• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/debug.h>
45 #include <linux/sunrpc/rpc_rdma.h>
46 #include <linux/interrupt.h>
47 #include <linux/sched.h>
48 #include <linux/slab.h>
49 #include <linux/spinlock.h>
50 #include <linux/workqueue.h>
51 #include <rdma/ib_verbs.h>
52 #include <rdma/rdma_cm.h>
53 #include <linux/sunrpc/svc_rdma.h>
54 #include <linux/export.h>
55 #include "xprt_rdma.h"
56 
57 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
58 
59 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *, int);
60 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
61 					struct net *net,
62 					struct sockaddr *sa, int salen,
63 					int flags);
64 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt);
65 static void svc_rdma_release_rqst(struct svc_rqst *);
66 static void svc_rdma_detach(struct svc_xprt *xprt);
67 static void svc_rdma_free(struct svc_xprt *xprt);
68 static int svc_rdma_has_wspace(struct svc_xprt *xprt);
69 static int svc_rdma_secure_port(struct svc_rqst *);
70 static void svc_rdma_kill_temp_xprt(struct svc_xprt *);
71 
72 static struct svc_xprt_ops svc_rdma_ops = {
73 	.xpo_create = svc_rdma_create,
74 	.xpo_recvfrom = svc_rdma_recvfrom,
75 	.xpo_sendto = svc_rdma_sendto,
76 	.xpo_release_rqst = svc_rdma_release_rqst,
77 	.xpo_detach = svc_rdma_detach,
78 	.xpo_free = svc_rdma_free,
79 	.xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
80 	.xpo_has_wspace = svc_rdma_has_wspace,
81 	.xpo_accept = svc_rdma_accept,
82 	.xpo_secure_port = svc_rdma_secure_port,
83 	.xpo_kill_temp_xprt = svc_rdma_kill_temp_xprt,
84 };
85 
86 struct svc_xprt_class svc_rdma_class = {
87 	.xcl_name = "rdma",
88 	.xcl_owner = THIS_MODULE,
89 	.xcl_ops = &svc_rdma_ops,
90 	.xcl_max_payload = RPCSVC_MAXPAYLOAD_RDMA,
91 	.xcl_ident = XPRT_TRANSPORT_RDMA,
92 };
93 
94 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
95 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *, struct net *,
96 					   struct sockaddr *, int, int);
97 static void svc_rdma_bc_detach(struct svc_xprt *);
98 static void svc_rdma_bc_free(struct svc_xprt *);
99 
100 static struct svc_xprt_ops svc_rdma_bc_ops = {
101 	.xpo_create = svc_rdma_bc_create,
102 	.xpo_detach = svc_rdma_bc_detach,
103 	.xpo_free = svc_rdma_bc_free,
104 	.xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
105 	.xpo_secure_port = svc_rdma_secure_port,
106 };
107 
108 struct svc_xprt_class svc_rdma_bc_class = {
109 	.xcl_name = "rdma-bc",
110 	.xcl_owner = THIS_MODULE,
111 	.xcl_ops = &svc_rdma_bc_ops,
112 	.xcl_max_payload = (1024 - RPCRDMA_HDRLEN_MIN)
113 };
114 
svc_rdma_bc_create(struct svc_serv * serv,struct net * net,struct sockaddr * sa,int salen,int flags)115 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *serv,
116 					   struct net *net,
117 					   struct sockaddr *sa, int salen,
118 					   int flags)
119 {
120 	struct svcxprt_rdma *cma_xprt;
121 	struct svc_xprt *xprt;
122 
123 	cma_xprt = rdma_create_xprt(serv, 0);
124 	if (!cma_xprt)
125 		return ERR_PTR(-ENOMEM);
126 	xprt = &cma_xprt->sc_xprt;
127 
128 	svc_xprt_init(net, &svc_rdma_bc_class, xprt, serv);
129 	serv->sv_bc_xprt = xprt;
130 
131 	dprintk("svcrdma: %s(%p)\n", __func__, xprt);
132 	return xprt;
133 }
134 
svc_rdma_bc_detach(struct svc_xprt * xprt)135 static void svc_rdma_bc_detach(struct svc_xprt *xprt)
136 {
137 	dprintk("svcrdma: %s(%p)\n", __func__, xprt);
138 }
139 
svc_rdma_bc_free(struct svc_xprt * xprt)140 static void svc_rdma_bc_free(struct svc_xprt *xprt)
141 {
142 	struct svcxprt_rdma *rdma =
143 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
144 
145 	dprintk("svcrdma: %s(%p)\n", __func__, xprt);
146 	if (xprt)
147 		kfree(rdma);
148 }
149 #endif	/* CONFIG_SUNRPC_BACKCHANNEL */
150 
alloc_ctxt(struct svcxprt_rdma * xprt,gfp_t flags)151 static struct svc_rdma_op_ctxt *alloc_ctxt(struct svcxprt_rdma *xprt,
152 					   gfp_t flags)
153 {
154 	struct svc_rdma_op_ctxt *ctxt;
155 
156 	ctxt = kmalloc(sizeof(*ctxt), flags);
157 	if (ctxt) {
158 		ctxt->xprt = xprt;
159 		INIT_LIST_HEAD(&ctxt->free);
160 		INIT_LIST_HEAD(&ctxt->dto_q);
161 	}
162 	return ctxt;
163 }
164 
svc_rdma_prealloc_ctxts(struct svcxprt_rdma * xprt)165 static bool svc_rdma_prealloc_ctxts(struct svcxprt_rdma *xprt)
166 {
167 	unsigned int i;
168 
169 	/* Each RPC/RDMA credit can consume a number of send
170 	 * and receive WQEs. One ctxt is allocated for each.
171 	 */
172 	i = xprt->sc_sq_depth + xprt->sc_rq_depth;
173 
174 	while (i--) {
175 		struct svc_rdma_op_ctxt *ctxt;
176 
177 		ctxt = alloc_ctxt(xprt, GFP_KERNEL);
178 		if (!ctxt) {
179 			dprintk("svcrdma: No memory for RDMA ctxt\n");
180 			return false;
181 		}
182 		list_add(&ctxt->free, &xprt->sc_ctxts);
183 	}
184 	return true;
185 }
186 
svc_rdma_get_context(struct svcxprt_rdma * xprt)187 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
188 {
189 	struct svc_rdma_op_ctxt *ctxt = NULL;
190 
191 	spin_lock_bh(&xprt->sc_ctxt_lock);
192 	xprt->sc_ctxt_used++;
193 	if (list_empty(&xprt->sc_ctxts))
194 		goto out_empty;
195 
196 	ctxt = list_first_entry(&xprt->sc_ctxts,
197 				struct svc_rdma_op_ctxt, free);
198 	list_del_init(&ctxt->free);
199 	spin_unlock_bh(&xprt->sc_ctxt_lock);
200 
201 out:
202 	ctxt->count = 0;
203 	ctxt->mapped_sges = 0;
204 	ctxt->frmr = NULL;
205 	return ctxt;
206 
207 out_empty:
208 	/* Either pre-allocation missed the mark, or send
209 	 * queue accounting is broken.
210 	 */
211 	spin_unlock_bh(&xprt->sc_ctxt_lock);
212 
213 	ctxt = alloc_ctxt(xprt, GFP_NOIO);
214 	if (ctxt)
215 		goto out;
216 
217 	spin_lock_bh(&xprt->sc_ctxt_lock);
218 	xprt->sc_ctxt_used--;
219 	spin_unlock_bh(&xprt->sc_ctxt_lock);
220 	WARN_ONCE(1, "svcrdma: empty RDMA ctxt list?\n");
221 	return NULL;
222 }
223 
svc_rdma_unmap_dma(struct svc_rdma_op_ctxt * ctxt)224 void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt)
225 {
226 	struct svcxprt_rdma *xprt = ctxt->xprt;
227 	struct ib_device *device = xprt->sc_cm_id->device;
228 	u32 lkey = xprt->sc_pd->local_dma_lkey;
229 	unsigned int i, count;
230 
231 	for (count = 0, i = 0; i < ctxt->mapped_sges; i++) {
232 		/*
233 		 * Unmap the DMA addr in the SGE if the lkey matches
234 		 * the local_dma_lkey, otherwise, ignore it since it is
235 		 * an FRMR lkey and will be unmapped later when the
236 		 * last WR that uses it completes.
237 		 */
238 		if (ctxt->sge[i].lkey == lkey) {
239 			count++;
240 			ib_dma_unmap_page(device,
241 					    ctxt->sge[i].addr,
242 					    ctxt->sge[i].length,
243 					    ctxt->direction);
244 		}
245 	}
246 	ctxt->mapped_sges = 0;
247 	atomic_sub(count, &xprt->sc_dma_used);
248 }
249 
svc_rdma_put_context(struct svc_rdma_op_ctxt * ctxt,int free_pages)250 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
251 {
252 	struct svcxprt_rdma *xprt = ctxt->xprt;
253 	int i;
254 
255 	if (free_pages)
256 		for (i = 0; i < ctxt->count; i++)
257 			put_page(ctxt->pages[i]);
258 
259 	spin_lock_bh(&xprt->sc_ctxt_lock);
260 	xprt->sc_ctxt_used--;
261 	list_add(&ctxt->free, &xprt->sc_ctxts);
262 	spin_unlock_bh(&xprt->sc_ctxt_lock);
263 }
264 
svc_rdma_destroy_ctxts(struct svcxprt_rdma * xprt)265 static void svc_rdma_destroy_ctxts(struct svcxprt_rdma *xprt)
266 {
267 	while (!list_empty(&xprt->sc_ctxts)) {
268 		struct svc_rdma_op_ctxt *ctxt;
269 
270 		ctxt = list_first_entry(&xprt->sc_ctxts,
271 					struct svc_rdma_op_ctxt, free);
272 		list_del(&ctxt->free);
273 		kfree(ctxt);
274 	}
275 }
276 
alloc_req_map(gfp_t flags)277 static struct svc_rdma_req_map *alloc_req_map(gfp_t flags)
278 {
279 	struct svc_rdma_req_map *map;
280 
281 	map = kmalloc(sizeof(*map), flags);
282 	if (map)
283 		INIT_LIST_HEAD(&map->free);
284 	return map;
285 }
286 
svc_rdma_prealloc_maps(struct svcxprt_rdma * xprt)287 static bool svc_rdma_prealloc_maps(struct svcxprt_rdma *xprt)
288 {
289 	unsigned int i;
290 
291 	/* One for each receive buffer on this connection. */
292 	i = xprt->sc_max_requests;
293 
294 	while (i--) {
295 		struct svc_rdma_req_map *map;
296 
297 		map = alloc_req_map(GFP_KERNEL);
298 		if (!map) {
299 			dprintk("svcrdma: No memory for request map\n");
300 			return false;
301 		}
302 		list_add(&map->free, &xprt->sc_maps);
303 	}
304 	return true;
305 }
306 
svc_rdma_get_req_map(struct svcxprt_rdma * xprt)307 struct svc_rdma_req_map *svc_rdma_get_req_map(struct svcxprt_rdma *xprt)
308 {
309 	struct svc_rdma_req_map *map = NULL;
310 
311 	spin_lock(&xprt->sc_map_lock);
312 	if (list_empty(&xprt->sc_maps))
313 		goto out_empty;
314 
315 	map = list_first_entry(&xprt->sc_maps,
316 			       struct svc_rdma_req_map, free);
317 	list_del_init(&map->free);
318 	spin_unlock(&xprt->sc_map_lock);
319 
320 out:
321 	map->count = 0;
322 	return map;
323 
324 out_empty:
325 	spin_unlock(&xprt->sc_map_lock);
326 
327 	/* Pre-allocation amount was incorrect */
328 	map = alloc_req_map(GFP_NOIO);
329 	if (map)
330 		goto out;
331 
332 	WARN_ONCE(1, "svcrdma: empty request map list?\n");
333 	return NULL;
334 }
335 
svc_rdma_put_req_map(struct svcxprt_rdma * xprt,struct svc_rdma_req_map * map)336 void svc_rdma_put_req_map(struct svcxprt_rdma *xprt,
337 			  struct svc_rdma_req_map *map)
338 {
339 	spin_lock(&xprt->sc_map_lock);
340 	list_add(&map->free, &xprt->sc_maps);
341 	spin_unlock(&xprt->sc_map_lock);
342 }
343 
svc_rdma_destroy_maps(struct svcxprt_rdma * xprt)344 static void svc_rdma_destroy_maps(struct svcxprt_rdma *xprt)
345 {
346 	while (!list_empty(&xprt->sc_maps)) {
347 		struct svc_rdma_req_map *map;
348 
349 		map = list_first_entry(&xprt->sc_maps,
350 				       struct svc_rdma_req_map, free);
351 		list_del(&map->free);
352 		kfree(map);
353 	}
354 }
355 
356 /* QP event handler */
qp_event_handler(struct ib_event * event,void * context)357 static void qp_event_handler(struct ib_event *event, void *context)
358 {
359 	struct svc_xprt *xprt = context;
360 
361 	switch (event->event) {
362 	/* These are considered benign events */
363 	case IB_EVENT_PATH_MIG:
364 	case IB_EVENT_COMM_EST:
365 	case IB_EVENT_SQ_DRAINED:
366 	case IB_EVENT_QP_LAST_WQE_REACHED:
367 		dprintk("svcrdma: QP event %s (%d) received for QP=%p\n",
368 			ib_event_msg(event->event), event->event,
369 			event->element.qp);
370 		break;
371 	/* These are considered fatal events */
372 	case IB_EVENT_PATH_MIG_ERR:
373 	case IB_EVENT_QP_FATAL:
374 	case IB_EVENT_QP_REQ_ERR:
375 	case IB_EVENT_QP_ACCESS_ERR:
376 	case IB_EVENT_DEVICE_FATAL:
377 	default:
378 		dprintk("svcrdma: QP ERROR event %s (%d) received for QP=%p, "
379 			"closing transport\n",
380 			ib_event_msg(event->event), event->event,
381 			event->element.qp);
382 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
383 		break;
384 	}
385 }
386 
387 /**
388  * svc_rdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
389  * @cq:        completion queue
390  * @wc:        completed WR
391  *
392  */
svc_rdma_wc_receive(struct ib_cq * cq,struct ib_wc * wc)393 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
394 {
395 	struct svcxprt_rdma *xprt = cq->cq_context;
396 	struct ib_cqe *cqe = wc->wr_cqe;
397 	struct svc_rdma_op_ctxt *ctxt;
398 
399 	/* WARNING: Only wc->wr_cqe and wc->status are reliable */
400 	ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
401 	ctxt->wc_status = wc->status;
402 	svc_rdma_unmap_dma(ctxt);
403 
404 	if (wc->status != IB_WC_SUCCESS)
405 		goto flushed;
406 
407 	/* All wc fields are now known to be valid */
408 	ctxt->byte_len = wc->byte_len;
409 	spin_lock(&xprt->sc_rq_dto_lock);
410 	list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
411 	spin_unlock(&xprt->sc_rq_dto_lock);
412 
413 	set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
414 	if (test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
415 		goto out;
416 	svc_xprt_enqueue(&xprt->sc_xprt);
417 	goto out;
418 
419 flushed:
420 	if (wc->status != IB_WC_WR_FLUSH_ERR)
421 		pr_warn("svcrdma: receive: %s (%u/0x%x)\n",
422 			ib_wc_status_msg(wc->status),
423 			wc->status, wc->vendor_err);
424 	set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
425 	svc_rdma_put_context(ctxt, 1);
426 
427 out:
428 	svc_xprt_put(&xprt->sc_xprt);
429 }
430 
svc_rdma_send_wc_common(struct svcxprt_rdma * xprt,struct ib_wc * wc,const char * opname)431 static void svc_rdma_send_wc_common(struct svcxprt_rdma *xprt,
432 				    struct ib_wc *wc,
433 				    const char *opname)
434 {
435 	if (wc->status != IB_WC_SUCCESS)
436 		goto err;
437 
438 out:
439 	atomic_dec(&xprt->sc_sq_count);
440 	wake_up(&xprt->sc_send_wait);
441 	return;
442 
443 err:
444 	set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
445 	if (wc->status != IB_WC_WR_FLUSH_ERR)
446 		pr_err("svcrdma: %s: %s (%u/0x%x)\n",
447 		       opname, ib_wc_status_msg(wc->status),
448 		       wc->status, wc->vendor_err);
449 	goto out;
450 }
451 
svc_rdma_send_wc_common_put(struct ib_cq * cq,struct ib_wc * wc,const char * opname)452 static void svc_rdma_send_wc_common_put(struct ib_cq *cq, struct ib_wc *wc,
453 					const char *opname)
454 {
455 	struct svcxprt_rdma *xprt = cq->cq_context;
456 
457 	svc_rdma_send_wc_common(xprt, wc, opname);
458 	svc_xprt_put(&xprt->sc_xprt);
459 }
460 
461 /**
462  * svc_rdma_wc_send - Invoked by RDMA provider for each polled Send WC
463  * @cq:        completion queue
464  * @wc:        completed WR
465  *
466  */
svc_rdma_wc_send(struct ib_cq * cq,struct ib_wc * wc)467 void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
468 {
469 	struct ib_cqe *cqe = wc->wr_cqe;
470 	struct svc_rdma_op_ctxt *ctxt;
471 
472 	svc_rdma_send_wc_common_put(cq, wc, "send");
473 
474 	ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
475 	svc_rdma_unmap_dma(ctxt);
476 	svc_rdma_put_context(ctxt, 1);
477 }
478 
479 /**
480  * svc_rdma_wc_write - Invoked by RDMA provider for each polled Write WC
481  * @cq:        completion queue
482  * @wc:        completed WR
483  *
484  */
svc_rdma_wc_write(struct ib_cq * cq,struct ib_wc * wc)485 void svc_rdma_wc_write(struct ib_cq *cq, struct ib_wc *wc)
486 {
487 	struct ib_cqe *cqe = wc->wr_cqe;
488 	struct svc_rdma_op_ctxt *ctxt;
489 
490 	svc_rdma_send_wc_common_put(cq, wc, "write");
491 
492 	ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
493 	svc_rdma_unmap_dma(ctxt);
494 	svc_rdma_put_context(ctxt, 0);
495 }
496 
497 /**
498  * svc_rdma_wc_reg - Invoked by RDMA provider for each polled FASTREG WC
499  * @cq:        completion queue
500  * @wc:        completed WR
501  *
502  */
svc_rdma_wc_reg(struct ib_cq * cq,struct ib_wc * wc)503 void svc_rdma_wc_reg(struct ib_cq *cq, struct ib_wc *wc)
504 {
505 	svc_rdma_send_wc_common_put(cq, wc, "fastreg");
506 }
507 
508 /**
509  * svc_rdma_wc_read - Invoked by RDMA provider for each polled Read WC
510  * @cq:        completion queue
511  * @wc:        completed WR
512  *
513  */
svc_rdma_wc_read(struct ib_cq * cq,struct ib_wc * wc)514 void svc_rdma_wc_read(struct ib_cq *cq, struct ib_wc *wc)
515 {
516 	struct svcxprt_rdma *xprt = cq->cq_context;
517 	struct ib_cqe *cqe = wc->wr_cqe;
518 	struct svc_rdma_op_ctxt *ctxt;
519 
520 	svc_rdma_send_wc_common(xprt, wc, "read");
521 
522 	ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
523 	svc_rdma_unmap_dma(ctxt);
524 	svc_rdma_put_frmr(xprt, ctxt->frmr);
525 
526 	if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
527 		struct svc_rdma_op_ctxt *read_hdr;
528 
529 		read_hdr = ctxt->read_hdr;
530 		spin_lock(&xprt->sc_rq_dto_lock);
531 		list_add_tail(&read_hdr->dto_q,
532 			      &xprt->sc_read_complete_q);
533 		spin_unlock(&xprt->sc_rq_dto_lock);
534 
535 		set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
536 		svc_xprt_enqueue(&xprt->sc_xprt);
537 	}
538 
539 	svc_rdma_put_context(ctxt, 0);
540 	svc_xprt_put(&xprt->sc_xprt);
541 }
542 
543 /**
544  * svc_rdma_wc_inv - Invoked by RDMA provider for each polled LOCAL_INV WC
545  * @cq:        completion queue
546  * @wc:        completed WR
547  *
548  */
svc_rdma_wc_inv(struct ib_cq * cq,struct ib_wc * wc)549 void svc_rdma_wc_inv(struct ib_cq *cq, struct ib_wc *wc)
550 {
551 	svc_rdma_send_wc_common_put(cq, wc, "localInv");
552 }
553 
rdma_create_xprt(struct svc_serv * serv,int listener)554 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
555 					     int listener)
556 {
557 	struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
558 
559 	if (!cma_xprt)
560 		return NULL;
561 	svc_xprt_init(&init_net, &svc_rdma_class, &cma_xprt->sc_xprt, serv);
562 	INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
563 	INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
564 	INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
565 	INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
566 	INIT_LIST_HEAD(&cma_xprt->sc_frmr_q);
567 	INIT_LIST_HEAD(&cma_xprt->sc_ctxts);
568 	INIT_LIST_HEAD(&cma_xprt->sc_maps);
569 	init_waitqueue_head(&cma_xprt->sc_send_wait);
570 
571 	spin_lock_init(&cma_xprt->sc_lock);
572 	spin_lock_init(&cma_xprt->sc_rq_dto_lock);
573 	spin_lock_init(&cma_xprt->sc_frmr_q_lock);
574 	spin_lock_init(&cma_xprt->sc_ctxt_lock);
575 	spin_lock_init(&cma_xprt->sc_map_lock);
576 
577 	if (listener)
578 		set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
579 
580 	return cma_xprt;
581 }
582 
svc_rdma_post_recv(struct svcxprt_rdma * xprt,gfp_t flags)583 int svc_rdma_post_recv(struct svcxprt_rdma *xprt, gfp_t flags)
584 {
585 	struct ib_recv_wr recv_wr, *bad_recv_wr;
586 	struct svc_rdma_op_ctxt *ctxt;
587 	struct page *page;
588 	dma_addr_t pa;
589 	int sge_no;
590 	int buflen;
591 	int ret;
592 
593 	ctxt = svc_rdma_get_context(xprt);
594 	buflen = 0;
595 	ctxt->direction = DMA_FROM_DEVICE;
596 	ctxt->cqe.done = svc_rdma_wc_receive;
597 	for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
598 		if (sge_no >= xprt->sc_max_sge) {
599 			pr_err("svcrdma: Too many sges (%d)\n", sge_no);
600 			goto err_put_ctxt;
601 		}
602 		page = alloc_page(flags);
603 		if (!page)
604 			goto err_put_ctxt;
605 		ctxt->pages[sge_no] = page;
606 		pa = ib_dma_map_page(xprt->sc_cm_id->device,
607 				     page, 0, PAGE_SIZE,
608 				     DMA_FROM_DEVICE);
609 		if (ib_dma_mapping_error(xprt->sc_cm_id->device, pa))
610 			goto err_put_ctxt;
611 		svc_rdma_count_mappings(xprt, ctxt);
612 		ctxt->sge[sge_no].addr = pa;
613 		ctxt->sge[sge_no].length = PAGE_SIZE;
614 		ctxt->sge[sge_no].lkey = xprt->sc_pd->local_dma_lkey;
615 		ctxt->count = sge_no + 1;
616 		buflen += PAGE_SIZE;
617 	}
618 	recv_wr.next = NULL;
619 	recv_wr.sg_list = &ctxt->sge[0];
620 	recv_wr.num_sge = ctxt->count;
621 	recv_wr.wr_cqe = &ctxt->cqe;
622 
623 	svc_xprt_get(&xprt->sc_xprt);
624 	ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
625 	if (ret) {
626 		svc_rdma_unmap_dma(ctxt);
627 		svc_rdma_put_context(ctxt, 1);
628 		svc_xprt_put(&xprt->sc_xprt);
629 	}
630 	return ret;
631 
632  err_put_ctxt:
633 	svc_rdma_unmap_dma(ctxt);
634 	svc_rdma_put_context(ctxt, 1);
635 	return -ENOMEM;
636 }
637 
svc_rdma_repost_recv(struct svcxprt_rdma * xprt,gfp_t flags)638 int svc_rdma_repost_recv(struct svcxprt_rdma *xprt, gfp_t flags)
639 {
640 	int ret = 0;
641 
642 	ret = svc_rdma_post_recv(xprt, flags);
643 	if (ret) {
644 		pr_err("svcrdma: could not post a receive buffer, err=%d.\n",
645 		       ret);
646 		pr_err("svcrdma: closing transport %p.\n", xprt);
647 		set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
648 		ret = -ENOTCONN;
649 	}
650 	return ret;
651 }
652 
653 static void
svc_rdma_parse_connect_private(struct svcxprt_rdma * newxprt,struct rdma_conn_param * param)654 svc_rdma_parse_connect_private(struct svcxprt_rdma *newxprt,
655 			       struct rdma_conn_param *param)
656 {
657 	const struct rpcrdma_connect_private *pmsg = param->private_data;
658 
659 	if (pmsg &&
660 	    pmsg->cp_magic == rpcrdma_cmp_magic &&
661 	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
662 		newxprt->sc_snd_w_inv = pmsg->cp_flags &
663 					RPCRDMA_CMP_F_SND_W_INV_OK;
664 
665 		dprintk("svcrdma: client send_size %u, recv_size %u "
666 			"remote inv %ssupported\n",
667 			rpcrdma_decode_buffer_size(pmsg->cp_send_size),
668 			rpcrdma_decode_buffer_size(pmsg->cp_recv_size),
669 			newxprt->sc_snd_w_inv ? "" : "un");
670 	}
671 }
672 
673 /*
674  * This function handles the CONNECT_REQUEST event on a listening
675  * endpoint. It is passed the cma_id for the _new_ connection. The context in
676  * this cma_id is inherited from the listening cma_id and is the svc_xprt
677  * structure for the listening endpoint.
678  *
679  * This function creates a new xprt for the new connection and enqueues it on
680  * the accept queue for the listent xprt. When the listen thread is kicked, it
681  * will call the recvfrom method on the listen xprt which will accept the new
682  * connection.
683  */
handle_connect_req(struct rdma_cm_id * new_cma_id,struct rdma_conn_param * param)684 static void handle_connect_req(struct rdma_cm_id *new_cma_id,
685 			       struct rdma_conn_param *param)
686 {
687 	struct svcxprt_rdma *listen_xprt = new_cma_id->context;
688 	struct svcxprt_rdma *newxprt;
689 	struct sockaddr *sa;
690 
691 	/* Create a new transport */
692 	newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
693 	if (!newxprt) {
694 		dprintk("svcrdma: failed to create new transport\n");
695 		return;
696 	}
697 	newxprt->sc_cm_id = new_cma_id;
698 	new_cma_id->context = newxprt;
699 	dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
700 		newxprt, newxprt->sc_cm_id, listen_xprt);
701 	svc_rdma_parse_connect_private(newxprt, param);
702 
703 	/* Save client advertised inbound read limit for use later in accept. */
704 	newxprt->sc_ord = param->initiator_depth;
705 
706 	/* Set the local and remote addresses in the transport */
707 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
708 	svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
709 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
710 	svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
711 
712 	/*
713 	 * Enqueue the new transport on the accept queue of the listening
714 	 * transport
715 	 */
716 	spin_lock_bh(&listen_xprt->sc_lock);
717 	list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
718 	spin_unlock_bh(&listen_xprt->sc_lock);
719 
720 	set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
721 	svc_xprt_enqueue(&listen_xprt->sc_xprt);
722 }
723 
724 /*
725  * Handles events generated on the listening endpoint. These events will be
726  * either be incoming connect requests or adapter removal  events.
727  */
rdma_listen_handler(struct rdma_cm_id * cma_id,struct rdma_cm_event * event)728 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
729 			       struct rdma_cm_event *event)
730 {
731 	struct svcxprt_rdma *xprt = cma_id->context;
732 	int ret = 0;
733 
734 	switch (event->event) {
735 	case RDMA_CM_EVENT_CONNECT_REQUEST:
736 		dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
737 			"event = %s (%d)\n", cma_id, cma_id->context,
738 			rdma_event_msg(event->event), event->event);
739 		handle_connect_req(cma_id, &event->param.conn);
740 		break;
741 
742 	case RDMA_CM_EVENT_ESTABLISHED:
743 		/* Accept complete */
744 		dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
745 			"cm_id=%p\n", xprt, cma_id);
746 		break;
747 
748 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
749 		dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
750 			xprt, cma_id);
751 		if (xprt)
752 			set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
753 		break;
754 
755 	default:
756 		dprintk("svcrdma: Unexpected event on listening endpoint %p, "
757 			"event = %s (%d)\n", cma_id,
758 			rdma_event_msg(event->event), event->event);
759 		break;
760 	}
761 
762 	return ret;
763 }
764 
rdma_cma_handler(struct rdma_cm_id * cma_id,struct rdma_cm_event * event)765 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
766 			    struct rdma_cm_event *event)
767 {
768 	struct svc_xprt *xprt = cma_id->context;
769 	struct svcxprt_rdma *rdma =
770 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
771 	switch (event->event) {
772 	case RDMA_CM_EVENT_ESTABLISHED:
773 		/* Accept complete */
774 		svc_xprt_get(xprt);
775 		dprintk("svcrdma: Connection completed on DTO xprt=%p, "
776 			"cm_id=%p\n", xprt, cma_id);
777 		clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
778 		svc_xprt_enqueue(xprt);
779 		break;
780 	case RDMA_CM_EVENT_DISCONNECTED:
781 		dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
782 			xprt, cma_id);
783 		if (xprt) {
784 			set_bit(XPT_CLOSE, &xprt->xpt_flags);
785 			svc_xprt_enqueue(xprt);
786 			svc_xprt_put(xprt);
787 		}
788 		break;
789 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
790 		dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
791 			"event = %s (%d)\n", cma_id, xprt,
792 			rdma_event_msg(event->event), event->event);
793 		if (xprt) {
794 			set_bit(XPT_CLOSE, &xprt->xpt_flags);
795 			svc_xprt_enqueue(xprt);
796 			svc_xprt_put(xprt);
797 		}
798 		break;
799 	default:
800 		dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
801 			"event = %s (%d)\n", cma_id,
802 			rdma_event_msg(event->event), event->event);
803 		break;
804 	}
805 	return 0;
806 }
807 
808 /*
809  * Create a listening RDMA service endpoint.
810  */
svc_rdma_create(struct svc_serv * serv,struct net * net,struct sockaddr * sa,int salen,int flags)811 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
812 					struct net *net,
813 					struct sockaddr *sa, int salen,
814 					int flags)
815 {
816 	struct rdma_cm_id *listen_id;
817 	struct svcxprt_rdma *cma_xprt;
818 	int ret;
819 
820 	dprintk("svcrdma: Creating RDMA socket\n");
821 	if ((sa->sa_family != AF_INET) && (sa->sa_family != AF_INET6)) {
822 		dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family);
823 		return ERR_PTR(-EAFNOSUPPORT);
824 	}
825 	cma_xprt = rdma_create_xprt(serv, 1);
826 	if (!cma_xprt)
827 		return ERR_PTR(-ENOMEM);
828 
829 	listen_id = rdma_create_id(&init_net, rdma_listen_handler, cma_xprt,
830 				   RDMA_PS_TCP, IB_QPT_RC);
831 	if (IS_ERR(listen_id)) {
832 		ret = PTR_ERR(listen_id);
833 		dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
834 		goto err0;
835 	}
836 
837 	/* Allow both IPv4 and IPv6 sockets to bind a single port
838 	 * at the same time.
839 	 */
840 #if IS_ENABLED(CONFIG_IPV6)
841 	ret = rdma_set_afonly(listen_id, 1);
842 	if (ret) {
843 		dprintk("svcrdma: rdma_set_afonly failed = %d\n", ret);
844 		goto err1;
845 	}
846 #endif
847 	ret = rdma_bind_addr(listen_id, sa);
848 	if (ret) {
849 		dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
850 		goto err1;
851 	}
852 	cma_xprt->sc_cm_id = listen_id;
853 
854 	ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
855 	if (ret) {
856 		dprintk("svcrdma: rdma_listen failed = %d\n", ret);
857 		goto err1;
858 	}
859 
860 	/*
861 	 * We need to use the address from the cm_id in case the
862 	 * caller specified 0 for the port number.
863 	 */
864 	sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
865 	svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
866 
867 	return &cma_xprt->sc_xprt;
868 
869  err1:
870 	rdma_destroy_id(listen_id);
871  err0:
872 	kfree(cma_xprt);
873 	return ERR_PTR(ret);
874 }
875 
rdma_alloc_frmr(struct svcxprt_rdma * xprt)876 static struct svc_rdma_fastreg_mr *rdma_alloc_frmr(struct svcxprt_rdma *xprt)
877 {
878 	struct ib_mr *mr;
879 	struct scatterlist *sg;
880 	struct svc_rdma_fastreg_mr *frmr;
881 	u32 num_sg;
882 
883 	frmr = kmalloc(sizeof(*frmr), GFP_KERNEL);
884 	if (!frmr)
885 		goto err;
886 
887 	num_sg = min_t(u32, RPCSVC_MAXPAGES, xprt->sc_frmr_pg_list_len);
888 	mr = ib_alloc_mr(xprt->sc_pd, IB_MR_TYPE_MEM_REG, num_sg);
889 	if (IS_ERR(mr))
890 		goto err_free_frmr;
891 
892 	sg = kcalloc(RPCSVC_MAXPAGES, sizeof(*sg), GFP_KERNEL);
893 	if (!sg)
894 		goto err_free_mr;
895 
896 	sg_init_table(sg, RPCSVC_MAXPAGES);
897 
898 	frmr->mr = mr;
899 	frmr->sg = sg;
900 	INIT_LIST_HEAD(&frmr->frmr_list);
901 	return frmr;
902 
903  err_free_mr:
904 	ib_dereg_mr(mr);
905  err_free_frmr:
906 	kfree(frmr);
907  err:
908 	return ERR_PTR(-ENOMEM);
909 }
910 
rdma_dealloc_frmr_q(struct svcxprt_rdma * xprt)911 static void rdma_dealloc_frmr_q(struct svcxprt_rdma *xprt)
912 {
913 	struct svc_rdma_fastreg_mr *frmr;
914 
915 	while (!list_empty(&xprt->sc_frmr_q)) {
916 		frmr = list_entry(xprt->sc_frmr_q.next,
917 				  struct svc_rdma_fastreg_mr, frmr_list);
918 		list_del_init(&frmr->frmr_list);
919 		kfree(frmr->sg);
920 		ib_dereg_mr(frmr->mr);
921 		kfree(frmr);
922 	}
923 }
924 
svc_rdma_get_frmr(struct svcxprt_rdma * rdma)925 struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma)
926 {
927 	struct svc_rdma_fastreg_mr *frmr = NULL;
928 
929 	spin_lock_bh(&rdma->sc_frmr_q_lock);
930 	if (!list_empty(&rdma->sc_frmr_q)) {
931 		frmr = list_entry(rdma->sc_frmr_q.next,
932 				  struct svc_rdma_fastreg_mr, frmr_list);
933 		list_del_init(&frmr->frmr_list);
934 		frmr->sg_nents = 0;
935 	}
936 	spin_unlock_bh(&rdma->sc_frmr_q_lock);
937 	if (frmr)
938 		return frmr;
939 
940 	return rdma_alloc_frmr(rdma);
941 }
942 
svc_rdma_put_frmr(struct svcxprt_rdma * rdma,struct svc_rdma_fastreg_mr * frmr)943 void svc_rdma_put_frmr(struct svcxprt_rdma *rdma,
944 		       struct svc_rdma_fastreg_mr *frmr)
945 {
946 	if (frmr) {
947 		ib_dma_unmap_sg(rdma->sc_cm_id->device,
948 				frmr->sg, frmr->sg_nents, frmr->direction);
949 		atomic_dec(&rdma->sc_dma_used);
950 		spin_lock_bh(&rdma->sc_frmr_q_lock);
951 		WARN_ON_ONCE(!list_empty(&frmr->frmr_list));
952 		list_add(&frmr->frmr_list, &rdma->sc_frmr_q);
953 		spin_unlock_bh(&rdma->sc_frmr_q_lock);
954 	}
955 }
956 
957 /*
958  * This is the xpo_recvfrom function for listening endpoints. Its
959  * purpose is to accept incoming connections. The CMA callback handler
960  * has already created a new transport and attached it to the new CMA
961  * ID.
962  *
963  * There is a queue of pending connections hung on the listening
964  * transport. This queue contains the new svc_xprt structure. This
965  * function takes svc_xprt structures off the accept_q and completes
966  * the connection.
967  */
svc_rdma_accept(struct svc_xprt * xprt)968 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
969 {
970 	struct svcxprt_rdma *listen_rdma;
971 	struct svcxprt_rdma *newxprt = NULL;
972 	struct rdma_conn_param conn_param;
973 	struct rpcrdma_connect_private pmsg;
974 	struct ib_qp_init_attr qp_attr;
975 	struct ib_device *dev;
976 	unsigned int i;
977 	int ret = 0;
978 
979 	listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
980 	clear_bit(XPT_CONN, &xprt->xpt_flags);
981 	/* Get the next entry off the accept list */
982 	spin_lock_bh(&listen_rdma->sc_lock);
983 	if (!list_empty(&listen_rdma->sc_accept_q)) {
984 		newxprt = list_entry(listen_rdma->sc_accept_q.next,
985 				     struct svcxprt_rdma, sc_accept_q);
986 		list_del_init(&newxprt->sc_accept_q);
987 	}
988 	if (!list_empty(&listen_rdma->sc_accept_q))
989 		set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
990 	spin_unlock_bh(&listen_rdma->sc_lock);
991 	if (!newxprt)
992 		return NULL;
993 
994 	dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
995 		newxprt, newxprt->sc_cm_id);
996 
997 	dev = newxprt->sc_cm_id->device;
998 
999 	/* Qualify the transport resource defaults with the
1000 	 * capabilities of this particular device */
1001 	newxprt->sc_max_sge = min((size_t)dev->attrs.max_sge,
1002 				  (size_t)RPCSVC_MAXPAGES);
1003 	newxprt->sc_max_sge_rd = min_t(size_t, dev->attrs.max_sge_rd,
1004 				       RPCSVC_MAXPAGES);
1005 	newxprt->sc_max_req_size = svcrdma_max_req_size;
1006 	newxprt->sc_max_requests = min_t(u32, dev->attrs.max_qp_wr,
1007 					 svcrdma_max_requests);
1008 	newxprt->sc_max_bc_requests = min_t(u32, dev->attrs.max_qp_wr,
1009 					    svcrdma_max_bc_requests);
1010 	newxprt->sc_rq_depth = newxprt->sc_max_requests +
1011 			       newxprt->sc_max_bc_requests;
1012 	newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_rq_depth;
1013 
1014 	if (!svc_rdma_prealloc_ctxts(newxprt))
1015 		goto errout;
1016 	if (!svc_rdma_prealloc_maps(newxprt))
1017 		goto errout;
1018 
1019 	/*
1020 	 * Limit ORD based on client limit, local device limit, and
1021 	 * configured svcrdma limit.
1022 	 */
1023 	newxprt->sc_ord = min_t(size_t, dev->attrs.max_qp_rd_atom, newxprt->sc_ord);
1024 	newxprt->sc_ord = min_t(size_t,	svcrdma_ord, newxprt->sc_ord);
1025 
1026 	newxprt->sc_pd = ib_alloc_pd(dev, 0);
1027 	if (IS_ERR(newxprt->sc_pd)) {
1028 		dprintk("svcrdma: error creating PD for connect request\n");
1029 		goto errout;
1030 	}
1031 	newxprt->sc_sq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_sq_depth,
1032 					0, IB_POLL_SOFTIRQ);
1033 	if (IS_ERR(newxprt->sc_sq_cq)) {
1034 		dprintk("svcrdma: error creating SQ CQ for connect request\n");
1035 		goto errout;
1036 	}
1037 	newxprt->sc_rq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_rq_depth,
1038 					0, IB_POLL_SOFTIRQ);
1039 	if (IS_ERR(newxprt->sc_rq_cq)) {
1040 		dprintk("svcrdma: error creating RQ CQ for connect request\n");
1041 		goto errout;
1042 	}
1043 
1044 	memset(&qp_attr, 0, sizeof qp_attr);
1045 	qp_attr.event_handler = qp_event_handler;
1046 	qp_attr.qp_context = &newxprt->sc_xprt;
1047 	qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
1048 	qp_attr.cap.max_recv_wr = newxprt->sc_rq_depth;
1049 	qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
1050 	qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
1051 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1052 	qp_attr.qp_type = IB_QPT_RC;
1053 	qp_attr.send_cq = newxprt->sc_sq_cq;
1054 	qp_attr.recv_cq = newxprt->sc_rq_cq;
1055 	dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
1056 		"    cm_id->device=%p, sc_pd->device=%p\n"
1057 		"    cap.max_send_wr = %d\n"
1058 		"    cap.max_recv_wr = %d\n"
1059 		"    cap.max_send_sge = %d\n"
1060 		"    cap.max_recv_sge = %d\n",
1061 		newxprt->sc_cm_id, newxprt->sc_pd,
1062 		dev, newxprt->sc_pd->device,
1063 		qp_attr.cap.max_send_wr,
1064 		qp_attr.cap.max_recv_wr,
1065 		qp_attr.cap.max_send_sge,
1066 		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 with the following "
1150 		"attributes:\n"
1151 		"    local_ip        : %pI4\n"
1152 		"    local_port	     : %d\n"
1153 		"    remote_ip       : %pI4\n"
1154 		"    remote_port     : %d\n"
1155 		"    max_sge         : %d\n"
1156 		"    max_sge_rd      : %d\n"
1157 		"    sq_depth        : %d\n"
1158 		"    max_requests    : %d\n"
1159 		"    ord             : %d\n",
1160 		newxprt,
1161 		&((struct sockaddr_in *)&newxprt->sc_cm_id->
1162 			 route.addr.src_addr)->sin_addr.s_addr,
1163 		ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
1164 		       route.addr.src_addr)->sin_port),
1165 		&((struct sockaddr_in *)&newxprt->sc_cm_id->
1166 			 route.addr.dst_addr)->sin_addr.s_addr,
1167 		ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
1168 		       route.addr.dst_addr)->sin_port),
1169 		newxprt->sc_max_sge,
1170 		newxprt->sc_max_sge_rd,
1171 		newxprt->sc_sq_depth,
1172 		newxprt->sc_max_requests,
1173 		newxprt->sc_ord);
1174 
1175 	return &newxprt->sc_xprt;
1176 
1177  errout:
1178 	dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
1179 	/* Take a reference in case the DTO handler runs */
1180 	svc_xprt_get(&newxprt->sc_xprt);
1181 	if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
1182 		ib_destroy_qp(newxprt->sc_qp);
1183 	rdma_destroy_id(newxprt->sc_cm_id);
1184 	/* This call to put will destroy the transport */
1185 	svc_xprt_put(&newxprt->sc_xprt);
1186 	return NULL;
1187 }
1188 
svc_rdma_release_rqst(struct svc_rqst * rqstp)1189 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
1190 {
1191 }
1192 
1193 /*
1194  * When connected, an svc_xprt has at least two references:
1195  *
1196  * - A reference held by the cm_id between the ESTABLISHED and
1197  *   DISCONNECTED events. If the remote peer disconnected first, this
1198  *   reference could be gone.
1199  *
1200  * - A reference held by the svc_recv code that called this function
1201  *   as part of close processing.
1202  *
1203  * At a minimum one references should still be held.
1204  */
svc_rdma_detach(struct svc_xprt * xprt)1205 static void svc_rdma_detach(struct svc_xprt *xprt)
1206 {
1207 	struct svcxprt_rdma *rdma =
1208 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
1209 	dprintk("svc: svc_rdma_detach(%p)\n", xprt);
1210 
1211 	/* Disconnect and flush posted WQE */
1212 	rdma_disconnect(rdma->sc_cm_id);
1213 }
1214 
__svc_rdma_free(struct work_struct * work)1215 static void __svc_rdma_free(struct work_struct *work)
1216 {
1217 	struct svcxprt_rdma *rdma =
1218 		container_of(work, struct svcxprt_rdma, sc_work);
1219 	struct svc_xprt *xprt = &rdma->sc_xprt;
1220 
1221 	dprintk("svcrdma: %s(%p)\n", __func__, rdma);
1222 
1223 	if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
1224 		ib_drain_qp(rdma->sc_qp);
1225 
1226 	/* We should only be called from kref_put */
1227 	if (atomic_read(&xprt->xpt_ref.refcount) != 0)
1228 		pr_err("svcrdma: sc_xprt still in use? (%d)\n",
1229 		       atomic_read(&xprt->xpt_ref.refcount));
1230 
1231 	/*
1232 	 * Destroy queued, but not processed read completions. Note
1233 	 * that this cleanup has to be done before destroying the
1234 	 * cm_id because the device ptr is needed to unmap the dma in
1235 	 * svc_rdma_put_context.
1236 	 */
1237 	while (!list_empty(&rdma->sc_read_complete_q)) {
1238 		struct svc_rdma_op_ctxt *ctxt;
1239 		ctxt = list_entry(rdma->sc_read_complete_q.next,
1240 				  struct svc_rdma_op_ctxt,
1241 				  dto_q);
1242 		list_del_init(&ctxt->dto_q);
1243 		svc_rdma_put_context(ctxt, 1);
1244 	}
1245 
1246 	/* Destroy queued, but not processed recv completions */
1247 	while (!list_empty(&rdma->sc_rq_dto_q)) {
1248 		struct svc_rdma_op_ctxt *ctxt;
1249 		ctxt = list_entry(rdma->sc_rq_dto_q.next,
1250 				  struct svc_rdma_op_ctxt,
1251 				  dto_q);
1252 		list_del_init(&ctxt->dto_q);
1253 		svc_rdma_put_context(ctxt, 1);
1254 	}
1255 
1256 	/* Warn if we leaked a resource or under-referenced */
1257 	if (rdma->sc_ctxt_used != 0)
1258 		pr_err("svcrdma: ctxt still in use? (%d)\n",
1259 		       rdma->sc_ctxt_used);
1260 	if (atomic_read(&rdma->sc_dma_used) != 0)
1261 		pr_err("svcrdma: dma still in use? (%d)\n",
1262 		       atomic_read(&rdma->sc_dma_used));
1263 
1264 	/* Final put of backchannel client transport */
1265 	if (xprt->xpt_bc_xprt) {
1266 		xprt_put(xprt->xpt_bc_xprt);
1267 		xprt->xpt_bc_xprt = NULL;
1268 	}
1269 
1270 	rdma_dealloc_frmr_q(rdma);
1271 	svc_rdma_destroy_ctxts(rdma);
1272 	svc_rdma_destroy_maps(rdma);
1273 
1274 	/* Destroy the QP if present (not a listener) */
1275 	if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
1276 		ib_destroy_qp(rdma->sc_qp);
1277 
1278 	if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
1279 		ib_free_cq(rdma->sc_sq_cq);
1280 
1281 	if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
1282 		ib_free_cq(rdma->sc_rq_cq);
1283 
1284 	if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
1285 		ib_dealloc_pd(rdma->sc_pd);
1286 
1287 	/* Destroy the CM ID */
1288 	rdma_destroy_id(rdma->sc_cm_id);
1289 
1290 	kfree(rdma);
1291 }
1292 
svc_rdma_free(struct svc_xprt * xprt)1293 static void svc_rdma_free(struct svc_xprt *xprt)
1294 {
1295 	struct svcxprt_rdma *rdma =
1296 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
1297 	INIT_WORK(&rdma->sc_work, __svc_rdma_free);
1298 	queue_work(svc_rdma_wq, &rdma->sc_work);
1299 }
1300 
svc_rdma_has_wspace(struct svc_xprt * xprt)1301 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
1302 {
1303 	struct svcxprt_rdma *rdma =
1304 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
1305 
1306 	/*
1307 	 * If there are already waiters on the SQ,
1308 	 * return false.
1309 	 */
1310 	if (waitqueue_active(&rdma->sc_send_wait))
1311 		return 0;
1312 
1313 	/* Otherwise return true. */
1314 	return 1;
1315 }
1316 
svc_rdma_secure_port(struct svc_rqst * rqstp)1317 static int svc_rdma_secure_port(struct svc_rqst *rqstp)
1318 {
1319 	return 1;
1320 }
1321 
svc_rdma_kill_temp_xprt(struct svc_xprt * xprt)1322 static void svc_rdma_kill_temp_xprt(struct svc_xprt *xprt)
1323 {
1324 }
1325 
svc_rdma_send(struct svcxprt_rdma * xprt,struct ib_send_wr * wr)1326 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1327 {
1328 	struct ib_send_wr *bad_wr, *n_wr;
1329 	int wr_count;
1330 	int i;
1331 	int ret;
1332 
1333 	if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1334 		return -ENOTCONN;
1335 
1336 	wr_count = 1;
1337 	for (n_wr = wr->next; n_wr; n_wr = n_wr->next)
1338 		wr_count++;
1339 
1340 	/* If the SQ is full, wait until an SQ entry is available */
1341 	while (1) {
1342 		spin_lock_bh(&xprt->sc_lock);
1343 		if (xprt->sc_sq_depth < atomic_read(&xprt->sc_sq_count) + wr_count) {
1344 			spin_unlock_bh(&xprt->sc_lock);
1345 			atomic_inc(&rdma_stat_sq_starve);
1346 
1347 			/* Wait until SQ WR available if SQ still full */
1348 			wait_event(xprt->sc_send_wait,
1349 				   atomic_read(&xprt->sc_sq_count) <
1350 				   xprt->sc_sq_depth);
1351 			if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1352 				return -ENOTCONN;
1353 			continue;
1354 		}
1355 		/* Take a transport ref for each WR posted */
1356 		for (i = 0; i < wr_count; i++)
1357 			svc_xprt_get(&xprt->sc_xprt);
1358 
1359 		/* Bump used SQ WR count and post */
1360 		atomic_add(wr_count, &xprt->sc_sq_count);
1361 		ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1362 		if (ret) {
1363 			set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
1364 			atomic_sub(wr_count, &xprt->sc_sq_count);
1365 			for (i = 0; i < wr_count; i ++)
1366 				svc_xprt_put(&xprt->sc_xprt);
1367 			dprintk("svcrdma: failed to post SQ WR rc=%d, "
1368 			       "sc_sq_count=%d, sc_sq_depth=%d\n",
1369 			       ret, atomic_read(&xprt->sc_sq_count),
1370 			       xprt->sc_sq_depth);
1371 		}
1372 		spin_unlock_bh(&xprt->sc_lock);
1373 		if (ret)
1374 			wake_up(&xprt->sc_send_wait);
1375 		break;
1376 	}
1377 	return ret;
1378 }
1379