Merge branch 'idr-2018-02-06' of git://git.infradead.org/users/willy/linux-dax
[muen/linux.git] / net / sunrpc / xprtrdma / svc_rdma_transport.c
1 /*
2  * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
3  * Copyright (c) 2005-2007 Network Appliance, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the BSD-type
9  * license below:
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  *
15  *      Redistributions of source code must retain the above copyright
16  *      notice, this list of conditions and the following disclaimer.
17  *
18  *      Redistributions in binary form must reproduce the above
19  *      copyright notice, this list of conditions and the following
20  *      disclaimer in the documentation and/or other materials provided
21  *      with the distribution.
22  *
23  *      Neither the name of the Network Appliance, Inc. nor the names of
24  *      its contributors may be used to endorse or promote products
25  *      derived from this software without specific prior written
26  *      permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39  *
40  * Author: Tom Tucker <tom@opengridcomputing.com>
41  */
42
43 #include <linux/sunrpc/svc_xprt.h>
44 #include <linux/sunrpc/addr.h>
45 #include <linux/sunrpc/debug.h>
46 #include <linux/sunrpc/rpc_rdma.h>
47 #include <linux/interrupt.h>
48 #include <linux/sched.h>
49 #include <linux/slab.h>
50 #include <linux/spinlock.h>
51 #include <linux/workqueue.h>
52 #include <rdma/ib_verbs.h>
53 #include <rdma/rdma_cm.h>
54 #include <rdma/rw.h>
55 #include <linux/sunrpc/svc_rdma.h>
56 #include <linux/export.h>
57 #include "xprt_rdma.h"
58
59 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
60
61 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *, int);
62 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
63                                         struct net *net,
64                                         struct sockaddr *sa, int salen,
65                                         int flags);
66 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt);
67 static void svc_rdma_release_rqst(struct svc_rqst *);
68 static void svc_rdma_detach(struct svc_xprt *xprt);
69 static void svc_rdma_free(struct svc_xprt *xprt);
70 static int svc_rdma_has_wspace(struct svc_xprt *xprt);
71 static int svc_rdma_secure_port(struct svc_rqst *);
72 static void svc_rdma_kill_temp_xprt(struct svc_xprt *);
73
74 static const struct svc_xprt_ops svc_rdma_ops = {
75         .xpo_create = svc_rdma_create,
76         .xpo_recvfrom = svc_rdma_recvfrom,
77         .xpo_sendto = svc_rdma_sendto,
78         .xpo_release_rqst = svc_rdma_release_rqst,
79         .xpo_detach = svc_rdma_detach,
80         .xpo_free = svc_rdma_free,
81         .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
82         .xpo_has_wspace = svc_rdma_has_wspace,
83         .xpo_accept = svc_rdma_accept,
84         .xpo_secure_port = svc_rdma_secure_port,
85         .xpo_kill_temp_xprt = svc_rdma_kill_temp_xprt,
86 };
87
88 struct svc_xprt_class svc_rdma_class = {
89         .xcl_name = "rdma",
90         .xcl_owner = THIS_MODULE,
91         .xcl_ops = &svc_rdma_ops,
92         .xcl_max_payload = RPCSVC_MAXPAYLOAD_RDMA,
93         .xcl_ident = XPRT_TRANSPORT_RDMA,
94 };
95
96 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
97 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *, struct net *,
98                                            struct sockaddr *, int, int);
99 static void svc_rdma_bc_detach(struct svc_xprt *);
100 static void svc_rdma_bc_free(struct svc_xprt *);
101
102 static const struct svc_xprt_ops svc_rdma_bc_ops = {
103         .xpo_create = svc_rdma_bc_create,
104         .xpo_detach = svc_rdma_bc_detach,
105         .xpo_free = svc_rdma_bc_free,
106         .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
107         .xpo_secure_port = svc_rdma_secure_port,
108 };
109
110 struct svc_xprt_class svc_rdma_bc_class = {
111         .xcl_name = "rdma-bc",
112         .xcl_owner = THIS_MODULE,
113         .xcl_ops = &svc_rdma_bc_ops,
114         .xcl_max_payload = (1024 - RPCRDMA_HDRLEN_MIN)
115 };
116
117 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *serv,
118                                            struct net *net,
119                                            struct sockaddr *sa, int salen,
120                                            int flags)
121 {
122         struct svcxprt_rdma *cma_xprt;
123         struct svc_xprt *xprt;
124
125         cma_xprt = rdma_create_xprt(serv, 0);
126         if (!cma_xprt)
127                 return ERR_PTR(-ENOMEM);
128         xprt = &cma_xprt->sc_xprt;
129
130         svc_xprt_init(net, &svc_rdma_bc_class, xprt, serv);
131         set_bit(XPT_CONG_CTRL, &xprt->xpt_flags);
132         serv->sv_bc_xprt = xprt;
133
134         dprintk("svcrdma: %s(%p)\n", __func__, xprt);
135         return xprt;
136 }
137
138 static void svc_rdma_bc_detach(struct svc_xprt *xprt)
139 {
140         dprintk("svcrdma: %s(%p)\n", __func__, xprt);
141 }
142
143 static void svc_rdma_bc_free(struct svc_xprt *xprt)
144 {
145         struct svcxprt_rdma *rdma =
146                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
147
148         dprintk("svcrdma: %s(%p)\n", __func__, xprt);
149         if (xprt)
150                 kfree(rdma);
151 }
152 #endif  /* CONFIG_SUNRPC_BACKCHANNEL */
153
154 static struct svc_rdma_op_ctxt *alloc_ctxt(struct svcxprt_rdma *xprt,
155                                            gfp_t flags)
156 {
157         struct svc_rdma_op_ctxt *ctxt;
158
159         ctxt = kmalloc(sizeof(*ctxt), flags);
160         if (ctxt) {
161                 ctxt->xprt = xprt;
162                 INIT_LIST_HEAD(&ctxt->list);
163         }
164         return ctxt;
165 }
166
167 static bool svc_rdma_prealloc_ctxts(struct svcxprt_rdma *xprt)
168 {
169         unsigned int i;
170
171         /* Each RPC/RDMA credit can consume one Receive and
172          * one Send WQE at the same time.
173          */
174         i = xprt->sc_sq_depth + xprt->sc_rq_depth;
175
176         while (i--) {
177                 struct svc_rdma_op_ctxt *ctxt;
178
179                 ctxt = alloc_ctxt(xprt, GFP_KERNEL);
180                 if (!ctxt) {
181                         dprintk("svcrdma: No memory for RDMA ctxt\n");
182                         return false;
183                 }
184                 list_add(&ctxt->list, &xprt->sc_ctxts);
185         }
186         return true;
187 }
188
189 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
190 {
191         struct svc_rdma_op_ctxt *ctxt = NULL;
192
193         spin_lock(&xprt->sc_ctxt_lock);
194         xprt->sc_ctxt_used++;
195         if (list_empty(&xprt->sc_ctxts))
196                 goto out_empty;
197
198         ctxt = list_first_entry(&xprt->sc_ctxts,
199                                 struct svc_rdma_op_ctxt, list);
200         list_del(&ctxt->list);
201         spin_unlock(&xprt->sc_ctxt_lock);
202
203 out:
204         ctxt->count = 0;
205         ctxt->mapped_sges = 0;
206         return ctxt;
207
208 out_empty:
209         /* Either pre-allocation missed the mark, or send
210          * queue accounting is broken.
211          */
212         spin_unlock(&xprt->sc_ctxt_lock);
213
214         ctxt = alloc_ctxt(xprt, GFP_NOIO);
215         if (ctxt)
216                 goto out;
217
218         spin_lock(&xprt->sc_ctxt_lock);
219         xprt->sc_ctxt_used--;
220         spin_unlock(&xprt->sc_ctxt_lock);
221         WARN_ONCE(1, "svcrdma: empty RDMA ctxt list?\n");
222         return NULL;
223 }
224
225 void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt)
226 {
227         struct svcxprt_rdma *xprt = ctxt->xprt;
228         struct ib_device *device = xprt->sc_cm_id->device;
229         unsigned int i;
230
231         for (i = 0; i < ctxt->mapped_sges; i++)
232                 ib_dma_unmap_page(device,
233                                   ctxt->sge[i].addr,
234                                   ctxt->sge[i].length,
235                                   ctxt->direction);
236         ctxt->mapped_sges = 0;
237 }
238
239 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
240 {
241         struct svcxprt_rdma *xprt = ctxt->xprt;
242         int i;
243
244         if (free_pages)
245                 for (i = 0; i < ctxt->count; i++)
246                         put_page(ctxt->pages[i]);
247
248         spin_lock(&xprt->sc_ctxt_lock);
249         xprt->sc_ctxt_used--;
250         list_add(&ctxt->list, &xprt->sc_ctxts);
251         spin_unlock(&xprt->sc_ctxt_lock);
252 }
253
254 static void svc_rdma_destroy_ctxts(struct svcxprt_rdma *xprt)
255 {
256         while (!list_empty(&xprt->sc_ctxts)) {
257                 struct svc_rdma_op_ctxt *ctxt;
258
259                 ctxt = list_first_entry(&xprt->sc_ctxts,
260                                         struct svc_rdma_op_ctxt, list);
261                 list_del(&ctxt->list);
262                 kfree(ctxt);
263         }
264 }
265
266 /* QP event handler */
267 static void qp_event_handler(struct ib_event *event, void *context)
268 {
269         struct svc_xprt *xprt = context;
270
271         switch (event->event) {
272         /* These are considered benign events */
273         case IB_EVENT_PATH_MIG:
274         case IB_EVENT_COMM_EST:
275         case IB_EVENT_SQ_DRAINED:
276         case IB_EVENT_QP_LAST_WQE_REACHED:
277                 dprintk("svcrdma: QP event %s (%d) received for QP=%p\n",
278                         ib_event_msg(event->event), event->event,
279                         event->element.qp);
280                 break;
281         /* These are considered fatal events */
282         case IB_EVENT_PATH_MIG_ERR:
283         case IB_EVENT_QP_FATAL:
284         case IB_EVENT_QP_REQ_ERR:
285         case IB_EVENT_QP_ACCESS_ERR:
286         case IB_EVENT_DEVICE_FATAL:
287         default:
288                 dprintk("svcrdma: QP ERROR event %s (%d) received for QP=%p, "
289                         "closing transport\n",
290                         ib_event_msg(event->event), event->event,
291                         event->element.qp);
292                 set_bit(XPT_CLOSE, &xprt->xpt_flags);
293                 svc_xprt_enqueue(xprt);
294                 break;
295         }
296 }
297
298 /**
299  * svc_rdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
300  * @cq:        completion queue
301  * @wc:        completed WR
302  *
303  */
304 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
305 {
306         struct svcxprt_rdma *xprt = cq->cq_context;
307         struct ib_cqe *cqe = wc->wr_cqe;
308         struct svc_rdma_op_ctxt *ctxt;
309
310         /* WARNING: Only wc->wr_cqe and wc->status are reliable */
311         ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
312         svc_rdma_unmap_dma(ctxt);
313
314         if (wc->status != IB_WC_SUCCESS)
315                 goto flushed;
316
317         /* All wc fields are now known to be valid */
318         ctxt->byte_len = wc->byte_len;
319         spin_lock(&xprt->sc_rq_dto_lock);
320         list_add_tail(&ctxt->list, &xprt->sc_rq_dto_q);
321         spin_unlock(&xprt->sc_rq_dto_lock);
322
323         set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
324         if (test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
325                 goto out;
326         goto out_enqueue;
327
328 flushed:
329         if (wc->status != IB_WC_WR_FLUSH_ERR)
330                 pr_warn("svcrdma: receive: %s (%u/0x%x)\n",
331                         ib_wc_status_msg(wc->status),
332                         wc->status, wc->vendor_err);
333         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
334         svc_rdma_put_context(ctxt, 1);
335
336 out_enqueue:
337         svc_xprt_enqueue(&xprt->sc_xprt);
338 out:
339         svc_xprt_put(&xprt->sc_xprt);
340 }
341
342 /**
343  * svc_rdma_wc_send - Invoked by RDMA provider for each polled Send WC
344  * @cq:        completion queue
345  * @wc:        completed WR
346  *
347  */
348 void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
349 {
350         struct svcxprt_rdma *xprt = cq->cq_context;
351         struct ib_cqe *cqe = wc->wr_cqe;
352         struct svc_rdma_op_ctxt *ctxt;
353
354         atomic_inc(&xprt->sc_sq_avail);
355         wake_up(&xprt->sc_send_wait);
356
357         ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
358         svc_rdma_unmap_dma(ctxt);
359         svc_rdma_put_context(ctxt, 1);
360
361         if (unlikely(wc->status != IB_WC_SUCCESS)) {
362                 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
363                 svc_xprt_enqueue(&xprt->sc_xprt);
364                 if (wc->status != IB_WC_WR_FLUSH_ERR)
365                         pr_err("svcrdma: Send: %s (%u/0x%x)\n",
366                                ib_wc_status_msg(wc->status),
367                                wc->status, wc->vendor_err);
368         }
369
370         svc_xprt_put(&xprt->sc_xprt);
371 }
372
373 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
374                                              int listener)
375 {
376         struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
377
378         if (!cma_xprt)
379                 return NULL;
380         svc_xprt_init(&init_net, &svc_rdma_class, &cma_xprt->sc_xprt, serv);
381         INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
382         INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
383         INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
384         INIT_LIST_HEAD(&cma_xprt->sc_ctxts);
385         INIT_LIST_HEAD(&cma_xprt->sc_rw_ctxts);
386         init_waitqueue_head(&cma_xprt->sc_send_wait);
387
388         spin_lock_init(&cma_xprt->sc_lock);
389         spin_lock_init(&cma_xprt->sc_rq_dto_lock);
390         spin_lock_init(&cma_xprt->sc_ctxt_lock);
391         spin_lock_init(&cma_xprt->sc_rw_ctxt_lock);
392
393         /*
394          * Note that this implies that the underlying transport support
395          * has some form of congestion control (see RFC 7530 section 3.1
396          * paragraph 2). For now, we assume that all supported RDMA
397          * transports are suitable here.
398          */
399         set_bit(XPT_CONG_CTRL, &cma_xprt->sc_xprt.xpt_flags);
400
401         if (listener)
402                 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
403
404         return cma_xprt;
405 }
406
407 int svc_rdma_post_recv(struct svcxprt_rdma *xprt, gfp_t flags)
408 {
409         struct ib_recv_wr recv_wr, *bad_recv_wr;
410         struct svc_rdma_op_ctxt *ctxt;
411         struct page *page;
412         dma_addr_t pa;
413         int sge_no;
414         int buflen;
415         int ret;
416
417         ctxt = svc_rdma_get_context(xprt);
418         buflen = 0;
419         ctxt->direction = DMA_FROM_DEVICE;
420         ctxt->cqe.done = svc_rdma_wc_receive;
421         for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
422                 if (sge_no >= xprt->sc_max_sge) {
423                         pr_err("svcrdma: Too many sges (%d)\n", sge_no);
424                         goto err_put_ctxt;
425                 }
426                 page = alloc_page(flags);
427                 if (!page)
428                         goto err_put_ctxt;
429                 ctxt->pages[sge_no] = page;
430                 pa = ib_dma_map_page(xprt->sc_cm_id->device,
431                                      page, 0, PAGE_SIZE,
432                                      DMA_FROM_DEVICE);
433                 if (ib_dma_mapping_error(xprt->sc_cm_id->device, pa))
434                         goto err_put_ctxt;
435                 svc_rdma_count_mappings(xprt, ctxt);
436                 ctxt->sge[sge_no].addr = pa;
437                 ctxt->sge[sge_no].length = PAGE_SIZE;
438                 ctxt->sge[sge_no].lkey = xprt->sc_pd->local_dma_lkey;
439                 ctxt->count = sge_no + 1;
440                 buflen += PAGE_SIZE;
441         }
442         recv_wr.next = NULL;
443         recv_wr.sg_list = &ctxt->sge[0];
444         recv_wr.num_sge = ctxt->count;
445         recv_wr.wr_cqe = &ctxt->cqe;
446
447         svc_xprt_get(&xprt->sc_xprt);
448         ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
449         if (ret) {
450                 svc_rdma_unmap_dma(ctxt);
451                 svc_rdma_put_context(ctxt, 1);
452                 svc_xprt_put(&xprt->sc_xprt);
453         }
454         return ret;
455
456  err_put_ctxt:
457         svc_rdma_unmap_dma(ctxt);
458         svc_rdma_put_context(ctxt, 1);
459         return -ENOMEM;
460 }
461
462 int svc_rdma_repost_recv(struct svcxprt_rdma *xprt, gfp_t flags)
463 {
464         int ret = 0;
465
466         ret = svc_rdma_post_recv(xprt, flags);
467         if (ret) {
468                 pr_err("svcrdma: could not post a receive buffer, err=%d.\n",
469                        ret);
470                 pr_err("svcrdma: closing transport %p.\n", xprt);
471                 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
472                 ret = -ENOTCONN;
473         }
474         return ret;
475 }
476
477 static void
478 svc_rdma_parse_connect_private(struct svcxprt_rdma *newxprt,
479                                struct rdma_conn_param *param)
480 {
481         const struct rpcrdma_connect_private *pmsg = param->private_data;
482
483         if (pmsg &&
484             pmsg->cp_magic == rpcrdma_cmp_magic &&
485             pmsg->cp_version == RPCRDMA_CMP_VERSION) {
486                 newxprt->sc_snd_w_inv = pmsg->cp_flags &
487                                         RPCRDMA_CMP_F_SND_W_INV_OK;
488
489                 dprintk("svcrdma: client send_size %u, recv_size %u "
490                         "remote inv %ssupported\n",
491                         rpcrdma_decode_buffer_size(pmsg->cp_send_size),
492                         rpcrdma_decode_buffer_size(pmsg->cp_recv_size),
493                         newxprt->sc_snd_w_inv ? "" : "un");
494         }
495 }
496
497 /*
498  * This function handles the CONNECT_REQUEST event on a listening
499  * endpoint. It is passed the cma_id for the _new_ connection. The context in
500  * this cma_id is inherited from the listening cma_id and is the svc_xprt
501  * structure for the listening endpoint.
502  *
503  * This function creates a new xprt for the new connection and enqueues it on
504  * the accept queue for the listent xprt. When the listen thread is kicked, it
505  * will call the recvfrom method on the listen xprt which will accept the new
506  * connection.
507  */
508 static void handle_connect_req(struct rdma_cm_id *new_cma_id,
509                                struct rdma_conn_param *param)
510 {
511         struct svcxprt_rdma *listen_xprt = new_cma_id->context;
512         struct svcxprt_rdma *newxprt;
513         struct sockaddr *sa;
514
515         /* Create a new transport */
516         newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
517         if (!newxprt) {
518                 dprintk("svcrdma: failed to create new transport\n");
519                 return;
520         }
521         newxprt->sc_cm_id = new_cma_id;
522         new_cma_id->context = newxprt;
523         dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
524                 newxprt, newxprt->sc_cm_id, listen_xprt);
525         svc_rdma_parse_connect_private(newxprt, param);
526
527         /* Save client advertised inbound read limit for use later in accept. */
528         newxprt->sc_ord = param->initiator_depth;
529
530         /* Set the local and remote addresses in the transport */
531         sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
532         svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
533         sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
534         svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
535
536         /*
537          * Enqueue the new transport on the accept queue of the listening
538          * transport
539          */
540         spin_lock_bh(&listen_xprt->sc_lock);
541         list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
542         spin_unlock_bh(&listen_xprt->sc_lock);
543
544         set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
545         svc_xprt_enqueue(&listen_xprt->sc_xprt);
546 }
547
548 /*
549  * Handles events generated on the listening endpoint. These events will be
550  * either be incoming connect requests or adapter removal  events.
551  */
552 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
553                                struct rdma_cm_event *event)
554 {
555         struct svcxprt_rdma *xprt = cma_id->context;
556         int ret = 0;
557
558         switch (event->event) {
559         case RDMA_CM_EVENT_CONNECT_REQUEST:
560                 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
561                         "event = %s (%d)\n", cma_id, cma_id->context,
562                         rdma_event_msg(event->event), event->event);
563                 handle_connect_req(cma_id, &event->param.conn);
564                 break;
565
566         case RDMA_CM_EVENT_ESTABLISHED:
567                 /* Accept complete */
568                 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
569                         "cm_id=%p\n", xprt, cma_id);
570                 break;
571
572         case RDMA_CM_EVENT_DEVICE_REMOVAL:
573                 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
574                         xprt, cma_id);
575                 if (xprt) {
576                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
577                         svc_xprt_enqueue(&xprt->sc_xprt);
578                 }
579                 break;
580
581         default:
582                 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
583                         "event = %s (%d)\n", cma_id,
584                         rdma_event_msg(event->event), event->event);
585                 break;
586         }
587
588         return ret;
589 }
590
591 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
592                             struct rdma_cm_event *event)
593 {
594         struct svc_xprt *xprt = cma_id->context;
595         struct svcxprt_rdma *rdma =
596                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
597         switch (event->event) {
598         case RDMA_CM_EVENT_ESTABLISHED:
599                 /* Accept complete */
600                 svc_xprt_get(xprt);
601                 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
602                         "cm_id=%p\n", xprt, cma_id);
603                 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
604                 svc_xprt_enqueue(xprt);
605                 break;
606         case RDMA_CM_EVENT_DISCONNECTED:
607                 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
608                         xprt, cma_id);
609                 if (xprt) {
610                         set_bit(XPT_CLOSE, &xprt->xpt_flags);
611                         svc_xprt_enqueue(xprt);
612                         svc_xprt_put(xprt);
613                 }
614                 break;
615         case RDMA_CM_EVENT_DEVICE_REMOVAL:
616                 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
617                         "event = %s (%d)\n", cma_id, xprt,
618                         rdma_event_msg(event->event), event->event);
619                 if (xprt) {
620                         set_bit(XPT_CLOSE, &xprt->xpt_flags);
621                         svc_xprt_enqueue(xprt);
622                         svc_xprt_put(xprt);
623                 }
624                 break;
625         default:
626                 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
627                         "event = %s (%d)\n", cma_id,
628                         rdma_event_msg(event->event), event->event);
629                 break;
630         }
631         return 0;
632 }
633
634 /*
635  * Create a listening RDMA service endpoint.
636  */
637 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
638                                         struct net *net,
639                                         struct sockaddr *sa, int salen,
640                                         int flags)
641 {
642         struct rdma_cm_id *listen_id;
643         struct svcxprt_rdma *cma_xprt;
644         int ret;
645
646         dprintk("svcrdma: Creating RDMA socket\n");
647         if ((sa->sa_family != AF_INET) && (sa->sa_family != AF_INET6)) {
648                 dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family);
649                 return ERR_PTR(-EAFNOSUPPORT);
650         }
651         cma_xprt = rdma_create_xprt(serv, 1);
652         if (!cma_xprt)
653                 return ERR_PTR(-ENOMEM);
654
655         listen_id = rdma_create_id(&init_net, rdma_listen_handler, cma_xprt,
656                                    RDMA_PS_TCP, IB_QPT_RC);
657         if (IS_ERR(listen_id)) {
658                 ret = PTR_ERR(listen_id);
659                 dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
660                 goto err0;
661         }
662
663         /* Allow both IPv4 and IPv6 sockets to bind a single port
664          * at the same time.
665          */
666 #if IS_ENABLED(CONFIG_IPV6)
667         ret = rdma_set_afonly(listen_id, 1);
668         if (ret) {
669                 dprintk("svcrdma: rdma_set_afonly failed = %d\n", ret);
670                 goto err1;
671         }
672 #endif
673         ret = rdma_bind_addr(listen_id, sa);
674         if (ret) {
675                 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
676                 goto err1;
677         }
678         cma_xprt->sc_cm_id = listen_id;
679
680         ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
681         if (ret) {
682                 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
683                 goto err1;
684         }
685
686         /*
687          * We need to use the address from the cm_id in case the
688          * caller specified 0 for the port number.
689          */
690         sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
691         svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
692
693         return &cma_xprt->sc_xprt;
694
695  err1:
696         rdma_destroy_id(listen_id);
697  err0:
698         kfree(cma_xprt);
699         return ERR_PTR(ret);
700 }
701
702 /*
703  * This is the xpo_recvfrom function for listening endpoints. Its
704  * purpose is to accept incoming connections. The CMA callback handler
705  * has already created a new transport and attached it to the new CMA
706  * ID.
707  *
708  * There is a queue of pending connections hung on the listening
709  * transport. This queue contains the new svc_xprt structure. This
710  * function takes svc_xprt structures off the accept_q and completes
711  * the connection.
712  */
713 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
714 {
715         struct svcxprt_rdma *listen_rdma;
716         struct svcxprt_rdma *newxprt = NULL;
717         struct rdma_conn_param conn_param;
718         struct rpcrdma_connect_private pmsg;
719         struct ib_qp_init_attr qp_attr;
720         struct ib_device *dev;
721         struct sockaddr *sap;
722         unsigned int i, ctxts;
723         int ret = 0;
724
725         listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
726         clear_bit(XPT_CONN, &xprt->xpt_flags);
727         /* Get the next entry off the accept list */
728         spin_lock_bh(&listen_rdma->sc_lock);
729         if (!list_empty(&listen_rdma->sc_accept_q)) {
730                 newxprt = list_entry(listen_rdma->sc_accept_q.next,
731                                      struct svcxprt_rdma, sc_accept_q);
732                 list_del_init(&newxprt->sc_accept_q);
733         }
734         if (!list_empty(&listen_rdma->sc_accept_q))
735                 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
736         spin_unlock_bh(&listen_rdma->sc_lock);
737         if (!newxprt)
738                 return NULL;
739
740         dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
741                 newxprt, newxprt->sc_cm_id);
742
743         dev = newxprt->sc_cm_id->device;
744         newxprt->sc_port_num = newxprt->sc_cm_id->port_num;
745
746         /* Qualify the transport resource defaults with the
747          * capabilities of this particular device */
748         newxprt->sc_max_sge = min((size_t)dev->attrs.max_sge,
749                                   (size_t)RPCSVC_MAXPAGES);
750         newxprt->sc_max_req_size = svcrdma_max_req_size;
751         newxprt->sc_max_requests = svcrdma_max_requests;
752         newxprt->sc_max_bc_requests = svcrdma_max_bc_requests;
753         newxprt->sc_rq_depth = newxprt->sc_max_requests +
754                                newxprt->sc_max_bc_requests;
755         if (newxprt->sc_rq_depth > dev->attrs.max_qp_wr) {
756                 pr_warn("svcrdma: reducing receive depth to %d\n",
757                         dev->attrs.max_qp_wr);
758                 newxprt->sc_rq_depth = dev->attrs.max_qp_wr;
759                 newxprt->sc_max_requests = newxprt->sc_rq_depth - 2;
760                 newxprt->sc_max_bc_requests = 2;
761         }
762         newxprt->sc_fc_credits = cpu_to_be32(newxprt->sc_max_requests);
763         ctxts = rdma_rw_mr_factor(dev, newxprt->sc_port_num, RPCSVC_MAXPAGES);
764         ctxts *= newxprt->sc_max_requests;
765         newxprt->sc_sq_depth = newxprt->sc_rq_depth + ctxts;
766         if (newxprt->sc_sq_depth > dev->attrs.max_qp_wr) {
767                 pr_warn("svcrdma: reducing send depth to %d\n",
768                         dev->attrs.max_qp_wr);
769                 newxprt->sc_sq_depth = dev->attrs.max_qp_wr;
770         }
771         atomic_set(&newxprt->sc_sq_avail, newxprt->sc_sq_depth);
772
773         if (!svc_rdma_prealloc_ctxts(newxprt))
774                 goto errout;
775
776         /*
777          * Limit ORD based on client limit, local device limit, and
778          * configured svcrdma limit.
779          */
780         newxprt->sc_ord = min_t(size_t, dev->attrs.max_qp_rd_atom, newxprt->sc_ord);
781         newxprt->sc_ord = min_t(size_t, svcrdma_ord, newxprt->sc_ord);
782
783         newxprt->sc_pd = ib_alloc_pd(dev, 0);
784         if (IS_ERR(newxprt->sc_pd)) {
785                 dprintk("svcrdma: error creating PD for connect request\n");
786                 goto errout;
787         }
788         newxprt->sc_sq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_sq_depth,
789                                         0, IB_POLL_WORKQUEUE);
790         if (IS_ERR(newxprt->sc_sq_cq)) {
791                 dprintk("svcrdma: error creating SQ CQ for connect request\n");
792                 goto errout;
793         }
794         newxprt->sc_rq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_rq_depth,
795                                         0, IB_POLL_WORKQUEUE);
796         if (IS_ERR(newxprt->sc_rq_cq)) {
797                 dprintk("svcrdma: error creating RQ CQ for connect request\n");
798                 goto errout;
799         }
800
801         memset(&qp_attr, 0, sizeof qp_attr);
802         qp_attr.event_handler = qp_event_handler;
803         qp_attr.qp_context = &newxprt->sc_xprt;
804         qp_attr.port_num = newxprt->sc_port_num;
805         qp_attr.cap.max_rdma_ctxs = ctxts;
806         qp_attr.cap.max_send_wr = newxprt->sc_sq_depth - ctxts;
807         qp_attr.cap.max_recv_wr = newxprt->sc_rq_depth;
808         qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
809         qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
810         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
811         qp_attr.qp_type = IB_QPT_RC;
812         qp_attr.send_cq = newxprt->sc_sq_cq;
813         qp_attr.recv_cq = newxprt->sc_rq_cq;
814         dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n",
815                 newxprt->sc_cm_id, newxprt->sc_pd);
816         dprintk("    cap.max_send_wr = %d, cap.max_recv_wr = %d\n",
817                 qp_attr.cap.max_send_wr, qp_attr.cap.max_recv_wr);
818         dprintk("    cap.max_send_sge = %d, cap.max_recv_sge = %d\n",
819                 qp_attr.cap.max_send_sge, qp_attr.cap.max_recv_sge);
820
821         ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
822         if (ret) {
823                 dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
824                 goto errout;
825         }
826         newxprt->sc_qp = newxprt->sc_cm_id->qp;
827
828         if (!(dev->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
829                 newxprt->sc_snd_w_inv = false;
830         if (!rdma_protocol_iwarp(dev, newxprt->sc_port_num) &&
831             !rdma_ib_or_roce(dev, newxprt->sc_port_num))
832                 goto errout;
833
834         /* Post receive buffers */
835         for (i = 0; i < newxprt->sc_max_requests; i++) {
836                 ret = svc_rdma_post_recv(newxprt, GFP_KERNEL);
837                 if (ret) {
838                         dprintk("svcrdma: failure posting receive buffers\n");
839                         goto errout;
840                 }
841         }
842
843         /* Swap out the handler */
844         newxprt->sc_cm_id->event_handler = rdma_cma_handler;
845
846         /* Construct RDMA-CM private message */
847         pmsg.cp_magic = rpcrdma_cmp_magic;
848         pmsg.cp_version = RPCRDMA_CMP_VERSION;
849         pmsg.cp_flags = 0;
850         pmsg.cp_send_size = pmsg.cp_recv_size =
851                 rpcrdma_encode_buffer_size(newxprt->sc_max_req_size);
852
853         /* Accept Connection */
854         set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
855         memset(&conn_param, 0, sizeof conn_param);
856         conn_param.responder_resources = 0;
857         conn_param.initiator_depth = newxprt->sc_ord;
858         conn_param.private_data = &pmsg;
859         conn_param.private_data_len = sizeof(pmsg);
860         ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
861         if (ret) {
862                 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
863                        ret);
864                 goto errout;
865         }
866
867         dprintk("svcrdma: new connection %p accepted:\n", newxprt);
868         sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
869         dprintk("    local address   : %pIS:%u\n", sap, rpc_get_port(sap));
870         sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
871         dprintk("    remote address  : %pIS:%u\n", sap, rpc_get_port(sap));
872         dprintk("    max_sge         : %d\n", newxprt->sc_max_sge);
873         dprintk("    sq_depth        : %d\n", newxprt->sc_sq_depth);
874         dprintk("    rdma_rw_ctxs    : %d\n", ctxts);
875         dprintk("    max_requests    : %d\n", newxprt->sc_max_requests);
876         dprintk("    ord             : %d\n", newxprt->sc_ord);
877
878         return &newxprt->sc_xprt;
879
880  errout:
881         dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
882         /* Take a reference in case the DTO handler runs */
883         svc_xprt_get(&newxprt->sc_xprt);
884         if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
885                 ib_destroy_qp(newxprt->sc_qp);
886         rdma_destroy_id(newxprt->sc_cm_id);
887         /* This call to put will destroy the transport */
888         svc_xprt_put(&newxprt->sc_xprt);
889         return NULL;
890 }
891
892 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
893 {
894 }
895
896 /*
897  * When connected, an svc_xprt has at least two references:
898  *
899  * - A reference held by the cm_id between the ESTABLISHED and
900  *   DISCONNECTED events. If the remote peer disconnected first, this
901  *   reference could be gone.
902  *
903  * - A reference held by the svc_recv code that called this function
904  *   as part of close processing.
905  *
906  * At a minimum one references should still be held.
907  */
908 static void svc_rdma_detach(struct svc_xprt *xprt)
909 {
910         struct svcxprt_rdma *rdma =
911                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
912         dprintk("svc: svc_rdma_detach(%p)\n", xprt);
913
914         /* Disconnect and flush posted WQE */
915         rdma_disconnect(rdma->sc_cm_id);
916 }
917
918 static void __svc_rdma_free(struct work_struct *work)
919 {
920         struct svcxprt_rdma *rdma =
921                 container_of(work, struct svcxprt_rdma, sc_work);
922         struct svc_xprt *xprt = &rdma->sc_xprt;
923
924         dprintk("svcrdma: %s(%p)\n", __func__, rdma);
925
926         if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
927                 ib_drain_qp(rdma->sc_qp);
928
929         /* We should only be called from kref_put */
930         if (kref_read(&xprt->xpt_ref) != 0)
931                 pr_err("svcrdma: sc_xprt still in use? (%d)\n",
932                        kref_read(&xprt->xpt_ref));
933
934         while (!list_empty(&rdma->sc_read_complete_q)) {
935                 struct svc_rdma_op_ctxt *ctxt;
936                 ctxt = list_first_entry(&rdma->sc_read_complete_q,
937                                         struct svc_rdma_op_ctxt, list);
938                 list_del(&ctxt->list);
939                 svc_rdma_put_context(ctxt, 1);
940         }
941         while (!list_empty(&rdma->sc_rq_dto_q)) {
942                 struct svc_rdma_op_ctxt *ctxt;
943                 ctxt = list_first_entry(&rdma->sc_rq_dto_q,
944                                         struct svc_rdma_op_ctxt, list);
945                 list_del(&ctxt->list);
946                 svc_rdma_put_context(ctxt, 1);
947         }
948
949         /* Warn if we leaked a resource or under-referenced */
950         if (rdma->sc_ctxt_used != 0)
951                 pr_err("svcrdma: ctxt still in use? (%d)\n",
952                        rdma->sc_ctxt_used);
953
954         /* Final put of backchannel client transport */
955         if (xprt->xpt_bc_xprt) {
956                 xprt_put(xprt->xpt_bc_xprt);
957                 xprt->xpt_bc_xprt = NULL;
958         }
959
960         svc_rdma_destroy_rw_ctxts(rdma);
961         svc_rdma_destroy_ctxts(rdma);
962
963         /* Destroy the QP if present (not a listener) */
964         if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
965                 ib_destroy_qp(rdma->sc_qp);
966
967         if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
968                 ib_free_cq(rdma->sc_sq_cq);
969
970         if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
971                 ib_free_cq(rdma->sc_rq_cq);
972
973         if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
974                 ib_dealloc_pd(rdma->sc_pd);
975
976         /* Destroy the CM ID */
977         rdma_destroy_id(rdma->sc_cm_id);
978
979         kfree(rdma);
980 }
981
982 static void svc_rdma_free(struct svc_xprt *xprt)
983 {
984         struct svcxprt_rdma *rdma =
985                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
986         INIT_WORK(&rdma->sc_work, __svc_rdma_free);
987         queue_work(svc_rdma_wq, &rdma->sc_work);
988 }
989
990 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
991 {
992         struct svcxprt_rdma *rdma =
993                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
994
995         /*
996          * If there are already waiters on the SQ,
997          * return false.
998          */
999         if (waitqueue_active(&rdma->sc_send_wait))
1000                 return 0;
1001
1002         /* Otherwise return true. */
1003         return 1;
1004 }
1005
1006 static int svc_rdma_secure_port(struct svc_rqst *rqstp)
1007 {
1008         return 1;
1009 }
1010
1011 static void svc_rdma_kill_temp_xprt(struct svc_xprt *xprt)
1012 {
1013 }
1014
1015 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1016 {
1017         struct ib_send_wr *bad_wr, *n_wr;
1018         int wr_count;
1019         int i;
1020         int ret;
1021
1022         if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1023                 return -ENOTCONN;
1024
1025         wr_count = 1;
1026         for (n_wr = wr->next; n_wr; n_wr = n_wr->next)
1027                 wr_count++;
1028
1029         /* If the SQ is full, wait until an SQ entry is available */
1030         while (1) {
1031                 if ((atomic_sub_return(wr_count, &xprt->sc_sq_avail) < 0)) {
1032                         atomic_inc(&rdma_stat_sq_starve);
1033
1034                         /* Wait until SQ WR available if SQ still full */
1035                         atomic_add(wr_count, &xprt->sc_sq_avail);
1036                         wait_event(xprt->sc_send_wait,
1037                                    atomic_read(&xprt->sc_sq_avail) > wr_count);
1038                         if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1039                                 return -ENOTCONN;
1040                         continue;
1041                 }
1042                 /* Take a transport ref for each WR posted */
1043                 for (i = 0; i < wr_count; i++)
1044                         svc_xprt_get(&xprt->sc_xprt);
1045
1046                 /* Bump used SQ WR count and post */
1047                 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1048                 if (ret) {
1049                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
1050                         for (i = 0; i < wr_count; i ++)
1051                                 svc_xprt_put(&xprt->sc_xprt);
1052                         dprintk("svcrdma: failed to post SQ WR rc=%d\n", ret);
1053                         dprintk("    sc_sq_avail=%d, sc_sq_depth=%d\n",
1054                                 atomic_read(&xprt->sc_sq_avail),
1055                                 xprt->sc_sq_depth);
1056                         wake_up(&xprt->sc_send_wait);
1057                 }
1058                 break;
1059         }
1060         return ret;
1061 }