• 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/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 
svc_rdma_bc_create(struct svc_serv * serv,struct net * net,struct sockaddr * sa,int salen,int flags)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 
svc_rdma_bc_detach(struct svc_xprt * xprt)138 static void svc_rdma_bc_detach(struct svc_xprt *xprt)
139 {
140 	dprintk("svcrdma: %s(%p)\n", __func__, xprt);
141 }
142 
svc_rdma_bc_free(struct svc_xprt * xprt)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 
alloc_ctxt(struct svcxprt_rdma * xprt,gfp_t flags)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 
svc_rdma_prealloc_ctxts(struct svcxprt_rdma * xprt)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 
svc_rdma_get_context(struct svcxprt_rdma * xprt)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 
svc_rdma_unmap_dma(struct svc_rdma_op_ctxt * ctxt)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 
svc_rdma_put_context(struct svc_rdma_op_ctxt * ctxt,int free_pages)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 
svc_rdma_destroy_ctxts(struct svcxprt_rdma * xprt)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 */
qp_event_handler(struct ib_event * event,void * context)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 		break;
294 	}
295 }
296 
297 /**
298  * svc_rdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
299  * @cq:        completion queue
300  * @wc:        completed WR
301  *
302  */
svc_rdma_wc_receive(struct ib_cq * cq,struct ib_wc * wc)303 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
304 {
305 	struct svcxprt_rdma *xprt = cq->cq_context;
306 	struct ib_cqe *cqe = wc->wr_cqe;
307 	struct svc_rdma_op_ctxt *ctxt;
308 
309 	/* WARNING: Only wc->wr_cqe and wc->status are reliable */
310 	ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
311 	svc_rdma_unmap_dma(ctxt);
312 
313 	if (wc->status != IB_WC_SUCCESS)
314 		goto flushed;
315 
316 	/* All wc fields are now known to be valid */
317 	ctxt->byte_len = wc->byte_len;
318 	spin_lock(&xprt->sc_rq_dto_lock);
319 	list_add_tail(&ctxt->list, &xprt->sc_rq_dto_q);
320 	spin_unlock(&xprt->sc_rq_dto_lock);
321 
322 	set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
323 	if (test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
324 		goto out;
325 	svc_xprt_enqueue(&xprt->sc_xprt);
326 	goto out;
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:
337 	svc_xprt_put(&xprt->sc_xprt);
338 }
339 
340 /**
341  * svc_rdma_wc_send - Invoked by RDMA provider for each polled Send WC
342  * @cq:        completion queue
343  * @wc:        completed WR
344  *
345  */
svc_rdma_wc_send(struct ib_cq * cq,struct ib_wc * wc)346 void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
347 {
348 	struct svcxprt_rdma *xprt = cq->cq_context;
349 	struct ib_cqe *cqe = wc->wr_cqe;
350 	struct svc_rdma_op_ctxt *ctxt;
351 
352 	atomic_inc(&xprt->sc_sq_avail);
353 	wake_up(&xprt->sc_send_wait);
354 
355 	ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
356 	svc_rdma_unmap_dma(ctxt);
357 	svc_rdma_put_context(ctxt, 1);
358 
359 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
360 		set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
361 		if (wc->status != IB_WC_WR_FLUSH_ERR)
362 			pr_err("svcrdma: Send: %s (%u/0x%x)\n",
363 			       ib_wc_status_msg(wc->status),
364 			       wc->status, wc->vendor_err);
365 	}
366 
367 	svc_xprt_put(&xprt->sc_xprt);
368 }
369 
rdma_create_xprt(struct svc_serv * serv,int listener)370 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
371 					     int listener)
372 {
373 	struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
374 
375 	if (!cma_xprt)
376 		return NULL;
377 	svc_xprt_init(&init_net, &svc_rdma_class, &cma_xprt->sc_xprt, serv);
378 	INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
379 	INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
380 	INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
381 	INIT_LIST_HEAD(&cma_xprt->sc_ctxts);
382 	INIT_LIST_HEAD(&cma_xprt->sc_rw_ctxts);
383 	init_waitqueue_head(&cma_xprt->sc_send_wait);
384 
385 	spin_lock_init(&cma_xprt->sc_lock);
386 	spin_lock_init(&cma_xprt->sc_rq_dto_lock);
387 	spin_lock_init(&cma_xprt->sc_ctxt_lock);
388 	spin_lock_init(&cma_xprt->sc_rw_ctxt_lock);
389 
390 	/*
391 	 * Note that this implies that the underlying transport support
392 	 * has some form of congestion control (see RFC 7530 section 3.1
393 	 * paragraph 2). For now, we assume that all supported RDMA
394 	 * transports are suitable here.
395 	 */
396 	set_bit(XPT_CONG_CTRL, &cma_xprt->sc_xprt.xpt_flags);
397 
398 	if (listener)
399 		set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
400 
401 	return cma_xprt;
402 }
403 
svc_rdma_post_recv(struct svcxprt_rdma * xprt,gfp_t flags)404 int svc_rdma_post_recv(struct svcxprt_rdma *xprt, gfp_t flags)
405 {
406 	struct ib_recv_wr recv_wr, *bad_recv_wr;
407 	struct svc_rdma_op_ctxt *ctxt;
408 	struct page *page;
409 	dma_addr_t pa;
410 	int sge_no;
411 	int buflen;
412 	int ret;
413 
414 	ctxt = svc_rdma_get_context(xprt);
415 	buflen = 0;
416 	ctxt->direction = DMA_FROM_DEVICE;
417 	ctxt->cqe.done = svc_rdma_wc_receive;
418 	for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
419 		if (sge_no >= xprt->sc_max_sge) {
420 			pr_err("svcrdma: Too many sges (%d)\n", sge_no);
421 			goto err_put_ctxt;
422 		}
423 		page = alloc_page(flags);
424 		if (!page)
425 			goto err_put_ctxt;
426 		ctxt->pages[sge_no] = page;
427 		pa = ib_dma_map_page(xprt->sc_cm_id->device,
428 				     page, 0, PAGE_SIZE,
429 				     DMA_FROM_DEVICE);
430 		if (ib_dma_mapping_error(xprt->sc_cm_id->device, pa))
431 			goto err_put_ctxt;
432 		svc_rdma_count_mappings(xprt, ctxt);
433 		ctxt->sge[sge_no].addr = pa;
434 		ctxt->sge[sge_no].length = PAGE_SIZE;
435 		ctxt->sge[sge_no].lkey = xprt->sc_pd->local_dma_lkey;
436 		ctxt->count = sge_no + 1;
437 		buflen += PAGE_SIZE;
438 	}
439 	recv_wr.next = NULL;
440 	recv_wr.sg_list = &ctxt->sge[0];
441 	recv_wr.num_sge = ctxt->count;
442 	recv_wr.wr_cqe = &ctxt->cqe;
443 
444 	svc_xprt_get(&xprt->sc_xprt);
445 	ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
446 	if (ret) {
447 		svc_rdma_unmap_dma(ctxt);
448 		svc_rdma_put_context(ctxt, 1);
449 		svc_xprt_put(&xprt->sc_xprt);
450 	}
451 	return ret;
452 
453  err_put_ctxt:
454 	svc_rdma_unmap_dma(ctxt);
455 	svc_rdma_put_context(ctxt, 1);
456 	return -ENOMEM;
457 }
458 
svc_rdma_repost_recv(struct svcxprt_rdma * xprt,gfp_t flags)459 int svc_rdma_repost_recv(struct svcxprt_rdma *xprt, gfp_t flags)
460 {
461 	int ret = 0;
462 
463 	ret = svc_rdma_post_recv(xprt, flags);
464 	if (ret) {
465 		pr_err("svcrdma: could not post a receive buffer, err=%d.\n",
466 		       ret);
467 		pr_err("svcrdma: closing transport %p.\n", xprt);
468 		set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
469 		ret = -ENOTCONN;
470 	}
471 	return ret;
472 }
473 
474 static void
svc_rdma_parse_connect_private(struct svcxprt_rdma * newxprt,struct rdma_conn_param * param)475 svc_rdma_parse_connect_private(struct svcxprt_rdma *newxprt,
476 			       struct rdma_conn_param *param)
477 {
478 	const struct rpcrdma_connect_private *pmsg = param->private_data;
479 
480 	if (pmsg &&
481 	    pmsg->cp_magic == rpcrdma_cmp_magic &&
482 	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
483 		newxprt->sc_snd_w_inv = pmsg->cp_flags &
484 					RPCRDMA_CMP_F_SND_W_INV_OK;
485 
486 		dprintk("svcrdma: client send_size %u, recv_size %u "
487 			"remote inv %ssupported\n",
488 			rpcrdma_decode_buffer_size(pmsg->cp_send_size),
489 			rpcrdma_decode_buffer_size(pmsg->cp_recv_size),
490 			newxprt->sc_snd_w_inv ? "" : "un");
491 	}
492 }
493 
494 /*
495  * This function handles the CONNECT_REQUEST event on a listening
496  * endpoint. It is passed the cma_id for the _new_ connection. The context in
497  * this cma_id is inherited from the listening cma_id and is the svc_xprt
498  * structure for the listening endpoint.
499  *
500  * This function creates a new xprt for the new connection and enqueues it on
501  * the accept queue for the listent xprt. When the listen thread is kicked, it
502  * will call the recvfrom method on the listen xprt which will accept the new
503  * connection.
504  */
handle_connect_req(struct rdma_cm_id * new_cma_id,struct rdma_conn_param * param)505 static void handle_connect_req(struct rdma_cm_id *new_cma_id,
506 			       struct rdma_conn_param *param)
507 {
508 	struct svcxprt_rdma *listen_xprt = new_cma_id->context;
509 	struct svcxprt_rdma *newxprt;
510 	struct sockaddr *sa;
511 
512 	/* Create a new transport */
513 	newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
514 	if (!newxprt) {
515 		dprintk("svcrdma: failed to create new transport\n");
516 		return;
517 	}
518 	newxprt->sc_cm_id = new_cma_id;
519 	new_cma_id->context = newxprt;
520 	dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
521 		newxprt, newxprt->sc_cm_id, listen_xprt);
522 	svc_rdma_parse_connect_private(newxprt, param);
523 
524 	/* Save client advertised inbound read limit for use later in accept. */
525 	newxprt->sc_ord = param->initiator_depth;
526 
527 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
528 	svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
529 	/* The remote port is arbitrary and not under the control of the
530 	 * client ULP. Set it to a fixed value so that the DRC continues
531 	 * to be effective after a reconnect.
532 	 */
533 	rpc_set_port((struct sockaddr *)&newxprt->sc_xprt.xpt_remote, 0);
534 
535 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
536 	svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
537 
538 	/*
539 	 * Enqueue the new transport on the accept queue of the listening
540 	 * transport
541 	 */
542 	spin_lock_bh(&listen_xprt->sc_lock);
543 	list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
544 	spin_unlock_bh(&listen_xprt->sc_lock);
545 
546 	set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
547 	svc_xprt_enqueue(&listen_xprt->sc_xprt);
548 }
549 
550 /*
551  * Handles events generated on the listening endpoint. These events will be
552  * either be incoming connect requests or adapter removal  events.
553  */
rdma_listen_handler(struct rdma_cm_id * cma_id,struct rdma_cm_event * event)554 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
555 			       struct rdma_cm_event *event)
556 {
557 	struct svcxprt_rdma *xprt = cma_id->context;
558 	int ret = 0;
559 
560 	switch (event->event) {
561 	case RDMA_CM_EVENT_CONNECT_REQUEST:
562 		dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
563 			"event = %s (%d)\n", cma_id, cma_id->context,
564 			rdma_event_msg(event->event), event->event);
565 		handle_connect_req(cma_id, &event->param.conn);
566 		break;
567 
568 	case RDMA_CM_EVENT_ESTABLISHED:
569 		/* Accept complete */
570 		dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
571 			"cm_id=%p\n", xprt, cma_id);
572 		break;
573 
574 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
575 		dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
576 			xprt, cma_id);
577 		if (xprt)
578 			set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
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 
rdma_cma_handler(struct rdma_cm_id * cma_id,struct rdma_cm_event * event)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  */
svc_rdma_create(struct svc_serv * serv,struct net * net,struct sockaddr * sa,int salen,int flags)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  */
svc_rdma_accept(struct svc_xprt * xprt)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 
svc_rdma_release_rqst(struct svc_rqst * rqstp)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  */
svc_rdma_detach(struct svc_xprt * xprt)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 
__svc_rdma_free(struct work_struct * work)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 
svc_rdma_free(struct svc_xprt * xprt)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 
svc_rdma_has_wspace(struct svc_xprt * xprt)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 
svc_rdma_secure_port(struct svc_rqst * rqstp)1006 static int svc_rdma_secure_port(struct svc_rqst *rqstp)
1007 {
1008 	return 1;
1009 }
1010 
svc_rdma_kill_temp_xprt(struct svc_xprt * xprt)1011 static void svc_rdma_kill_temp_xprt(struct svc_xprt *xprt)
1012 {
1013 }
1014 
svc_rdma_send(struct svcxprt_rdma * xprt,struct ib_send_wr * wr)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 }
1062