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 <linux/sunrpc/svc_rdma.h>
55 #include <linux/export.h>
56 #include "xprt_rdma.h"
57 
58 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
59 
60 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *, int);
61 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
62 					struct net *net,
63 					struct sockaddr *sa, int salen,
64 					int flags);
65 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt);
66 static void svc_rdma_release_rqst(struct svc_rqst *);
67 static void svc_rdma_detach(struct svc_xprt *xprt);
68 static void svc_rdma_free(struct svc_xprt *xprt);
69 static int svc_rdma_has_wspace(struct svc_xprt *xprt);
70 static int svc_rdma_secure_port(struct svc_rqst *);
71 static void svc_rdma_kill_temp_xprt(struct svc_xprt *);
72 
73 static struct svc_xprt_ops svc_rdma_ops = {
74 	.xpo_create = svc_rdma_create,
75 	.xpo_recvfrom = svc_rdma_recvfrom,
76 	.xpo_sendto = svc_rdma_sendto,
77 	.xpo_release_rqst = svc_rdma_release_rqst,
78 	.xpo_detach = svc_rdma_detach,
79 	.xpo_free = svc_rdma_free,
80 	.xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
81 	.xpo_has_wspace = svc_rdma_has_wspace,
82 	.xpo_accept = svc_rdma_accept,
83 	.xpo_secure_port = svc_rdma_secure_port,
84 	.xpo_kill_temp_xprt = svc_rdma_kill_temp_xprt,
85 };
86 
87 struct svc_xprt_class svc_rdma_class = {
88 	.xcl_name = "rdma",
89 	.xcl_owner = THIS_MODULE,
90 	.xcl_ops = &svc_rdma_ops,
91 	.xcl_max_payload = RPCSVC_MAXPAYLOAD_RDMA,
92 	.xcl_ident = XPRT_TRANSPORT_RDMA,
93 };
94 
95 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
96 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *, struct net *,
97 					   struct sockaddr *, int, int);
98 static void svc_rdma_bc_detach(struct svc_xprt *);
99 static void svc_rdma_bc_free(struct svc_xprt *);
100 
101 static struct svc_xprt_ops svc_rdma_bc_ops = {
102 	.xpo_create = svc_rdma_bc_create,
103 	.xpo_detach = svc_rdma_bc_detach,
104 	.xpo_free = svc_rdma_bc_free,
105 	.xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
106 	.xpo_secure_port = svc_rdma_secure_port,
107 };
108 
109 struct svc_xprt_class svc_rdma_bc_class = {
110 	.xcl_name = "rdma-bc",
111 	.xcl_owner = THIS_MODULE,
112 	.xcl_ops = &svc_rdma_bc_ops,
113 	.xcl_max_payload = (1024 - RPCRDMA_HDRLEN_MIN)
114 };
115 
116 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *serv,
117 					   struct net *net,
118 					   struct sockaddr *sa, int salen,
119 					   int flags)
120 {
121 	struct svcxprt_rdma *cma_xprt;
122 	struct svc_xprt *xprt;
123 
124 	cma_xprt = rdma_create_xprt(serv, 0);
125 	if (!cma_xprt)
126 		return ERR_PTR(-ENOMEM);
127 	xprt = &cma_xprt->sc_xprt;
128 
129 	svc_xprt_init(net, &svc_rdma_bc_class, xprt, serv);
130 	set_bit(XPT_CONG_CTRL, &xprt->xpt_flags);
131 	serv->sv_bc_xprt = xprt;
132 
133 	dprintk("svcrdma: %s(%p)\n", __func__, xprt);
134 	return xprt;
135 }
136 
137 static void svc_rdma_bc_detach(struct svc_xprt *xprt)
138 {
139 	dprintk("svcrdma: %s(%p)\n", __func__, xprt);
140 }
141 
142 static void svc_rdma_bc_free(struct svc_xprt *xprt)
143 {
144 	struct svcxprt_rdma *rdma =
145 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
146 
147 	dprintk("svcrdma: %s(%p)\n", __func__, xprt);
148 	if (xprt)
149 		kfree(rdma);
150 }
151 #endif	/* CONFIG_SUNRPC_BACKCHANNEL */
152 
153 static struct svc_rdma_op_ctxt *alloc_ctxt(struct svcxprt_rdma *xprt,
154 					   gfp_t flags)
155 {
156 	struct svc_rdma_op_ctxt *ctxt;
157 
158 	ctxt = kmalloc(sizeof(*ctxt), flags);
159 	if (ctxt) {
160 		ctxt->xprt = xprt;
161 		INIT_LIST_HEAD(&ctxt->list);
162 	}
163 	return ctxt;
164 }
165 
166 static bool svc_rdma_prealloc_ctxts(struct svcxprt_rdma *xprt)
167 {
168 	unsigned int i;
169 
170 	/* Each RPC/RDMA credit can consume a number of send
171 	 * and receive WQEs. One ctxt is allocated for each.
172 	 */
173 	i = xprt->sc_sq_depth + xprt->sc_rq_depth;
174 
175 	while (i--) {
176 		struct svc_rdma_op_ctxt *ctxt;
177 
178 		ctxt = alloc_ctxt(xprt, GFP_KERNEL);
179 		if (!ctxt) {
180 			dprintk("svcrdma: No memory for RDMA ctxt\n");
181 			return false;
182 		}
183 		list_add(&ctxt->list, &xprt->sc_ctxts);
184 	}
185 	return true;
186 }
187 
188 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
189 {
190 	struct svc_rdma_op_ctxt *ctxt = NULL;
191 
192 	spin_lock(&xprt->sc_ctxt_lock);
193 	xprt->sc_ctxt_used++;
194 	if (list_empty(&xprt->sc_ctxts))
195 		goto out_empty;
196 
197 	ctxt = list_first_entry(&xprt->sc_ctxts,
198 				struct svc_rdma_op_ctxt, list);
199 	list_del(&ctxt->list);
200 	spin_unlock(&xprt->sc_ctxt_lock);
201 
202 out:
203 	ctxt->count = 0;
204 	ctxt->mapped_sges = 0;
205 	ctxt->frmr = NULL;
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 	u32 lkey = xprt->sc_pd->local_dma_lkey;
230 	unsigned int i;
231 
232 	for (i = 0; i < ctxt->mapped_sges; i++) {
233 		/*
234 		 * Unmap the DMA addr in the SGE if the lkey matches
235 		 * the local_dma_lkey, otherwise, ignore it since it is
236 		 * an FRMR lkey and will be unmapped later when the
237 		 * last WR that uses it completes.
238 		 */
239 		if (ctxt->sge[i].lkey == lkey)
240 			ib_dma_unmap_page(device,
241 					    ctxt->sge[i].addr,
242 					    ctxt->sge[i].length,
243 					    ctxt->direction);
244 	}
245 	ctxt->mapped_sges = 0;
246 }
247 
248 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
249 {
250 	struct svcxprt_rdma *xprt = ctxt->xprt;
251 	int i;
252 
253 	if (free_pages)
254 		for (i = 0; i < ctxt->count; i++)
255 			put_page(ctxt->pages[i]);
256 
257 	spin_lock(&xprt->sc_ctxt_lock);
258 	xprt->sc_ctxt_used--;
259 	list_add(&ctxt->list, &xprt->sc_ctxts);
260 	spin_unlock(&xprt->sc_ctxt_lock);
261 }
262 
263 static void svc_rdma_destroy_ctxts(struct svcxprt_rdma *xprt)
264 {
265 	while (!list_empty(&xprt->sc_ctxts)) {
266 		struct svc_rdma_op_ctxt *ctxt;
267 
268 		ctxt = list_first_entry(&xprt->sc_ctxts,
269 					struct svc_rdma_op_ctxt, list);
270 		list_del(&ctxt->list);
271 		kfree(ctxt);
272 	}
273 }
274 
275 static struct svc_rdma_req_map *alloc_req_map(gfp_t flags)
276 {
277 	struct svc_rdma_req_map *map;
278 
279 	map = kmalloc(sizeof(*map), flags);
280 	if (map)
281 		INIT_LIST_HEAD(&map->free);
282 	return map;
283 }
284 
285 static bool svc_rdma_prealloc_maps(struct svcxprt_rdma *xprt)
286 {
287 	unsigned int i;
288 
289 	/* One for each receive buffer on this connection. */
290 	i = xprt->sc_max_requests;
291 
292 	while (i--) {
293 		struct svc_rdma_req_map *map;
294 
295 		map = alloc_req_map(GFP_KERNEL);
296 		if (!map) {
297 			dprintk("svcrdma: No memory for request map\n");
298 			return false;
299 		}
300 		list_add(&map->free, &xprt->sc_maps);
301 	}
302 	return true;
303 }
304 
305 struct svc_rdma_req_map *svc_rdma_get_req_map(struct svcxprt_rdma *xprt)
306 {
307 	struct svc_rdma_req_map *map = NULL;
308 
309 	spin_lock(&xprt->sc_map_lock);
310 	if (list_empty(&xprt->sc_maps))
311 		goto out_empty;
312 
313 	map = list_first_entry(&xprt->sc_maps,
314 			       struct svc_rdma_req_map, free);
315 	list_del_init(&map->free);
316 	spin_unlock(&xprt->sc_map_lock);
317 
318 out:
319 	map->count = 0;
320 	return map;
321 
322 out_empty:
323 	spin_unlock(&xprt->sc_map_lock);
324 
325 	/* Pre-allocation amount was incorrect */
326 	map = alloc_req_map(GFP_NOIO);
327 	if (map)
328 		goto out;
329 
330 	WARN_ONCE(1, "svcrdma: empty request map list?\n");
331 	return NULL;
332 }
333 
334 void svc_rdma_put_req_map(struct svcxprt_rdma *xprt,
335 			  struct svc_rdma_req_map *map)
336 {
337 	spin_lock(&xprt->sc_map_lock);
338 	list_add(&map->free, &xprt->sc_maps);
339 	spin_unlock(&xprt->sc_map_lock);
340 }
341 
342 static void svc_rdma_destroy_maps(struct svcxprt_rdma *xprt)
343 {
344 	while (!list_empty(&xprt->sc_maps)) {
345 		struct svc_rdma_req_map *map;
346 
347 		map = list_first_entry(&xprt->sc_maps,
348 				       struct svc_rdma_req_map, free);
349 		list_del(&map->free);
350 		kfree(map);
351 	}
352 }
353 
354 /* QP event handler */
355 static void qp_event_handler(struct ib_event *event, void *context)
356 {
357 	struct svc_xprt *xprt = context;
358 
359 	switch (event->event) {
360 	/* These are considered benign events */
361 	case IB_EVENT_PATH_MIG:
362 	case IB_EVENT_COMM_EST:
363 	case IB_EVENT_SQ_DRAINED:
364 	case IB_EVENT_QP_LAST_WQE_REACHED:
365 		dprintk("svcrdma: QP event %s (%d) received for QP=%p\n",
366 			ib_event_msg(event->event), event->event,
367 			event->element.qp);
368 		break;
369 	/* These are considered fatal events */
370 	case IB_EVENT_PATH_MIG_ERR:
371 	case IB_EVENT_QP_FATAL:
372 	case IB_EVENT_QP_REQ_ERR:
373 	case IB_EVENT_QP_ACCESS_ERR:
374 	case IB_EVENT_DEVICE_FATAL:
375 	default:
376 		dprintk("svcrdma: QP ERROR event %s (%d) received for QP=%p, "
377 			"closing transport\n",
378 			ib_event_msg(event->event), event->event,
379 			event->element.qp);
380 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
381 		break;
382 	}
383 }
384 
385 /**
386  * svc_rdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
387  * @cq:        completion queue
388  * @wc:        completed WR
389  *
390  */
391 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
392 {
393 	struct svcxprt_rdma *xprt = cq->cq_context;
394 	struct ib_cqe *cqe = wc->wr_cqe;
395 	struct svc_rdma_op_ctxt *ctxt;
396 
397 	/* WARNING: Only wc->wr_cqe and wc->status are reliable */
398 	ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
399 	svc_rdma_unmap_dma(ctxt);
400 
401 	if (wc->status != IB_WC_SUCCESS)
402 		goto flushed;
403 
404 	/* All wc fields are now known to be valid */
405 	ctxt->byte_len = wc->byte_len;
406 	spin_lock(&xprt->sc_rq_dto_lock);
407 	list_add_tail(&ctxt->list, &xprt->sc_rq_dto_q);
408 	spin_unlock(&xprt->sc_rq_dto_lock);
409 
410 	set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
411 	if (test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
412 		goto out;
413 	svc_xprt_enqueue(&xprt->sc_xprt);
414 	goto out;
415 
416 flushed:
417 	if (wc->status != IB_WC_WR_FLUSH_ERR)
418 		pr_warn("svcrdma: receive: %s (%u/0x%x)\n",
419 			ib_wc_status_msg(wc->status),
420 			wc->status, wc->vendor_err);
421 	set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
422 	svc_rdma_put_context(ctxt, 1);
423 
424 out:
425 	svc_xprt_put(&xprt->sc_xprt);
426 }
427 
428 static void svc_rdma_send_wc_common(struct svcxprt_rdma *xprt,
429 				    struct ib_wc *wc,
430 				    const char *opname)
431 {
432 	if (wc->status != IB_WC_SUCCESS)
433 		goto err;
434 
435 out:
436 	atomic_inc(&xprt->sc_sq_avail);
437 	wake_up(&xprt->sc_send_wait);
438 	return;
439 
440 err:
441 	set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
442 	if (wc->status != IB_WC_WR_FLUSH_ERR)
443 		pr_err("svcrdma: %s: %s (%u/0x%x)\n",
444 		       opname, ib_wc_status_msg(wc->status),
445 		       wc->status, wc->vendor_err);
446 	goto out;
447 }
448 
449 static void svc_rdma_send_wc_common_put(struct ib_cq *cq, struct ib_wc *wc,
450 					const char *opname)
451 {
452 	struct svcxprt_rdma *xprt = cq->cq_context;
453 
454 	svc_rdma_send_wc_common(xprt, wc, opname);
455 	svc_xprt_put(&xprt->sc_xprt);
456 }
457 
458 /**
459  * svc_rdma_wc_send - Invoked by RDMA provider for each polled Send WC
460  * @cq:        completion queue
461  * @wc:        completed WR
462  *
463  */
464 void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
465 {
466 	struct ib_cqe *cqe = wc->wr_cqe;
467 	struct svc_rdma_op_ctxt *ctxt;
468 
469 	svc_rdma_send_wc_common_put(cq, wc, "send");
470 
471 	ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
472 	svc_rdma_unmap_dma(ctxt);
473 	svc_rdma_put_context(ctxt, 1);
474 }
475 
476 /**
477  * svc_rdma_wc_write - Invoked by RDMA provider for each polled Write WC
478  * @cq:        completion queue
479  * @wc:        completed WR
480  *
481  */
482 void svc_rdma_wc_write(struct ib_cq *cq, struct ib_wc *wc)
483 {
484 	struct ib_cqe *cqe = wc->wr_cqe;
485 	struct svc_rdma_op_ctxt *ctxt;
486 
487 	svc_rdma_send_wc_common_put(cq, wc, "write");
488 
489 	ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
490 	svc_rdma_unmap_dma(ctxt);
491 	svc_rdma_put_context(ctxt, 0);
492 }
493 
494 /**
495  * svc_rdma_wc_reg - Invoked by RDMA provider for each polled FASTREG WC
496  * @cq:        completion queue
497  * @wc:        completed WR
498  *
499  */
500 void svc_rdma_wc_reg(struct ib_cq *cq, struct ib_wc *wc)
501 {
502 	svc_rdma_send_wc_common_put(cq, wc, "fastreg");
503 }
504 
505 /**
506  * svc_rdma_wc_read - Invoked by RDMA provider for each polled Read WC
507  * @cq:        completion queue
508  * @wc:        completed WR
509  *
510  */
511 void svc_rdma_wc_read(struct ib_cq *cq, struct ib_wc *wc)
512 {
513 	struct svcxprt_rdma *xprt = cq->cq_context;
514 	struct ib_cqe *cqe = wc->wr_cqe;
515 	struct svc_rdma_op_ctxt *ctxt;
516 
517 	svc_rdma_send_wc_common(xprt, wc, "read");
518 
519 	ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
520 	svc_rdma_unmap_dma(ctxt);
521 	svc_rdma_put_frmr(xprt, ctxt->frmr);
522 
523 	if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
524 		struct svc_rdma_op_ctxt *read_hdr;
525 
526 		read_hdr = ctxt->read_hdr;
527 		spin_lock(&xprt->sc_rq_dto_lock);
528 		list_add_tail(&read_hdr->list,
529 			      &xprt->sc_read_complete_q);
530 		spin_unlock(&xprt->sc_rq_dto_lock);
531 
532 		set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
533 		svc_xprt_enqueue(&xprt->sc_xprt);
534 	}
535 
536 	svc_rdma_put_context(ctxt, 0);
537 	svc_xprt_put(&xprt->sc_xprt);
538 }
539 
540 /**
541  * svc_rdma_wc_inv - Invoked by RDMA provider for each polled LOCAL_INV WC
542  * @cq:        completion queue
543  * @wc:        completed WR
544  *
545  */
546 void svc_rdma_wc_inv(struct ib_cq *cq, struct ib_wc *wc)
547 {
548 	svc_rdma_send_wc_common_put(cq, wc, "localInv");
549 }
550 
551 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
552 					     int listener)
553 {
554 	struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
555 
556 	if (!cma_xprt)
557 		return NULL;
558 	svc_xprt_init(&init_net, &svc_rdma_class, &cma_xprt->sc_xprt, serv);
559 	INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
560 	INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
561 	INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
562 	INIT_LIST_HEAD(&cma_xprt->sc_frmr_q);
563 	INIT_LIST_HEAD(&cma_xprt->sc_ctxts);
564 	INIT_LIST_HEAD(&cma_xprt->sc_maps);
565 	init_waitqueue_head(&cma_xprt->sc_send_wait);
566 
567 	spin_lock_init(&cma_xprt->sc_lock);
568 	spin_lock_init(&cma_xprt->sc_rq_dto_lock);
569 	spin_lock_init(&cma_xprt->sc_frmr_q_lock);
570 	spin_lock_init(&cma_xprt->sc_ctxt_lock);
571 	spin_lock_init(&cma_xprt->sc_map_lock);
572 
573 	/*
574 	 * Note that this implies that the underlying transport support
575 	 * has some form of congestion control (see RFC 7530 section 3.1
576 	 * paragraph 2). For now, we assume that all supported RDMA
577 	 * transports are suitable here.
578 	 */
579 	set_bit(XPT_CONG_CTRL, &cma_xprt->sc_xprt.xpt_flags);
580 
581 	if (listener)
582 		set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
583 
584 	return cma_xprt;
585 }
586 
587 int svc_rdma_post_recv(struct svcxprt_rdma *xprt, gfp_t flags)
588 {
589 	struct ib_recv_wr recv_wr, *bad_recv_wr;
590 	struct svc_rdma_op_ctxt *ctxt;
591 	struct page *page;
592 	dma_addr_t pa;
593 	int sge_no;
594 	int buflen;
595 	int ret;
596 
597 	ctxt = svc_rdma_get_context(xprt);
598 	buflen = 0;
599 	ctxt->direction = DMA_FROM_DEVICE;
600 	ctxt->cqe.done = svc_rdma_wc_receive;
601 	for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
602 		if (sge_no >= xprt->sc_max_sge) {
603 			pr_err("svcrdma: Too many sges (%d)\n", sge_no);
604 			goto err_put_ctxt;
605 		}
606 		page = alloc_page(flags);
607 		if (!page)
608 			goto err_put_ctxt;
609 		ctxt->pages[sge_no] = page;
610 		pa = ib_dma_map_page(xprt->sc_cm_id->device,
611 				     page, 0, PAGE_SIZE,
612 				     DMA_FROM_DEVICE);
613 		if (ib_dma_mapping_error(xprt->sc_cm_id->device, pa))
614 			goto err_put_ctxt;
615 		svc_rdma_count_mappings(xprt, ctxt);
616 		ctxt->sge[sge_no].addr = pa;
617 		ctxt->sge[sge_no].length = PAGE_SIZE;
618 		ctxt->sge[sge_no].lkey = xprt->sc_pd->local_dma_lkey;
619 		ctxt->count = sge_no + 1;
620 		buflen += PAGE_SIZE;
621 	}
622 	recv_wr.next = NULL;
623 	recv_wr.sg_list = &ctxt->sge[0];
624 	recv_wr.num_sge = ctxt->count;
625 	recv_wr.wr_cqe = &ctxt->cqe;
626 
627 	svc_xprt_get(&xprt->sc_xprt);
628 	ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
629 	if (ret) {
630 		svc_rdma_unmap_dma(ctxt);
631 		svc_rdma_put_context(ctxt, 1);
632 		svc_xprt_put(&xprt->sc_xprt);
633 	}
634 	return ret;
635 
636  err_put_ctxt:
637 	svc_rdma_unmap_dma(ctxt);
638 	svc_rdma_put_context(ctxt, 1);
639 	return -ENOMEM;
640 }
641 
642 int svc_rdma_repost_recv(struct svcxprt_rdma *xprt, gfp_t flags)
643 {
644 	int ret = 0;
645 
646 	ret = svc_rdma_post_recv(xprt, flags);
647 	if (ret) {
648 		pr_err("svcrdma: could not post a receive buffer, err=%d.\n",
649 		       ret);
650 		pr_err("svcrdma: closing transport %p.\n", xprt);
651 		set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
652 		ret = -ENOTCONN;
653 	}
654 	return ret;
655 }
656 
657 static void
658 svc_rdma_parse_connect_private(struct svcxprt_rdma *newxprt,
659 			       struct rdma_conn_param *param)
660 {
661 	const struct rpcrdma_connect_private *pmsg = param->private_data;
662 
663 	if (pmsg &&
664 	    pmsg->cp_magic == rpcrdma_cmp_magic &&
665 	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
666 		newxprt->sc_snd_w_inv = pmsg->cp_flags &
667 					RPCRDMA_CMP_F_SND_W_INV_OK;
668 
669 		dprintk("svcrdma: client send_size %u, recv_size %u "
670 			"remote inv %ssupported\n",
671 			rpcrdma_decode_buffer_size(pmsg->cp_send_size),
672 			rpcrdma_decode_buffer_size(pmsg->cp_recv_size),
673 			newxprt->sc_snd_w_inv ? "" : "un");
674 	}
675 }
676 
677 /*
678  * This function handles the CONNECT_REQUEST event on a listening
679  * endpoint. It is passed the cma_id for the _new_ connection. The context in
680  * this cma_id is inherited from the listening cma_id and is the svc_xprt
681  * structure for the listening endpoint.
682  *
683  * This function creates a new xprt for the new connection and enqueues it on
684  * the accept queue for the listent xprt. When the listen thread is kicked, it
685  * will call the recvfrom method on the listen xprt which will accept the new
686  * connection.
687  */
688 static void handle_connect_req(struct rdma_cm_id *new_cma_id,
689 			       struct rdma_conn_param *param)
690 {
691 	struct svcxprt_rdma *listen_xprt = new_cma_id->context;
692 	struct svcxprt_rdma *newxprt;
693 	struct sockaddr *sa;
694 
695 	/* Create a new transport */
696 	newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
697 	if (!newxprt) {
698 		dprintk("svcrdma: failed to create new transport\n");
699 		return;
700 	}
701 	newxprt->sc_cm_id = new_cma_id;
702 	new_cma_id->context = newxprt;
703 	dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
704 		newxprt, newxprt->sc_cm_id, listen_xprt);
705 	svc_rdma_parse_connect_private(newxprt, param);
706 
707 	/* Save client advertised inbound read limit for use later in accept. */
708 	newxprt->sc_ord = param->initiator_depth;
709 
710 	/* Set the local and remote addresses in the transport */
711 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
712 	svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
713 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
714 	svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
715 
716 	/*
717 	 * Enqueue the new transport on the accept queue of the listening
718 	 * transport
719 	 */
720 	spin_lock_bh(&listen_xprt->sc_lock);
721 	list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
722 	spin_unlock_bh(&listen_xprt->sc_lock);
723 
724 	set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
725 	svc_xprt_enqueue(&listen_xprt->sc_xprt);
726 }
727 
728 /*
729  * Handles events generated on the listening endpoint. These events will be
730  * either be incoming connect requests or adapter removal  events.
731  */
732 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
733 			       struct rdma_cm_event *event)
734 {
735 	struct svcxprt_rdma *xprt = cma_id->context;
736 	int ret = 0;
737 
738 	switch (event->event) {
739 	case RDMA_CM_EVENT_CONNECT_REQUEST:
740 		dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
741 			"event = %s (%d)\n", cma_id, cma_id->context,
742 			rdma_event_msg(event->event), event->event);
743 		handle_connect_req(cma_id, &event->param.conn);
744 		break;
745 
746 	case RDMA_CM_EVENT_ESTABLISHED:
747 		/* Accept complete */
748 		dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
749 			"cm_id=%p\n", xprt, cma_id);
750 		break;
751 
752 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
753 		dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
754 			xprt, cma_id);
755 		if (xprt)
756 			set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
757 		break;
758 
759 	default:
760 		dprintk("svcrdma: Unexpected event on listening endpoint %p, "
761 			"event = %s (%d)\n", cma_id,
762 			rdma_event_msg(event->event), event->event);
763 		break;
764 	}
765 
766 	return ret;
767 }
768 
769 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
770 			    struct rdma_cm_event *event)
771 {
772 	struct svc_xprt *xprt = cma_id->context;
773 	struct svcxprt_rdma *rdma =
774 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
775 	switch (event->event) {
776 	case RDMA_CM_EVENT_ESTABLISHED:
777 		/* Accept complete */
778 		svc_xprt_get(xprt);
779 		dprintk("svcrdma: Connection completed on DTO xprt=%p, "
780 			"cm_id=%p\n", xprt, cma_id);
781 		clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
782 		svc_xprt_enqueue(xprt);
783 		break;
784 	case RDMA_CM_EVENT_DISCONNECTED:
785 		dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
786 			xprt, cma_id);
787 		if (xprt) {
788 			set_bit(XPT_CLOSE, &xprt->xpt_flags);
789 			svc_xprt_enqueue(xprt);
790 			svc_xprt_put(xprt);
791 		}
792 		break;
793 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
794 		dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
795 			"event = %s (%d)\n", cma_id, xprt,
796 			rdma_event_msg(event->event), event->event);
797 		if (xprt) {
798 			set_bit(XPT_CLOSE, &xprt->xpt_flags);
799 			svc_xprt_enqueue(xprt);
800 			svc_xprt_put(xprt);
801 		}
802 		break;
803 	default:
804 		dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
805 			"event = %s (%d)\n", cma_id,
806 			rdma_event_msg(event->event), event->event);
807 		break;
808 	}
809 	return 0;
810 }
811 
812 /*
813  * Create a listening RDMA service endpoint.
814  */
815 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
816 					struct net *net,
817 					struct sockaddr *sa, int salen,
818 					int flags)
819 {
820 	struct rdma_cm_id *listen_id;
821 	struct svcxprt_rdma *cma_xprt;
822 	int ret;
823 
824 	dprintk("svcrdma: Creating RDMA socket\n");
825 	if ((sa->sa_family != AF_INET) && (sa->sa_family != AF_INET6)) {
826 		dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family);
827 		return ERR_PTR(-EAFNOSUPPORT);
828 	}
829 	cma_xprt = rdma_create_xprt(serv, 1);
830 	if (!cma_xprt)
831 		return ERR_PTR(-ENOMEM);
832 
833 	listen_id = rdma_create_id(&init_net, rdma_listen_handler, cma_xprt,
834 				   RDMA_PS_TCP, IB_QPT_RC);
835 	if (IS_ERR(listen_id)) {
836 		ret = PTR_ERR(listen_id);
837 		dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
838 		goto err0;
839 	}
840 
841 	/* Allow both IPv4 and IPv6 sockets to bind a single port
842 	 * at the same time.
843 	 */
844 #if IS_ENABLED(CONFIG_IPV6)
845 	ret = rdma_set_afonly(listen_id, 1);
846 	if (ret) {
847 		dprintk("svcrdma: rdma_set_afonly failed = %d\n", ret);
848 		goto err1;
849 	}
850 #endif
851 	ret = rdma_bind_addr(listen_id, sa);
852 	if (ret) {
853 		dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
854 		goto err1;
855 	}
856 	cma_xprt->sc_cm_id = listen_id;
857 
858 	ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
859 	if (ret) {
860 		dprintk("svcrdma: rdma_listen failed = %d\n", ret);
861 		goto err1;
862 	}
863 
864 	/*
865 	 * We need to use the address from the cm_id in case the
866 	 * caller specified 0 for the port number.
867 	 */
868 	sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
869 	svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
870 
871 	return &cma_xprt->sc_xprt;
872 
873  err1:
874 	rdma_destroy_id(listen_id);
875  err0:
876 	kfree(cma_xprt);
877 	return ERR_PTR(ret);
878 }
879 
880 static struct svc_rdma_fastreg_mr *rdma_alloc_frmr(struct svcxprt_rdma *xprt)
881 {
882 	struct ib_mr *mr;
883 	struct scatterlist *sg;
884 	struct svc_rdma_fastreg_mr *frmr;
885 	u32 num_sg;
886 
887 	frmr = kmalloc(sizeof(*frmr), GFP_KERNEL);
888 	if (!frmr)
889 		goto err;
890 
891 	num_sg = min_t(u32, RPCSVC_MAXPAGES, xprt->sc_frmr_pg_list_len);
892 	mr = ib_alloc_mr(xprt->sc_pd, IB_MR_TYPE_MEM_REG, num_sg);
893 	if (IS_ERR(mr))
894 		goto err_free_frmr;
895 
896 	sg = kcalloc(RPCSVC_MAXPAGES, sizeof(*sg), GFP_KERNEL);
897 	if (!sg)
898 		goto err_free_mr;
899 
900 	sg_init_table(sg, RPCSVC_MAXPAGES);
901 
902 	frmr->mr = mr;
903 	frmr->sg = sg;
904 	INIT_LIST_HEAD(&frmr->frmr_list);
905 	return frmr;
906 
907  err_free_mr:
908 	ib_dereg_mr(mr);
909  err_free_frmr:
910 	kfree(frmr);
911  err:
912 	return ERR_PTR(-ENOMEM);
913 }
914 
915 static void rdma_dealloc_frmr_q(struct svcxprt_rdma *xprt)
916 {
917 	struct svc_rdma_fastreg_mr *frmr;
918 
919 	while (!list_empty(&xprt->sc_frmr_q)) {
920 		frmr = list_entry(xprt->sc_frmr_q.next,
921 				  struct svc_rdma_fastreg_mr, frmr_list);
922 		list_del_init(&frmr->frmr_list);
923 		kfree(frmr->sg);
924 		ib_dereg_mr(frmr->mr);
925 		kfree(frmr);
926 	}
927 }
928 
929 struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma)
930 {
931 	struct svc_rdma_fastreg_mr *frmr = NULL;
932 
933 	spin_lock(&rdma->sc_frmr_q_lock);
934 	if (!list_empty(&rdma->sc_frmr_q)) {
935 		frmr = list_entry(rdma->sc_frmr_q.next,
936 				  struct svc_rdma_fastreg_mr, frmr_list);
937 		list_del_init(&frmr->frmr_list);
938 		frmr->sg_nents = 0;
939 	}
940 	spin_unlock(&rdma->sc_frmr_q_lock);
941 	if (frmr)
942 		return frmr;
943 
944 	return rdma_alloc_frmr(rdma);
945 }
946 
947 void svc_rdma_put_frmr(struct svcxprt_rdma *rdma,
948 		       struct svc_rdma_fastreg_mr *frmr)
949 {
950 	if (frmr) {
951 		ib_dma_unmap_sg(rdma->sc_cm_id->device,
952 				frmr->sg, frmr->sg_nents, frmr->direction);
953 		spin_lock(&rdma->sc_frmr_q_lock);
954 		WARN_ON_ONCE(!list_empty(&frmr->frmr_list));
955 		list_add(&frmr->frmr_list, &rdma->sc_frmr_q);
956 		spin_unlock(&rdma->sc_frmr_q_lock);
957 	}
958 }
959 
960 /*
961  * This is the xpo_recvfrom function for listening endpoints. Its
962  * purpose is to accept incoming connections. The CMA callback handler
963  * has already created a new transport and attached it to the new CMA
964  * ID.
965  *
966  * There is a queue of pending connections hung on the listening
967  * transport. This queue contains the new svc_xprt structure. This
968  * function takes svc_xprt structures off the accept_q and completes
969  * the connection.
970  */
971 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
972 {
973 	struct svcxprt_rdma *listen_rdma;
974 	struct svcxprt_rdma *newxprt = NULL;
975 	struct rdma_conn_param conn_param;
976 	struct rpcrdma_connect_private pmsg;
977 	struct ib_qp_init_attr qp_attr;
978 	struct ib_device *dev;
979 	struct sockaddr *sap;
980 	unsigned int i;
981 	int ret = 0;
982 
983 	listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
984 	clear_bit(XPT_CONN, &xprt->xpt_flags);
985 	/* Get the next entry off the accept list */
986 	spin_lock_bh(&listen_rdma->sc_lock);
987 	if (!list_empty(&listen_rdma->sc_accept_q)) {
988 		newxprt = list_entry(listen_rdma->sc_accept_q.next,
989 				     struct svcxprt_rdma, sc_accept_q);
990 		list_del_init(&newxprt->sc_accept_q);
991 	}
992 	if (!list_empty(&listen_rdma->sc_accept_q))
993 		set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
994 	spin_unlock_bh(&listen_rdma->sc_lock);
995 	if (!newxprt)
996 		return NULL;
997 
998 	dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
999 		newxprt, newxprt->sc_cm_id);
1000 
1001 	dev = newxprt->sc_cm_id->device;
1002 
1003 	/* Qualify the transport resource defaults with the
1004 	 * capabilities of this particular device */
1005 	newxprt->sc_max_sge = min((size_t)dev->attrs.max_sge,
1006 				  (size_t)RPCSVC_MAXPAGES);
1007 	newxprt->sc_max_sge_rd = min_t(size_t, dev->attrs.max_sge_rd,
1008 				       RPCSVC_MAXPAGES);
1009 	newxprt->sc_max_req_size = svcrdma_max_req_size;
1010 	newxprt->sc_max_requests = min_t(u32, dev->attrs.max_qp_wr,
1011 					 svcrdma_max_requests);
1012 	newxprt->sc_fc_credits = cpu_to_be32(newxprt->sc_max_requests);
1013 	newxprt->sc_max_bc_requests = min_t(u32, dev->attrs.max_qp_wr,
1014 					    svcrdma_max_bc_requests);
1015 	newxprt->sc_rq_depth = newxprt->sc_max_requests +
1016 			       newxprt->sc_max_bc_requests;
1017 	newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_rq_depth;
1018 	atomic_set(&newxprt->sc_sq_avail, newxprt->sc_sq_depth);
1019 
1020 	if (!svc_rdma_prealloc_ctxts(newxprt))
1021 		goto errout;
1022 	if (!svc_rdma_prealloc_maps(newxprt))
1023 		goto errout;
1024 
1025 	/*
1026 	 * Limit ORD based on client limit, local device limit, and
1027 	 * configured svcrdma limit.
1028 	 */
1029 	newxprt->sc_ord = min_t(size_t, dev->attrs.max_qp_rd_atom, newxprt->sc_ord);
1030 	newxprt->sc_ord = min_t(size_t,	svcrdma_ord, newxprt->sc_ord);
1031 
1032 	newxprt->sc_pd = ib_alloc_pd(dev, 0);
1033 	if (IS_ERR(newxprt->sc_pd)) {
1034 		dprintk("svcrdma: error creating PD for connect request\n");
1035 		goto errout;
1036 	}
1037 	newxprt->sc_sq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_sq_depth,
1038 					0, IB_POLL_WORKQUEUE);
1039 	if (IS_ERR(newxprt->sc_sq_cq)) {
1040 		dprintk("svcrdma: error creating SQ CQ for connect request\n");
1041 		goto errout;
1042 	}
1043 	newxprt->sc_rq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_rq_depth,
1044 					0, IB_POLL_WORKQUEUE);
1045 	if (IS_ERR(newxprt->sc_rq_cq)) {
1046 		dprintk("svcrdma: error creating RQ CQ for connect request\n");
1047 		goto errout;
1048 	}
1049 
1050 	memset(&qp_attr, 0, sizeof qp_attr);
1051 	qp_attr.event_handler = qp_event_handler;
1052 	qp_attr.qp_context = &newxprt->sc_xprt;
1053 	qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
1054 	qp_attr.cap.max_recv_wr = newxprt->sc_rq_depth;
1055 	qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
1056 	qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
1057 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1058 	qp_attr.qp_type = IB_QPT_RC;
1059 	qp_attr.send_cq = newxprt->sc_sq_cq;
1060 	qp_attr.recv_cq = newxprt->sc_rq_cq;
1061 	dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n",
1062 		newxprt->sc_cm_id, newxprt->sc_pd);
1063 	dprintk("    cap.max_send_wr = %d, cap.max_recv_wr = %d\n",
1064 		qp_attr.cap.max_send_wr, qp_attr.cap.max_recv_wr);
1065 	dprintk("    cap.max_send_sge = %d, cap.max_recv_sge = %d\n",
1066 		qp_attr.cap.max_send_sge, qp_attr.cap.max_recv_sge);
1067 
1068 	ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
1069 	if (ret) {
1070 		dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
1071 		goto errout;
1072 	}
1073 	newxprt->sc_qp = newxprt->sc_cm_id->qp;
1074 
1075 	/*
1076 	 * Use the most secure set of MR resources based on the
1077 	 * transport type and available memory management features in
1078 	 * the device. Here's the table implemented below:
1079 	 *
1080 	 *		Fast	Global	DMA	Remote WR
1081 	 *		Reg	LKEY	MR	Access
1082 	 *		Sup'd	Sup'd	Needed	Needed
1083 	 *
1084 	 * IWARP	N	N	Y	Y
1085 	 *		N	Y	Y	Y
1086 	 *		Y	N	Y	N
1087 	 *		Y	Y	N	-
1088 	 *
1089 	 * IB		N	N	Y	N
1090 	 *		N	Y	N	-
1091 	 *		Y	N	Y	N
1092 	 *		Y	Y	N	-
1093 	 *
1094 	 * NB:	iWARP requires remote write access for the data sink
1095 	 *	of an RDMA_READ. IB does not.
1096 	 */
1097 	newxprt->sc_reader = rdma_read_chunk_lcl;
1098 	if (dev->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1099 		newxprt->sc_frmr_pg_list_len =
1100 			dev->attrs.max_fast_reg_page_list_len;
1101 		newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_FAST_REG;
1102 		newxprt->sc_reader = rdma_read_chunk_frmr;
1103 	} else
1104 		newxprt->sc_snd_w_inv = false;
1105 
1106 	/*
1107 	 * Determine if a DMA MR is required and if so, what privs are required
1108 	 */
1109 	if (!rdma_protocol_iwarp(dev, newxprt->sc_cm_id->port_num) &&
1110 	    !rdma_ib_or_roce(dev, newxprt->sc_cm_id->port_num))
1111 		goto errout;
1112 
1113 	if (rdma_protocol_iwarp(dev, newxprt->sc_cm_id->port_num))
1114 		newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV;
1115 
1116 	/* Post receive buffers */
1117 	for (i = 0; i < newxprt->sc_max_requests; i++) {
1118 		ret = svc_rdma_post_recv(newxprt, GFP_KERNEL);
1119 		if (ret) {
1120 			dprintk("svcrdma: failure posting receive buffers\n");
1121 			goto errout;
1122 		}
1123 	}
1124 
1125 	/* Swap out the handler */
1126 	newxprt->sc_cm_id->event_handler = rdma_cma_handler;
1127 
1128 	/* Construct RDMA-CM private message */
1129 	pmsg.cp_magic = rpcrdma_cmp_magic;
1130 	pmsg.cp_version = RPCRDMA_CMP_VERSION;
1131 	pmsg.cp_flags = 0;
1132 	pmsg.cp_send_size = pmsg.cp_recv_size =
1133 		rpcrdma_encode_buffer_size(newxprt->sc_max_req_size);
1134 
1135 	/* Accept Connection */
1136 	set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
1137 	memset(&conn_param, 0, sizeof conn_param);
1138 	conn_param.responder_resources = 0;
1139 	conn_param.initiator_depth = newxprt->sc_ord;
1140 	conn_param.private_data = &pmsg;
1141 	conn_param.private_data_len = sizeof(pmsg);
1142 	ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
1143 	if (ret) {
1144 		dprintk("svcrdma: failed to accept new connection, ret=%d\n",
1145 		       ret);
1146 		goto errout;
1147 	}
1148 
1149 	dprintk("svcrdma: new connection %p accepted:\n", newxprt);
1150 	sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
1151 	dprintk("    local address   : %pIS:%u\n", sap, rpc_get_port(sap));
1152 	sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
1153 	dprintk("    remote address  : %pIS:%u\n", sap, rpc_get_port(sap));
1154 	dprintk("    max_sge         : %d\n", newxprt->sc_max_sge);
1155 	dprintk("    max_sge_rd      : %d\n", newxprt->sc_max_sge_rd);
1156 	dprintk("    sq_depth        : %d\n", newxprt->sc_sq_depth);
1157 	dprintk("    max_requests    : %d\n", newxprt->sc_max_requests);
1158 	dprintk("    ord             : %d\n", newxprt->sc_ord);
1159 
1160 	return &newxprt->sc_xprt;
1161 
1162  errout:
1163 	dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
1164 	/* Take a reference in case the DTO handler runs */
1165 	svc_xprt_get(&newxprt->sc_xprt);
1166 	if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
1167 		ib_destroy_qp(newxprt->sc_qp);
1168 	rdma_destroy_id(newxprt->sc_cm_id);
1169 	/* This call to put will destroy the transport */
1170 	svc_xprt_put(&newxprt->sc_xprt);
1171 	return NULL;
1172 }
1173 
1174 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
1175 {
1176 }
1177 
1178 /*
1179  * When connected, an svc_xprt has at least two references:
1180  *
1181  * - A reference held by the cm_id between the ESTABLISHED and
1182  *   DISCONNECTED events. If the remote peer disconnected first, this
1183  *   reference could be gone.
1184  *
1185  * - A reference held by the svc_recv code that called this function
1186  *   as part of close processing.
1187  *
1188  * At a minimum one references should still be held.
1189  */
1190 static void svc_rdma_detach(struct svc_xprt *xprt)
1191 {
1192 	struct svcxprt_rdma *rdma =
1193 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
1194 	dprintk("svc: svc_rdma_detach(%p)\n", xprt);
1195 
1196 	/* Disconnect and flush posted WQE */
1197 	rdma_disconnect(rdma->sc_cm_id);
1198 }
1199 
1200 static void __svc_rdma_free(struct work_struct *work)
1201 {
1202 	struct svcxprt_rdma *rdma =
1203 		container_of(work, struct svcxprt_rdma, sc_work);
1204 	struct svc_xprt *xprt = &rdma->sc_xprt;
1205 
1206 	dprintk("svcrdma: %s(%p)\n", __func__, rdma);
1207 
1208 	if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
1209 		ib_drain_qp(rdma->sc_qp);
1210 
1211 	/* We should only be called from kref_put */
1212 	if (kref_read(&xprt->xpt_ref) != 0)
1213 		pr_err("svcrdma: sc_xprt still in use? (%d)\n",
1214 		       kref_read(&xprt->xpt_ref));
1215 
1216 	/*
1217 	 * Destroy queued, but not processed read completions. Note
1218 	 * that this cleanup has to be done before destroying the
1219 	 * cm_id because the device ptr is needed to unmap the dma in
1220 	 * svc_rdma_put_context.
1221 	 */
1222 	while (!list_empty(&rdma->sc_read_complete_q)) {
1223 		struct svc_rdma_op_ctxt *ctxt;
1224 		ctxt = list_first_entry(&rdma->sc_read_complete_q,
1225 					struct svc_rdma_op_ctxt, list);
1226 		list_del(&ctxt->list);
1227 		svc_rdma_put_context(ctxt, 1);
1228 	}
1229 
1230 	/* Destroy queued, but not processed recv completions */
1231 	while (!list_empty(&rdma->sc_rq_dto_q)) {
1232 		struct svc_rdma_op_ctxt *ctxt;
1233 		ctxt = list_first_entry(&rdma->sc_rq_dto_q,
1234 					struct svc_rdma_op_ctxt, list);
1235 		list_del(&ctxt->list);
1236 		svc_rdma_put_context(ctxt, 1);
1237 	}
1238 
1239 	/* Warn if we leaked a resource or under-referenced */
1240 	if (rdma->sc_ctxt_used != 0)
1241 		pr_err("svcrdma: ctxt still in use? (%d)\n",
1242 		       rdma->sc_ctxt_used);
1243 
1244 	/* Final put of backchannel client transport */
1245 	if (xprt->xpt_bc_xprt) {
1246 		xprt_put(xprt->xpt_bc_xprt);
1247 		xprt->xpt_bc_xprt = NULL;
1248 	}
1249 
1250 	rdma_dealloc_frmr_q(rdma);
1251 	svc_rdma_destroy_ctxts(rdma);
1252 	svc_rdma_destroy_maps(rdma);
1253 
1254 	/* Destroy the QP if present (not a listener) */
1255 	if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
1256 		ib_destroy_qp(rdma->sc_qp);
1257 
1258 	if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
1259 		ib_free_cq(rdma->sc_sq_cq);
1260 
1261 	if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
1262 		ib_free_cq(rdma->sc_rq_cq);
1263 
1264 	if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
1265 		ib_dealloc_pd(rdma->sc_pd);
1266 
1267 	/* Destroy the CM ID */
1268 	rdma_destroy_id(rdma->sc_cm_id);
1269 
1270 	kfree(rdma);
1271 }
1272 
1273 static void svc_rdma_free(struct svc_xprt *xprt)
1274 {
1275 	struct svcxprt_rdma *rdma =
1276 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
1277 	INIT_WORK(&rdma->sc_work, __svc_rdma_free);
1278 	queue_work(svc_rdma_wq, &rdma->sc_work);
1279 }
1280 
1281 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
1282 {
1283 	struct svcxprt_rdma *rdma =
1284 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
1285 
1286 	/*
1287 	 * If there are already waiters on the SQ,
1288 	 * return false.
1289 	 */
1290 	if (waitqueue_active(&rdma->sc_send_wait))
1291 		return 0;
1292 
1293 	/* Otherwise return true. */
1294 	return 1;
1295 }
1296 
1297 static int svc_rdma_secure_port(struct svc_rqst *rqstp)
1298 {
1299 	return 1;
1300 }
1301 
1302 static void svc_rdma_kill_temp_xprt(struct svc_xprt *xprt)
1303 {
1304 }
1305 
1306 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1307 {
1308 	struct ib_send_wr *bad_wr, *n_wr;
1309 	int wr_count;
1310 	int i;
1311 	int ret;
1312 
1313 	if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1314 		return -ENOTCONN;
1315 
1316 	wr_count = 1;
1317 	for (n_wr = wr->next; n_wr; n_wr = n_wr->next)
1318 		wr_count++;
1319 
1320 	/* If the SQ is full, wait until an SQ entry is available */
1321 	while (1) {
1322 		if ((atomic_sub_return(wr_count, &xprt->sc_sq_avail) < 0)) {
1323 			atomic_inc(&rdma_stat_sq_starve);
1324 
1325 			/* Wait until SQ WR available if SQ still full */
1326 			atomic_add(wr_count, &xprt->sc_sq_avail);
1327 			wait_event(xprt->sc_send_wait,
1328 				   atomic_read(&xprt->sc_sq_avail) > wr_count);
1329 			if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1330 				return -ENOTCONN;
1331 			continue;
1332 		}
1333 		/* Take a transport ref for each WR posted */
1334 		for (i = 0; i < wr_count; i++)
1335 			svc_xprt_get(&xprt->sc_xprt);
1336 
1337 		/* Bump used SQ WR count and post */
1338 		ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1339 		if (ret) {
1340 			set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
1341 			for (i = 0; i < wr_count; i ++)
1342 				svc_xprt_put(&xprt->sc_xprt);
1343 			dprintk("svcrdma: failed to post SQ WR rc=%d\n", ret);
1344 			dprintk("    sc_sq_avail=%d, sc_sq_depth=%d\n",
1345 				atomic_read(&xprt->sc_sq_avail),
1346 				xprt->sc_sq_depth);
1347 			wake_up(&xprt->sc_send_wait);
1348 		}
1349 		break;
1350 	}
1351 	return ret;
1352 }
1353