurred, do not retry
646  */
647 static int
648 xprt_rdma_allocate(struct rpc_task *task)
649 {
650         struct rpc_rqst *rqst = task->tk_rqstp;
651         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
652         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
653         gfp_t flags;
654
655         flags = RPCRDMA_DEF_GFP;
656         if (RPC_IS_SWAPPER(task))
657                 flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
658
659         if (!rpcrdma_get_sendbuf(r_xprt, req, rqst->rq_callsize, flags))
660                 goto out_fail;
661         if (!rpcrdma_get_recvbuf(r_xprt, req, rqst->rq_rcvsize, flags))
662                 goto out_fail;
663
664         rqst->rq_buffer = req->rl_sendbuf->rg_base;
665         rqst->rq_rbuffer = req->rl_recvbuf->rg_base;
666         trace_xprtrdma_op_allocate(task, req);
667         return 0;
668
669 out_fail:
670         trace_xprtrdma_op_allocate(task, NULL);
671         return -ENOMEM;
672 }
673
674 /**
675  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
676  * @task: RPC task
677  *
678  * Caller guarantees rqst->rq_buffer is non-NULL.
679  */
680 static void
681 xprt_rdma_free(struct rpc_task *task)
682 {
683         struct rpc_rqst *rqst = task->tk_rqstp;
684         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
685         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
686
687         if (test_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags))
688                 rpcrdma_release_rqst(r_xprt, req);
689         trace_xprtrdma_op_free(task, req);
690 }
691
692 /**
693  * xprt_rdma_send_request - marshal and send an RPC request
694  * @rqst: RPC message in rq_snd_buf
695  *
696  * Caller holds the transport's write lock.
697  *
698  * Returns:
699  *      %0 if the RPC message has been sent
700  *      %-ENOTCONN if the caller should reconnect and call again
701  *      %-EAGAIN if the caller should call again
702  *      %-ENOBUFS if the caller should call again after a delay
703  *      %-EMSGSIZE if encoding ran out of buffer space. The request
704  *              was not sent. Do not try to send this message again.
705  *      %-EIO if an I/O error occurred. The request was not sent.
706  *              Do not try to send this message again.
707  */
708 static int
709 xprt_rdma_send_request(struct rpc_rqst *rqst)
710 {
711         struct rpc_xprt *xprt = rqst->rq_xprt;
712         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
713         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
714         int rc = 0;
715
716 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
717         if (unlikely(!rqst->rq_buffer))
718                 return xprt_rdma_bc_send_reply(rqst);
719 #endif  /* CONFIG_SUNRPC_BACKCHANNEL */
720
721         if (!xprt_connected(xprt))
722                 return -ENOTCONN;
723
724         if (!xprt_request_get_cong(xprt, rqst))
725                 return -EBADSLT;
726
727         rc = rpcrdma_marshal_req(r_xprt, rqst);
728         if (rc < 0)
729                 goto failed_marshal;
730
731         /* Must suppress retransmit to maintain credits */
732         if (rqst->rq_connect_cookie == xprt->connect_cookie)
733                 goto drop_connection;
734         rqst->rq_xtime = ktime_get();
735
736         __set_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags);
737         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
738                 goto drop_connection;
739
740         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
741         rqst->rq_bytes_sent = 0;
742
743         /* An RPC with no reply will throw off credit accounting,
744          * so drop the connection to reset the credit grant.
745          */
746         if (!rpc_reply_expected(rqst->rq_task))
747                 goto drop_connection;
748         return 0;
749
750 failed_marshal:
751         if (rc != -ENOTCONN)
752                 return rc;
753 drop_connection:
754         xprt_rdma_close(xprt);
755         return -ENOTCONN;
756 }
757
758 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
759 {
760         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
761         long idle_time = 0;
762
763         if (xprt_connected(xprt))
764                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
765
766         seq_puts(seq, "\txprt:\trdma ");
767         seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
768                    0,   /* need a local port? */
769                    xprt->stat.bind_count,
770                    xprt->stat.connect_count,
771                    xprt->stat.connect_time / HZ,
772                    idle_time,
773                    xprt->stat.sends,
774                    xprt->stat.recvs,
775                    xprt->stat.bad_xids,
776                    xprt->stat.req_u,
777                    xprt->stat.bklog_u);
778         seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
779                    r_xprt->rx_stats.read_chunk_count,
780                    r_xprt->rx_stats.write_chunk_count,
781                    r_xprt->rx_stats.reply_chunk_count,
782                    r_xprt->rx_stats.total_rdma_request,
783                    r_xprt->rx_stats.total_rdma_reply,
784                    r_xprt->rx_stats.pullup_copy_count,
785                    r_xprt->rx_stats.fixup_copy_count,
786                    r_xprt->rx_stats.hardway_register_count,
787                    r_xprt->rx_stats.failed_marshal_count,
788                    r_xprt->rx_stats.bad_reply_count,
789                    r_xprt->rx_stats.nomsg_call_count);
790         seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n",
791                    r_xprt->rx_stats.mrs_recycled,
792                    r_xprt->rx_stats.mrs_orphaned,
793                    r_xprt->rx_stats.mrs_allocated,
794                    r_xprt->rx_stats.local_inv_needed,
795                    r_xprt->rx_stats.empty_sendctx_q,
796                    r_xprt->rx_stats.reply_waits_for_send);
797 }
798
799 static int
800 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
801 {
802         return 0;
803 }
804
805 static void
806 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
807 {
808 }
809
810 /*
811  * Plumbing for rpc transport switch and kernel module
812  */
813
814 static const struct rpc_xprt_ops xprt_rdma_procs = {
815         .reserve_xprt           = xprt_reserve_xprt_cong,
816         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
817         .alloc_slot             = xprt_rdma_alloc_slot,
818         .free_slot              = xprt_rdma_free_slot,
819         .release_request        = xprt_release_rqst_cong,       /* ditto */
820         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
821         .timer                  = xprt_rdma_timer,
822         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
823         .set_port               = xprt_rdma_set_port,
824         .connect                = xprt_rdma_connect,
825         .buf_alloc              = xprt_rdma_allocate,
826         .buf_free               = xprt_rdma_free,
827         .send_request           = xprt_rdma_send_request,
828         .close                  = xprt_rdma_close,
829         .destroy                = xprt_rdma_destroy,
830         .print_stats            = xprt_rdma_print_stats,
831         .enable_swap            = xprt_rdma_enable_swap,
832         .disable_swap           = xprt_rdma_disable_swap,
833         .inject_disconnect      = xprt_rdma_inject_disconnect,
834 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
835         .bc_setup               = xprt_rdma_bc_setup,
836         .bc_maxpayload          = xprt_rdma_bc_maxpayload,
837         .bc_free_rqst           = xprt_rdma_bc_free_rqst,
838         .bc_destroy             = xprt_rdma_bc_destroy,
839 #endif
840 };
841
842 static struct xprt_class xprt_rdma = {
843         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
844         .name                   = "rdma",
845         .owner                  = THIS_MODULE,
846         .ident                  = XPRT_TRANSPORT_RDMA,
847         .setup                  = xprt_setup_rdma,
848 };
849
850 void xprt_rdma_cleanup(void)
851 {
852 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
853         if (sunrpc_table_header) {
854                 unregister_sysctl_table(sunrpc_table_header);
855                 sunrpc_table_header = NULL;
856         }
857 #endif
858
859         xprt_unregister_transport(&xprt_rdma);
860         xprt_unregister_transport(&xprt_rdma_bc);
861 }
862
863 int xprt_rdma_init(void)
864 {
865         int rc;
866
867         rc = xprt_register_transport(&xprt_rdma);
868         if (rc)
869                 return rc;
870
871         rc = xprt_register_transport(&xprt_rdma_bc);
872         if (rc) {
873                 xprt_unregister_transport(&xprt_rdma);
874                 return rc;
875         }
876
877 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
878         if (!sunrpc_table_header)
879                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
880 #endif
881         return 0;
882 }