xref: /freebsd/sys/dev/iser/iser_verbs.c (revision 3416500aef140042c64bc149cb1ec6620483bc44)
1 /* $FreeBSD$ */
2 /*-
3  * Copyright (c) 2015, Mellanox Technologies, Inc. All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 
27 #include "icl_iser.h"
28 
29 static MALLOC_DEFINE(M_ISER_VERBS, "iser_verbs", "iser verbs backend");
30 static int iser_cq_poll_limit = 512;
31 
32 static void
33 iser_cq_event_callback(struct ib_event *cause, void *context)
34 {
35 	ISER_ERR("got cq event %d", cause->event);
36 }
37 
38 static void
39 iser_qp_event_callback(struct ib_event *cause, void *context)
40 {
41 	ISER_ERR("got qp event %d", cause->event);
42 }
43 
44 static void
45 iser_event_handler(struct ib_event_handler *handler,
46 				struct ib_event *event)
47 {
48 	ISER_ERR("async event %d on device %s port %d",
49 		 event->event, event->device->name,
50 		 event->element.port_num);
51 }
52 
53 /**
54  * is_iser_tx_desc - Indicate if the completion wr_id
55  *     is a TX descriptor or not.
56  * @iser_conn: iser connection
57  * @wr_id: completion WR identifier
58  *
59  * Since we cannot rely on wc opcode in FLUSH errors
60  * we must work around it by checking if the wr_id address
61  * falls in the iser connection rx_descs buffer. If so
62  * it is an RX descriptor, otherwize it is a TX.
63  */
64 static inline bool
65 is_iser_tx_desc(struct iser_conn *iser_conn, void *wr_id)
66 {
67 	void *start = iser_conn->rx_descs;
68 	u64 len = iser_conn->num_rx_descs * sizeof(*iser_conn->rx_descs);
69 	void *end = (void *)((uintptr_t)start + (uintptr_t)len);
70 
71 	if (start) {
72 		if (wr_id >= start && wr_id < end)
73 			return false;
74 	} else {
75 		return ((uintptr_t)wr_id != (uintptr_t)iser_conn->login_resp_buf);
76 	}
77 
78 	return true;
79 }
80 
81 /**
82  * iser_handle_comp_error() - Handle error completion
83  * @ib_conn:   connection RDMA resources
84  * @wc:        work completion
85  *
86  * Notes: Update post_recv_buf_count in case of recv error completion.
87  *        For non-FLUSH error completion we should also notify iscsi layer that
88  *        connection is failed (in case we passed bind stage).
89  */
90 static void
91 iser_handle_comp_error(struct ib_conn *ib_conn,
92 		       struct ib_wc *wc)
93 {
94 	void *wr_id = (void *)(uintptr_t)wc->wr_id;
95 	struct iser_conn *iser_conn = container_of(ib_conn, struct iser_conn,
96 						   ib_conn);
97 
98 	if (is_iser_tx_desc(iser_conn, wr_id)) {
99 		ISER_DBG("conn %p got send comp error", iser_conn);
100 	} else {
101 		ISER_DBG("conn %p got recv comp error", iser_conn);
102 		ib_conn->post_recv_buf_count--;
103 	}
104 	if (wc->status != IB_WC_WR_FLUSH_ERR)
105 		iser_conn->icl_conn.ic_error(&iser_conn->icl_conn);
106 }
107 
108 /**
109  * iser_handle_wc - handle a single work completion
110  * @wc: work completion
111  *
112  * Soft-IRQ context, work completion can be either
113  * SEND or RECV, and can turn out successful or
114  * with error (or flush error).
115  */
116 static void iser_handle_wc(struct ib_wc *wc)
117 {
118 	struct ib_conn *ib_conn;
119 	struct iser_tx_desc *tx_desc;
120 	struct iser_rx_desc *rx_desc;
121 
122 	ib_conn = wc->qp->qp_context;
123 	if (likely(wc->status == IB_WC_SUCCESS)) {
124 		if (wc->opcode == IB_WC_RECV) {
125 			rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
126 			iser_rcv_completion(rx_desc, wc->byte_len,
127 					    ib_conn);
128 		} else
129 		if (wc->opcode == IB_WC_SEND) {
130 			tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
131 			iser_snd_completion(tx_desc, ib_conn);
132 		} else {
133 			ISER_ERR("Unknown wc opcode %d", wc->opcode);
134 		}
135 	} else {
136 		struct iser_conn *iser_conn = container_of(ib_conn, struct iser_conn,
137 					ib_conn);
138 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
139 			ISER_ERR("conn %p wr id %llx status %d vend_err %x",
140 				 iser_conn, (unsigned long long)wc->wr_id,
141 				 wc->status, wc->vendor_err);
142 		} else {
143 			ISER_DBG("flush error: conn %p wr id %llx",
144 				 iser_conn, (unsigned long long)wc->wr_id);
145 		}
146 
147 		if (wc->wr_id == ISER_BEACON_WRID) {
148 			/* all flush errors were consumed */
149 			mtx_lock(&ib_conn->beacon.flush_lock);
150 			ISER_DBG("conn %p got ISER_BEACON_WRID", iser_conn);
151 			cv_signal(&ib_conn->beacon.flush_cv);
152 			mtx_unlock(&ib_conn->beacon.flush_lock);
153 		} else {
154 			iser_handle_comp_error(ib_conn, wc);
155 		}
156 	}
157 }
158 
159 static void
160 iser_cq_tasklet_fn(void *data, int pending)
161 {
162 	struct iser_comp *comp = (struct iser_comp *)data;
163 	struct ib_cq *cq = comp->cq;
164 	struct ib_wc *const wcs = comp->wcs;
165 	int completed = 0;
166 	int i;
167 	int n;
168 
169 	while ((n = ib_poll_cq(cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
170 		for (i = 0; i < n; i++)
171 			iser_handle_wc(&wcs[i]);
172 
173 		completed += n;
174 		if (completed >= iser_cq_poll_limit)
175 			break;
176 	}
177 
178 	/*
179 	 * It is assumed here that arming CQ only once its empty
180 	 * would not cause interrupts to be missed.
181 	 */
182 	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
183 }
184 
185 static void
186 iser_cq_callback(struct ib_cq *cq, void *cq_context)
187 {
188 	struct iser_comp *comp = cq_context;
189 
190 	taskqueue_enqueue(comp->tq, &comp->task);
191 }
192 
193 /**
194  * iser_create_device_ib_res - creates Protection Domain (PD), Completion
195  * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
196  * the adapator.
197  *
198  * returns 0 on success, -1 on failure
199  */
200 static int
201 iser_create_device_ib_res(struct iser_device *device)
202 {
203 	struct ib_device_attr *dev_attr = &device->dev_attr;
204 	int ret, i, max_cqe;
205 
206 	ret = ib_query_device(device->ib_device, dev_attr);
207 	if (ret) {
208 		ISER_ERR("Query device failed for %s", device->ib_device->name);
209 		return (ret);
210 	}
211 
212 	if (!(dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS)) {
213 		ISER_ERR("device %s doesn't support Fastreg, "
214 			 "can't register memory", device->ib_device->name);
215 		return (1);
216 	}
217 
218 	device->comps_used = min(mp_ncpus, device->ib_device->num_comp_vectors);
219 
220 	device->comps = malloc(device->comps_used * sizeof(*device->comps),
221 		M_ISER_VERBS, M_WAITOK | M_ZERO);
222 	if (!device->comps)
223 		goto comps_err;
224 
225 	max_cqe = min(ISER_MAX_CQ_LEN, dev_attr->max_cqe);
226 
227 	ISER_DBG("using %d CQs, device %s supports %d vectors max_cqe %d",
228 		 device->comps_used, device->ib_device->name,
229 		 device->ib_device->num_comp_vectors, max_cqe);
230 
231 	device->pd = ib_alloc_pd(device->ib_device);
232 	if (IS_ERR(device->pd))
233 		goto pd_err;
234 
235 	for (i = 0; i < device->comps_used; i++) {
236 		struct iser_comp *comp = &device->comps[i];
237 
238 		comp->device = device;
239 		comp->cq = ib_create_cq(device->ib_device,
240 					iser_cq_callback,
241 					iser_cq_event_callback,
242 					(void *)comp,
243 					max_cqe, i);
244 		if (IS_ERR(comp->cq)) {
245 			comp->cq = NULL;
246 			goto cq_err;
247 		}
248 
249 		if (ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP))
250 			goto cq_err;
251 
252 		TASK_INIT(&comp->task, 0, iser_cq_tasklet_fn, comp);
253 		comp->tq = taskqueue_create_fast("iser_taskq", M_NOWAIT,
254 				taskqueue_thread_enqueue, &comp->tq);
255 		if (!comp->tq)
256 			goto tq_err;
257 		taskqueue_start_threads(&comp->tq, 1, PI_NET, "iser taskq");
258 	}
259 
260 	device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
261 				   IB_ACCESS_REMOTE_WRITE |
262 				   IB_ACCESS_REMOTE_READ);
263 	if (IS_ERR(device->mr))
264 		goto tq_err;
265 
266 	INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
267 				iser_event_handler);
268 	if (ib_register_event_handler(&device->event_handler))
269 		goto handler_err;
270 
271 	return (0);
272 
273 handler_err:
274 	ib_dereg_mr(device->mr);
275 tq_err:
276 	for (i = 0; i < device->comps_used; i++) {
277 		struct iser_comp *comp = &device->comps[i];
278 		if (comp->tq)
279 			taskqueue_free(comp->tq);
280 	}
281 cq_err:
282 	for (i = 0; i < device->comps_used; i++) {
283 		struct iser_comp *comp = &device->comps[i];
284 		if (comp->cq)
285 			ib_destroy_cq(comp->cq);
286 	}
287 	ib_dealloc_pd(device->pd);
288 pd_err:
289 	free(device->comps, M_ISER_VERBS);
290 comps_err:
291 	ISER_ERR("failed to allocate an IB resource");
292 	return (1);
293 }
294 
295 /**
296  * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
297  * CQ and PD created with the device associated with the adapator.
298  */
299 static void
300 iser_free_device_ib_res(struct iser_device *device)
301 {
302 	int i;
303 
304 	for (i = 0; i < device->comps_used; i++) {
305 		struct iser_comp *comp = &device->comps[i];
306 
307 		taskqueue_free(comp->tq);
308 		ib_destroy_cq(comp->cq);
309 		comp->cq = NULL;
310 	}
311 
312 	(void)ib_unregister_event_handler(&device->event_handler);
313 	(void)ib_dereg_mr(device->mr);
314 	(void)ib_dealloc_pd(device->pd);
315 
316 	free(device->comps, M_ISER_VERBS);
317 	device->comps = NULL;
318 
319 	device->mr = NULL;
320 	device->pd = NULL;
321 }
322 
323 static int
324 iser_alloc_reg_res(struct ib_device *ib_device,
325 		   struct ib_pd *pd,
326 		   struct iser_reg_resources *res)
327 {
328 	int ret;
329 
330 	res->frpl = ib_alloc_fast_reg_page_list(ib_device,
331 						ISCSI_ISER_SG_TABLESIZE + 1);
332 	if (IS_ERR(res->frpl)) {
333 		ret = -PTR_ERR(res->frpl);
334 		ISER_ERR("Failed to allocate fast reg page list err=%d", ret);
335 		return (ret);
336 	}
337 
338 	res->mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE + 1);
339 	if (IS_ERR(res->mr)) {
340 		ret = -PTR_ERR(res->mr);
341 		ISER_ERR("Failed to allocate  fast reg mr err=%d", ret);
342 		goto fast_reg_mr_failure;
343 	}
344 	res->mr_valid = 1;
345 
346 	return (0);
347 
348 fast_reg_mr_failure:
349 	ib_free_fast_reg_page_list(res->frpl);
350 
351 	return (ret);
352 }
353 
354 static void
355 iser_free_reg_res(struct iser_reg_resources *rsc)
356 {
357 	ib_dereg_mr(rsc->mr);
358 	ib_free_fast_reg_page_list(rsc->frpl);
359 }
360 
361 static struct fast_reg_descriptor *
362 iser_create_fastreg_desc(struct ib_device *ib_device, struct ib_pd *pd)
363 {
364 	struct fast_reg_descriptor *desc;
365 	int ret;
366 
367 	desc = malloc(sizeof(*desc), M_ISER_VERBS, M_WAITOK | M_ZERO);
368 	if (!desc) {
369 		ISER_ERR("Failed to allocate a new fastreg descriptor");
370 		return (NULL);
371 	}
372 
373 	ret = iser_alloc_reg_res(ib_device, pd, &desc->rsc);
374 	if (ret) {
375 		ISER_ERR("failed to allocate reg_resources");
376 		goto err;
377 	}
378 
379 	return (desc);
380 err:
381 	free(desc, M_ISER_VERBS);
382 	return (NULL);
383 }
384 
385 /**
386  * iser_create_fmr_pool - Creates FMR pool and page_vector
387  *
388  * returns 0 on success, or errno code on failure
389  */
390 int
391 iser_create_fastreg_pool(struct ib_conn *ib_conn, unsigned cmds_max)
392 {
393 	struct iser_device *device = ib_conn->device;
394 	struct fast_reg_descriptor *desc;
395 	int i;
396 
397 	INIT_LIST_HEAD(&ib_conn->fastreg.pool);
398 	ib_conn->fastreg.pool_size = 0;
399 	for (i = 0; i < cmds_max; i++) {
400 		desc = iser_create_fastreg_desc(device->ib_device, device->pd);
401 		if (!desc) {
402 			ISER_ERR("Failed to create fastreg descriptor");
403 			goto err;
404 		}
405 
406 		list_add_tail(&desc->list, &ib_conn->fastreg.pool);
407 		ib_conn->fastreg.pool_size++;
408 	}
409 
410 	return (0);
411 
412 err:
413 	iser_free_fastreg_pool(ib_conn);
414 	return (ENOMEM);
415 }
416 
417 /**
418  * iser_free_fmr_pool - releases the FMR pool and page vec
419  */
420 void
421 iser_free_fastreg_pool(struct ib_conn *ib_conn)
422 {
423 	struct fast_reg_descriptor *desc, *tmp;
424 	int i = 0;
425 
426 	if (list_empty(&ib_conn->fastreg.pool))
427 		return;
428 
429 	ISER_DBG("freeing conn %p fr pool", ib_conn);
430 
431 	list_for_each_entry_safe(desc, tmp, &ib_conn->fastreg.pool, list) {
432 		list_del(&desc->list);
433 		iser_free_reg_res(&desc->rsc);
434 		free(desc, M_ISER_VERBS);
435 		++i;
436 	}
437 
438 	if (i < ib_conn->fastreg.pool_size)
439 		ISER_WARN("pool still has %d regions registered",
440 			  ib_conn->fastreg.pool_size - i);
441 }
442 
443 /**
444  * iser_create_ib_conn_res - Queue-Pair (QP)
445  *
446  * returns 0 on success, 1 on failure
447  */
448 static int
449 iser_create_ib_conn_res(struct ib_conn *ib_conn)
450 {
451 	struct iser_conn *iser_conn;
452 	struct iser_device *device;
453 	struct ib_device_attr *dev_attr;
454 	struct ib_qp_init_attr init_attr;
455 	int index, min_index = 0;
456 	int ret = -ENOMEM;
457 
458 	iser_conn = container_of(ib_conn, struct iser_conn, ib_conn);
459 	device = ib_conn->device;
460 	dev_attr = &device->dev_attr;
461 
462 	mtx_lock(&ig.connlist_mutex);
463 	/* select the CQ with the minimal number of usages */
464 	for (index = 0; index < device->comps_used; index++) {
465 		if (device->comps[index].active_qps <
466 		    device->comps[min_index].active_qps)
467 			min_index = index;
468 	}
469 	ib_conn->comp = &device->comps[min_index];
470 	ib_conn->comp->active_qps++;
471 	mtx_unlock(&ig.connlist_mutex);
472 	ISER_INFO("cq index %d used for ib_conn %p", min_index, ib_conn);
473 
474 	memset(&init_attr, 0, sizeof init_attr);
475 	init_attr.event_handler = iser_qp_event_callback;
476 	init_attr.qp_context	= (void *)ib_conn;
477 	init_attr.send_cq	= ib_conn->comp->cq;
478 	init_attr.recv_cq	= ib_conn->comp->cq;
479 	init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
480 	init_attr.cap.max_send_sge = 2;
481 	init_attr.cap.max_recv_sge = 1;
482 	init_attr.sq_sig_type	= IB_SIGNAL_REQ_WR;
483 	init_attr.qp_type	= IB_QPT_RC;
484 
485 	if (dev_attr->max_qp_wr > ISER_QP_MAX_REQ_DTOS) {
486 		init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS;
487 		iser_conn->max_cmds =
488 			ISER_GET_MAX_XMIT_CMDS(ISER_QP_MAX_REQ_DTOS);
489 	} else {
490 		init_attr.cap.max_send_wr = dev_attr->max_qp_wr;
491 		iser_conn->max_cmds =
492 			ISER_GET_MAX_XMIT_CMDS(dev_attr->max_qp_wr);
493 	}
494 	ISER_DBG("device %s supports max_send_wr %d",
495 	         device->ib_device->name, dev_attr->max_qp_wr);
496 
497 	ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
498 	if (ret)
499 		goto out_err;
500 
501 	ib_conn->qp = ib_conn->cma_id->qp;
502 	ISER_DBG("setting conn %p cma_id %p qp %p",
503 		 ib_conn, ib_conn->cma_id,
504 		 ib_conn->cma_id->qp);
505 
506 	return (ret);
507 
508 out_err:
509 	mtx_lock(&ig.connlist_mutex);
510 	ib_conn->comp->active_qps--;
511 	mtx_unlock(&ig.connlist_mutex);
512 	ISER_ERR("unable to alloc mem or create resource, err %d", ret);
513 
514 	return (ret);
515 }
516 
517 /**
518  * based on the resolved device node GUID see if there already allocated
519  * device for this device. If there's no such, create one.
520  */
521 static struct iser_device *
522 iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
523 {
524 	struct iser_device *device;
525 
526 	sx_xlock(&ig.device_list_mutex);
527 
528 	list_for_each_entry(device, &ig.device_list, ig_list)
529 		/* find if there's a match using the node GUID */
530 		if (device->ib_device->node_guid == cma_id->device->node_guid)
531 			goto inc_refcnt;
532 
533 	device = malloc(sizeof *device, M_ISER_VERBS, M_WAITOK | M_ZERO);
534 	if (device == NULL)
535 		goto out;
536 
537 	/* assign this device to the device */
538 	device->ib_device = cma_id->device;
539 	/* init the device and link it into ig device list */
540 	if (iser_create_device_ib_res(device)) {
541 		free(device, M_ISER_VERBS);
542 		device = NULL;
543 		goto out;
544 	}
545 	list_add(&device->ig_list, &ig.device_list);
546 
547 inc_refcnt:
548 	device->refcount++;
549 	ISER_INFO("device %p refcount %d", device, device->refcount);
550 out:
551 	sx_xunlock(&ig.device_list_mutex);
552 	return (device);
553 }
554 
555 /* if there's no demand for this device, release it */
556 static void
557 iser_device_try_release(struct iser_device *device)
558 {
559 	sx_xlock(&ig.device_list_mutex);
560 	device->refcount--;
561 	ISER_INFO("device %p refcount %d", device, device->refcount);
562 	if (!device->refcount) {
563 		iser_free_device_ib_res(device);
564 		list_del(&device->ig_list);
565 		free(device, M_ISER_VERBS);
566 		device = NULL;
567 	}
568 	sx_xunlock(&ig.device_list_mutex);
569 }
570 
571 /**
572  * Called with state mutex held
573  **/
574 static int iser_conn_state_comp_exch(struct iser_conn *iser_conn,
575 				     enum iser_conn_state comp,
576 				     enum iser_conn_state exch)
577 {
578 	int ret;
579 
580 	ret = (iser_conn->state == comp);
581 	if (ret)
582 		iser_conn->state = exch;
583 
584 	return ret;
585 }
586 
587 /**
588  * iser_free_ib_conn_res - release IB related resources
589  * @iser_conn: iser connection struct
590  * @destroy: indicator if we need to try to release the
591  *     iser device and memory regoins pool (only iscsi
592  *     shutdown and DEVICE_REMOVAL will use this).
593  *
594  * This routine is called with the iser state mutex held
595  * so the cm_id removal is out of here. It is Safe to
596  * be invoked multiple times.
597  */
598 void
599 iser_free_ib_conn_res(struct iser_conn *iser_conn,
600 				  bool destroy)
601 {
602 	struct ib_conn *ib_conn = &iser_conn->ib_conn;
603 	struct iser_device *device = ib_conn->device;
604 
605 	ISER_INFO("freeing conn %p cma_id %p qp %p",
606 		  iser_conn, ib_conn->cma_id, ib_conn->qp);
607 
608 	if (ib_conn->qp != NULL) {
609 		mtx_lock(&ig.connlist_mutex);
610 		ib_conn->comp->active_qps--;
611 		mtx_unlock(&ig.connlist_mutex);
612 		rdma_destroy_qp(ib_conn->cma_id);
613 		ib_conn->qp = NULL;
614 	}
615 
616 	if (destroy) {
617 		if (iser_conn->login_buf)
618 			iser_free_login_buf(iser_conn);
619 
620 		if (iser_conn->rx_descs)
621 			iser_free_rx_descriptors(iser_conn);
622 
623 		if (device != NULL) {
624 			iser_device_try_release(device);
625 			ib_conn->device = NULL;
626 		}
627 	}
628 }
629 
630 /**
631  * triggers start of the disconnect procedures and wait for them to be done
632  * Called with state mutex held
633  */
634 int
635 iser_conn_terminate(struct iser_conn *iser_conn)
636 {
637 	struct ib_conn *ib_conn = &iser_conn->ib_conn;
638 	struct ib_send_wr *bad_send_wr;
639 	struct ib_recv_wr *bad_recv_wr;
640 	int err = 0;
641 
642 	/* terminate the iser conn only if the conn state is UP */
643 	if (!iser_conn_state_comp_exch(iser_conn, ISER_CONN_UP,
644 					   ISER_CONN_TERMINATING))
645 		return (0);
646 
647 	ISER_INFO("iser_conn %p state %d\n", iser_conn, iser_conn->state);
648 
649 	if (ib_conn->qp == NULL) {
650 		/* HOW can this be??? */
651 		ISER_WARN("qp wasn't created");
652 		return (1);
653 	}
654 
655 	/*
656 	 * Todo: This is a temporary workaround.
657 	 * We serialize the connection closure using global lock in order to
658 	 * receive all posted beacons completions.
659 	 * Without Serialization, in case we open many connections (QPs) on
660 	 * the same CQ, we might miss beacons because of missing interrupts.
661 	 */
662 	sx_xlock(&ig.close_conns_mutex);
663 
664 	/*
665 	 * In case we didn't already clean up the cma_id (peer initiated
666 	 * a disconnection), we need to Cause the CMA to change the QP
667 	 * state to ERROR.
668 	 */
669 	if (ib_conn->cma_id) {
670 		err = rdma_disconnect(ib_conn->cma_id);
671 		if (err)
672 			ISER_ERR("Failed to disconnect, conn: 0x%p err %d",
673 				iser_conn, err);
674 
675 		mtx_lock(&ib_conn->beacon.flush_lock);
676 		memset(&ib_conn->beacon.send, 0, sizeof(struct ib_send_wr));
677 		ib_conn->beacon.send.wr_id = ISER_BEACON_WRID;
678 		ib_conn->beacon.send.opcode = IB_WR_SEND;
679 		/* post an indication that all send flush errors were consumed */
680 		err = ib_post_send(ib_conn->qp, &ib_conn->beacon.send, &bad_send_wr);
681 		if (err) {
682 			ISER_ERR("conn %p failed to post send_beacon", ib_conn);
683 			mtx_unlock(&ib_conn->beacon.flush_lock);
684 			goto out;
685 		}
686 
687 		ISER_DBG("before send cv_wait: %p", iser_conn);
688 		cv_wait(&ib_conn->beacon.flush_cv, &ib_conn->beacon.flush_lock);
689 		ISER_DBG("after send cv_wait: %p", iser_conn);
690 
691 		memset(&ib_conn->beacon.recv, 0, sizeof(struct ib_recv_wr));
692 		ib_conn->beacon.recv.wr_id = ISER_BEACON_WRID;
693 		/* post an indication that all recv flush errors were consumed */
694 		err = ib_post_recv(ib_conn->qp, &ib_conn->beacon.recv, &bad_recv_wr);
695 		if (err) {
696 			ISER_ERR("conn %p failed to post recv_beacon", ib_conn);
697 			mtx_unlock(&ib_conn->beacon.flush_lock);
698 			goto out;
699 		}
700 
701 		ISER_DBG("before recv cv_wait: %p", iser_conn);
702 		cv_wait(&ib_conn->beacon.flush_cv, &ib_conn->beacon.flush_lock);
703 		mtx_unlock(&ib_conn->beacon.flush_lock);
704 		ISER_DBG("after recv cv_wait: %p", iser_conn);
705 	}
706 out:
707 	sx_xunlock(&ig.close_conns_mutex);
708 	return (1);
709 }
710 
711 /**
712  * Called with state mutex held
713  **/
714 static void
715 iser_connect_error(struct rdma_cm_id *cma_id)
716 {
717 	struct iser_conn *iser_conn;
718 
719 	iser_conn = cma_id->context;
720 
721 	ISER_ERR("conn %p", iser_conn);
722 
723 	iser_conn->state = ISER_CONN_TERMINATING;
724 
725 	cv_signal(&iser_conn->up_cv);
726 }
727 
728 /**
729  * Called with state mutex held
730  **/
731 static void
732 iser_addr_handler(struct rdma_cm_id *cma_id)
733 {
734 	struct iser_device *device;
735 	struct iser_conn   *iser_conn;
736 	struct ib_conn   *ib_conn;
737 	int    ret;
738 
739 	iser_conn = cma_id->context;
740 
741 	ib_conn = &iser_conn->ib_conn;
742 	device = iser_device_find_by_ib_device(cma_id);
743 	if (!device) {
744 		ISER_ERR("conn %p device lookup/creation failed",
745 			 iser_conn);
746 		iser_connect_error(cma_id);
747 		return;
748 	}
749 
750 	ib_conn->device = device;
751 
752 	ret = rdma_resolve_route(cma_id, 1000);
753 	if (ret) {
754 		ISER_ERR("conn %p resolve route failed: %d", iser_conn, ret);
755 		iser_connect_error(cma_id);
756 		return;
757 	}
758 }
759 
760 /**
761  * Called with state mutex held
762  **/
763 static void
764 iser_route_handler(struct rdma_cm_id *cma_id)
765 {
766 	struct rdma_conn_param conn_param;
767 	int    ret;
768 	struct iser_cm_hdr req_hdr;
769 	struct iser_conn *iser_conn = cma_id->context;
770 	struct ib_conn *ib_conn = &iser_conn->ib_conn;
771 	struct iser_device *device = ib_conn->device;
772 
773 	ret = iser_create_ib_conn_res(ib_conn);
774 	if (ret)
775 		goto failure;
776 
777 	memset(&conn_param, 0, sizeof conn_param);
778 	conn_param.responder_resources = device->dev_attr.max_qp_rd_atom;
779 	conn_param.retry_count	       = 7;
780 	conn_param.rnr_retry_count     = 6;
781 	/*
782 	 * Initiaotr depth should not be set, but in order to compat
783 	 * with old targets, we keep this value set.
784 	 */
785 	conn_param.initiator_depth     = 1;
786 
787 	memset(&req_hdr, 0, sizeof(req_hdr));
788 	req_hdr.flags = (ISER_ZBVA_NOT_SUPPORTED |
789 			ISER_SEND_W_INV_NOT_SUPPORTED);
790 	conn_param.private_data		= (void *)&req_hdr;
791 	conn_param.private_data_len	= sizeof(struct iser_cm_hdr);
792 
793 	ret = rdma_connect(cma_id, &conn_param);
794 	if (ret) {
795 		ISER_ERR("conn %p failure connecting: %d", iser_conn, ret);
796 		goto failure;
797 	}
798 
799 	return;
800 failure:
801 	iser_connect_error(cma_id);
802 }
803 
804 /**
805  * Called with state mutex held
806  **/
807 static void
808 iser_connected_handler(struct rdma_cm_id *cma_id)
809 {
810 	struct iser_conn *iser_conn;
811 	struct ib_qp_attr attr;
812 	struct ib_qp_init_attr init_attr;
813 
814 	iser_conn = cma_id->context;
815 
816 	(void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
817 
818 	ISER_INFO("remote qpn:%x my qpn:%x",
819 		  attr.dest_qp_num, cma_id->qp->qp_num);
820 
821 	iser_conn->state = ISER_CONN_UP;
822 
823 	cv_signal(&iser_conn->up_cv);
824 }
825 
826 /**
827  * Called with state mutex held
828  **/
829 static void
830 iser_cleanup_handler(struct rdma_cm_id *cma_id, bool destroy)
831 {
832 	struct iser_conn *iser_conn = cma_id->context;
833 
834 	if (iser_conn_terminate(iser_conn))
835 		iser_conn->icl_conn.ic_error(&iser_conn->icl_conn);
836 
837 }
838 
839 int
840 iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
841 {
842 	struct iser_conn *iser_conn;
843 	int ret = 0;
844 
845 	iser_conn = cma_id->context;
846 	ISER_INFO("event %d status %d conn %p id %p",
847 		  event->event, event->status, cma_id->context, cma_id);
848 
849 	sx_xlock(&iser_conn->state_mutex);
850 	switch (event->event) {
851 	case RDMA_CM_EVENT_ADDR_RESOLVED:
852 		iser_addr_handler(cma_id);
853 		break;
854 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
855 		iser_route_handler(cma_id);
856 		break;
857 	case RDMA_CM_EVENT_ESTABLISHED:
858 		iser_connected_handler(cma_id);
859 		break;
860 	case RDMA_CM_EVENT_ADDR_ERROR:
861 	case RDMA_CM_EVENT_ROUTE_ERROR:
862 	case RDMA_CM_EVENT_CONNECT_ERROR:
863 	case RDMA_CM_EVENT_UNREACHABLE:
864 	case RDMA_CM_EVENT_REJECTED:
865 		iser_connect_error(cma_id);
866 		break;
867 	case RDMA_CM_EVENT_DISCONNECTED:
868 	case RDMA_CM_EVENT_ADDR_CHANGE:
869 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
870 		iser_cleanup_handler(cma_id, false);
871 		break;
872 	default:
873 		ISER_ERR("Unexpected RDMA CM event (%d)", event->event);
874 		break;
875 	}
876 	sx_xunlock(&iser_conn->state_mutex);
877 
878 	return (ret);
879 }
880 
881 int
882 iser_post_recvl(struct iser_conn *iser_conn)
883 {
884 	struct ib_recv_wr rx_wr, *rx_wr_failed;
885 	struct ib_conn *ib_conn = &iser_conn->ib_conn;
886 	struct ib_sge	  sge;
887 	int ib_ret;
888 
889 	sge.addr   = iser_conn->login_resp_dma;
890 	sge.length = ISER_RX_LOGIN_SIZE;
891 	sge.lkey   = ib_conn->device->mr->lkey;
892 
893 	rx_wr.wr_id   = (uintptr_t)iser_conn->login_resp_buf;
894 	rx_wr.sg_list = &sge;
895 	rx_wr.num_sge = 1;
896 	rx_wr.next    = NULL;
897 
898 	ib_conn->post_recv_buf_count++;
899 	ib_ret	= ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
900 	if (ib_ret) {
901 		ISER_ERR("ib_post_recv failed ret=%d", ib_ret);
902 		ib_conn->post_recv_buf_count--;
903 	}
904 
905 	return (ib_ret);
906 }
907 
908 int
909 iser_post_recvm(struct iser_conn *iser_conn, int count)
910 {
911 	struct ib_recv_wr *rx_wr, *rx_wr_failed;
912 	int i, ib_ret;
913 	struct ib_conn *ib_conn = &iser_conn->ib_conn;
914 	unsigned int my_rx_head = iser_conn->rx_desc_head;
915 	struct iser_rx_desc *rx_desc;
916 
917 	for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
918 		rx_desc		= &iser_conn->rx_descs[my_rx_head];
919 		rx_wr->wr_id	= (uintptr_t)rx_desc;
920 		rx_wr->sg_list	= &rx_desc->rx_sg;
921 		rx_wr->num_sge	= 1;
922 		rx_wr->next	= rx_wr + 1;
923 		my_rx_head = (my_rx_head + 1) % iser_conn->qp_max_recv_dtos;
924 	}
925 
926 	rx_wr--;
927 	rx_wr->next = NULL; /* mark end of work requests list */
928 
929 	ib_conn->post_recv_buf_count += count;
930 	ib_ret	= ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
931 	if (ib_ret) {
932 		ISER_ERR("ib_post_recv failed ret=%d", ib_ret);
933 		ib_conn->post_recv_buf_count -= count;
934 	} else
935 		iser_conn->rx_desc_head = my_rx_head;
936 
937 	return (ib_ret);
938 }
939 
940 /**
941  * iser_start_send - Initiate a Send DTO operation
942  *
943  * returns 0 on success, -1 on failure
944  */
945 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
946 		   bool signal)
947 {
948 	int		  ib_ret;
949 	struct ib_send_wr send_wr, *send_wr_failed;
950 
951 	ib_dma_sync_single_for_device(ib_conn->device->ib_device,
952 				      tx_desc->dma_addr, ISER_HEADERS_LEN,
953 				      DMA_TO_DEVICE);
954 
955 	send_wr.next	   = NULL;
956 	send_wr.wr_id	   = (uintptr_t)tx_desc;
957 	send_wr.sg_list	   = tx_desc->tx_sg;
958 	send_wr.num_sge	   = tx_desc->num_sge;
959 	send_wr.opcode	   = IB_WR_SEND;
960 	send_wr.send_flags = signal ? IB_SEND_SIGNALED : 0;
961 
962 	ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
963 	if (ib_ret)
964 		ISER_ERR("ib_post_send failed, ret:%d", ib_ret);
965 
966 	return (ib_ret);
967 }
968